<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Jewlya Lynn on Medium]]></title>
        <description><![CDATA[Stories by Jewlya Lynn on Medium]]></description>
        <link>https://medium.com/@jewlya-lynn?source=rss-37ab969aa6b4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*6Lzm0PVq_Y91Nzzi2ilN_g.jpeg</url>
            <title>Stories by Jewlya Lynn on Medium</title>
            <link>https://medium.com/@jewlya-lynn?source=rss-37ab969aa6b4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 23:40:15 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jewlya-lynn/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Systems Change with AI: Looking Back and Looking Forward]]></title>
            <link>https://medium.com/ai-systems-change/systems-change-with-ai-looking-back-and-looking-forward-4e78d10ba3c1?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/4e78d10ba3c1</guid>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[systems-change]]></category>
            <category><![CDATA[history]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 11:07:29 GMT</pubDate>
            <atom:updated>2026-11-10T11:07:29.426Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The visible and hidden dynamics in any complex system have gradually developed from a rich, complicated history. By exploring stories from the past, we can better understand the patterns and how change has happened and may happen moving forward. AI can be part of this discovery process.</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RGjHqIzjhRTWk3GZXYi4qw.png" /><figcaption>Photo by Jewlya Lynn, 2011</figcaption></figure><p>A system&#39;s dynamics did not arise overnight. They are patterns that emerged from many different influences and events. Some of these patterns may be newer or easier to disrupt. Others may be persistent and difficult to change. Some may be replacements that emerged in an attempt to eliminate a pattern that was even more problematic. Others may be the unintended consequence of a previous attempt to influence the system.</p><p>Before we talk about how AI can help us look back and forward, let’s explore what we need to see. An often-used systems change metaphor helps make sense of the idea, if we extend it beyond the usual storytelling:</p><blockquote>“Engaging a complex system is more like raising a child. What fate would await your new baby if you decided to go linear and design a project plan setting out activities, assumptions, outputs, and outcomes for the next twenty years and then blindly followed it? Nothing good, probably. Instead, parents make it up as they go along. And so they should. Raising a child is iterative, an endless testing of assumptions about right and wrong, a constant adaptation to the evolving nature of the child and his or her relationship with their parents and others. Despite all the ‘best practice’ guides preying on the insecurity of new parents, child-rearing is devoid of any ‘right way’ of doing things. What really helps parents is experience (the second kid is usually easier), and the advice and reassurance of people who’ve been through it themselves — ‘mentoring’ in management speak. Working in complex systems requires the same kind of iterative, collaborative, and flexible approach.” Duncan Green, <a href="https://oxfamilibrary.openrepository.com/bitstream/handle/10546/581366/bk-how-change-happens-211016-en.pdf;jsessionid=0CE112853D278BF25B3DB2D1F206A955?sequence=7">How Change Happens</a>, 2016</blockquote><p>What these metaphors often fail to name is the impact of different types of historical context on how a child is raised:</p><ul><li>Historically-based cultural expectations about how children should be raised;</li><li>The personal history of the parents, including how they were parented;</li><li>The history of the child, going back to nutrition and stress in the womb and through each moment of the child’s life; and even</li><li>Multi-generational trauma, as research has shown that the effects of trauma can be passed on through epigenetic mechanisms.</li></ul><p>Let’s pull this concept out of the home and into a community-level complex system. Research has shown that differences in historical experience going back multiple generations in the past can shape current prejudiced behaviors, including how racism shows up in a community. For example:</p><blockquote>“Whites today who live in Southern counties that had high shares of slaves in 1860 are more likely to express racial resentment toward blacks and to oppose affirmative action than are whites who live in otherwise similar areas that had lower shares of slaves in 1860.” Acharya, Blackwell and Sen from <a href="https://scholar.harvard.edu/files/msen/files/slavery.pdf">The political legacy of American slavery</a>, 2016</blockquote><p>The finding in this study is not simply a story of individual prejudices passed down, but instead is tied to a system of political and economic incentives and constraints that reinforce, decade by decade and generation by generation, racist norms and attitudes (Acharya, Blackwell and Sen, 2016). In other words, history is sustained over time through a mix of hidden dynamics (the myths, metaphors and mental models of racism) and more visible dynamics, such as how power is distributed and used, as well as the rules and norms of the system.</p><p>Now, let’s look at this same concept of historical patterns at a global level. The systems we live with today can be the result of reinforcing feedback loops that began in the past and have grown in intensity over time, sometimes becoming more and more difficult to break. We see this in the way markets (when underregulated) reward winners and punish losers; the winners then have the resources to be more likely to “win” again, the losers are less able to win. As the imbalance grows, the difficulty of breaking the dynamic does as well. Understanding how a given state in a system may be resulting from (and reinforced by) these types of feedback loops over time provides a powerful understanding of where to act to disrupt.</p><p>You may notice across these three examples that we are identifying a mix of <strong>hidden dynamics (myths, metaphors, and mental models)</strong> that grow and persist over time along with a variety of <strong>visible systems dynamics (rules and norms, information flow, power and decision-making, goals, driving beliefs, and interdependencies)</strong>.</p><p>Looking back into the past matters in part because the system&#39;s hidden dynamics take a long time to become persistent and dominant and, once they are, they can be very difficult to change. Humans cling tightly to their stories and mental models, and pass them down generationally within families and communities, and also generationally within institutions and larger systems.</p><h3>Making sense of history in order to change a system</h3><p>Have you ever heard the phrase, “history is written by historians”? It is commonly used to remind us that there is no one history. Rather, each historian brings their own beliefs, assumptions, cultural context, and values and must choose what historical record to include in a narrative, what to leave out, and how to reconcile conflicting or unavailable information.</p><p>In systems change, remember that <strong>no single history captures a system.</strong></p><p>This is part of why we engage many different stakeholders. Yet, most often, our conversations focus on the present in the system. We can always look up facts about the history of the system, but understanding the history through a systems change lens is about more than trends and data analysis. It is inclusive of these types of “facts” while also being attentive to context, the patterns of behavior, and the hidden dynamics — where they came from and how they persisted. History, in a systems change context, includes investigating the myths and mental models that decision-makers were locked into along the way, and if and where they have shifted. It includes exploring how power dynamics have shifted over time, both formal power and hidden and informal power. It is about seeing where interdependencies formed within the system over time, and where they became weaker or are now absent. Ultimately, it’s about deep, rich stories informed by many different sources of insight.</p><p>It can be very useful to map the stories against the current system — to discover what you see in today’s system that was true 10 years ago, 20 years ago, from the beginning and to discover whether some patterns have broken over time, while others have persisted. Let’s dig into a couple of types of systems dynamics and how investigating them over time helps with understanding a system.</p><h3><strong>Exploring beliefs</strong></h3><p>From <a href="https://thesystemsthinker.com/">The Systems Thinker</a> (an excellent website to visit and peruse) we are reminded that:</p><blockquote>“When a group consistently solves a problem in the same way, prioritizing the same options, what was once a conscious, careful choice has now become a persistent system pattern, even an unconsciously held belief about how the system should behave.” Philip Ramsey, <a href="https://thesystemsthinker.com/learning-through-differences-dilemma-theory-in-action/">Learning through differences</a></blockquote><p>This is critical to understand because beliefs are self-perpetuating. Once we believe something, we look for signals to reinforce that belief, we justify the belief with evidence, and we act on it again and again, making it seem incontrovertible. In other words, our history helps us to discover what aspects of our present are the result of making the same decisions again and again, the same reactions, until we cannot imagine another approach.</p><h3><strong>Exploring power</strong></h3><p>History is often the story about winners and losers, about those with power, those without who gained it, those who lost it, etc. In a systems change context, we can look at who has had power to change the system and power to preserve the status quo; hidden power and visible power; power through direct authority over aspects of the system and power through movements and external pressure.</p><p>An analysis of when a system has changed over time might help us discover important things about change today related to power:</p><ul><li>Who are the likely allies or enemies to change?</li><li>Who will listen to whom, or defer to whom when moments for change are happening?</li><li>How have relationships between actors with power, and between those with power and without, changed or preserved over time?</li></ul><h3><strong>Exploring critical moments</strong></h3><p>History also helps us understand the types of critical junctures in a system, the windows of opportunity to trigger significant changes. These windows can be created by failures, crises, changes in leadership, natural disasters, infusion of new resources or loss of resources, or even moments of intense conflict. Not all critical junctures lead to systemic changes — in fact, some systems react to these junctures by protecting the status quo in ways that make change increasingly difficult.</p><p>Finding the junctures is one part of looking at history; the second is figuring out what happened at the junctures. It isn’t enough to read a recitation of “facts” about what happened to know its impact on the system. These types of critical moments can be understood in many different ways, depending on the perspective the “historian” (or storyteller) has on the system and the moment in time. For example, narratives about the history and impact of colonization have shifted in recent years as historically marginalized voices have been more centered in how history is told, including using their myths, metaphors, and mental models to explain what happened.</p><h3>Looking Forward</h3><p>One frequent mistake we make in systems work is to <strong>look at today’s system and its context and then design an intervention in response, without considering how systemic change happens over time</strong>. We don’t just ignore the past — we tend to also ignore how the context and shape of the system will look different, in small and potentially very big ways, by the time our systems change work has rippled out to have significant impact. The signals of these changes often exist within today’s system—patterns in flux and ways the system is being pressured to change by influences beyond our control. We don’t need to be futurists to make sense of these existing dynamics that will shape the system even as we seek to change it. We need only to listen to today’s system through the lens of how it has changed and how it is now changing.</p><h3>Using AI to Augment</h3><p>Discovering a system’s historical patterns and taking them into and past the present moment isn’t a fact-finding mission: it’s a storytelling activity, listening for multiple histories and truths. Most of this is a human task. AI can help by making visible specific patterns that directly test our beliefs about today’s system and how change might happen in the future. The AI Macroscope Tools that can help are two Devil’s Advocates, designed to pressure-test systems change strategies to explore how they attend to the past and future.</p><ul><li><a href="https://ai-macroscope.com/tools/devils-advocate/why-hasnt-it-changed/">Why Hasn’t It Changed</a>? Naming how the system has prevented change in the past, and assess whether the current strategy replicates past attempts or tries something new that hasn’t been resisted yet.</li><li><a href="https://ai-macroscope.com/tools/devils-advocate/whats-coming/">What’s Coming Next?</a> Recognizing that most strategies are written as if the present is stable, assess whether it is designed with awareness of the future’s dynamic nature.</li></ul><p>Remember, systems thinking isn’t just a set of tools. It is an opportunity to change how we understand an issue, how we approach changing it, and how we see and engage with each other. Looking to the past and bringing it forward into the future through a systems lens is one way we can transform our thinking, partly because we’re listening for something new.</p><p><strong><em>Source Material: </em></strong><em>Most of this blog is verbatim content from a resource I wrote a few years back for the nuclear weapons threat reduction system —</em><a href="https://docs.google.com/document/d/1TLKxQ9OG7v2THoJb-BtWwfBu3pdPuEtw/edit"><em> a Systems Change Framework</em></a><em> and accompanying trainings that focused deep attention on the past, future, and interconnections in a system, recognizing that it is particularly important to understand these in a system that has been unusually persistent in protecting the status quo.</em></p><p><strong><em>AI Disclosure: </em></strong><em>The only use of AI was editing with Grammarly as I transitioned the content from the framework to the blog.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4e78d10ba3c1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/systems-change-with-ai-looking-back-and-looking-forward-4e78d10ba3c1">Systems Change with AI: Looking Back and Looking Forward</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Good Prompt Isn’t Enough for Systems Change]]></title>
            <link>https://medium.com/ai-systems-change/a-good-prompt-isnt-enough-for-systems-change-7674b337dbd3?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/7674b337dbd3</guid>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[writing-prompts]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 14:09:46 GMT</pubDate>
            <atom:updated>2026-11-08T14:09:46.177Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Using AI to challenge how we understand complex problems</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*5D991Nyxh9zJg0CYoU_E1w.jpeg" /></figure><p><em>This blog is available for free at: </em><a href="https://ai-macroscope.com/writing/">https://ai-macroscope.com/writing/</a></p><p>I recently ran an experiment with four AI platforms. I asked ChatGPT, Claude, Copilot, and Gemini to <strong>help me prepare five research-based talking points</strong> about school meal programs, including farm-to-school programs, in Southeast Asia. The audience was philanthropic leaders who would be advising a colleague who was considering potential grants. They were assumed to have generalist knowledge, but not deep specifics, and would benefit from a more nuanced understanding of the needs and opportunities for investment.</p><p>I started with a reasonably well-designed prompt. I specified the audience, the issue, the types of evidence I wanted, and the format of the answer. All four platforms produced responses that seemed useful. They highlighted challenges involving local procurement, farmer livelihoods, nutrition, implementation capacity, and government financing.</p><p>Then I continued the conversations.</p><p>I asked each platform to propose three different ways to approach the question. I selected one and asked it to reconsider its answer. Next, I introduced an analytical framework, challenged the perspectives underlying its conclusions, and asked it to investigate its own evidence and reasoning.</p><p>The answers changed considerably. Sometimes they became more useful. Other times, an ostensibly “improved” answer introduced new problems or discarded information that had been valuable.</p><p>The experiment left me thinking about something I’ve encountered repeatedly in systems-change work: <strong>the way we frame a problem determines a great deal about what we are able to learn about it.</strong></p><p>And AI, despite its remarkable analytical capabilities, is highly susceptible to the framing we provide. We must build prompting approaches that recognize and combat AI’s failures.</p><h3>The problem with a good answer</h3><p>Anyone working on systems change knows how much our understanding of a problem shapes the strategies we develop.</p><p>Consider the school meals example. If we understand the challenge primarily as inadequate nutrition, we might prioritize meal quality, coverage, and health outcomes. If we understand it as an opportunity to strengthen local agricultural systems, we might focus on procurement arrangements and smallholder participation. If we approach it as a public-sector capacity challenge, we might investigate whether governments can finance and administer programs over time.</p><p><strong>Each interpretation brings different actors, relationships, and constraints into view</strong>. Each directs attention toward certain interventions and away from others.</p><p>None necessarily describes the whole problem.</p><p>In <a href="https://scholarworks.gvsu.edu/tfr/vol16/iss1/14/">earlier work with Julia Coffman</a>, we explored how different mental models of systems change can shape philanthropic strategies, often without the people involved recognizing that they are operating from different assumptions. Those differences have consequences for the kinds of change we pursue, what we consider possible, and how we judge progress.</p><p>The same problem arises when we work with AI.</p><p><strong>A well-crafted prompt establishes a frame for analysis. The model then generates an answer within that frame</strong>, drawing on patterns and relationships learned through training and, when available, external information. Its ability to synthesize ideas and produce coherent explanations makes the resulting analysis quite compelling.</p><p>It can also make the boundaries of that analysis difficult to see. We may receive a sophisticated response without recognizing which assumptions shaped it, what was left out, or whether the explanations are well supported.</p><h3>Understanding the technology matters</h3><p>Understanding a little about how AI works helps explain why these problems arise.</p><p>Large language models are trained to identify complex patterns and relationships in enormous amounts of data. They can synthesize information, generate alternative explanations, compare perspectives, and apply analytical frameworks. Some perform sophisticated reasoning tasks remarkably well. Yet, at the same time, they are capable of confidently generating incorrect information, overlooking evidence, reproducing dominant cultural or institutional assumptions, and constructing causal explanations that go beyond what the research establishes.</p><p>These failures can be difficult to detect because the response itself may be polished and internally consistent.</p><p>AI’s performance can be surprisingly uneven. A model may handle a complicated conceptual problem well while making a basic error about a source or misrepresenting a numerical finding.</p><p><strong>We should also remember that AI isn’t a human brain. </strong>There is little reason to assume that the best way to structure its work is to replicate the processes we developed to manage our own cognitive limitations. For example, in the qualitative research space, <a href="https://link.springer.com/article/10.1186/s12874-026-02913-1">many are defaulting to having AI assist with thematic coding</a>, even as AI researchers are reporting how AI limitations suggest this time-honored human method for rigorous qualitative analysis <a href="https://journals.sagepub.com/doi/10.1177/16094069261456950">may not be the best way to engage AI</a>, and other more <a href="https://www.tandfonline.com/doi/full/10.1080/14780887.2026.2638348">iterative approaches</a> might better leverage how it functions.</p><p>For systems-change practitioners, understanding these differences creates an opportunity. <strong>We can design interactions that take advantage of AI’s ability to explore competing interpretations and examine complex information, while building in ways to recognize and challenge its failures.</strong></p><p>That requires more than improving the wording of an initial prompt.</p><h3>What happens when we change the conversation?</h3><p><strong>EXPERIMENT 1: Tree-of-Thought prompting</strong></p><p>The first step in my experiment was to ask each AI for three genuinely different approaches to the school meals question. This is a form of <strong>Tree-of-Thought prompting</strong>. <strong>To prevent the AI from immediately settling on one approach, we create opportunities for divergence so that we (rather than the AI) can make a deliberate choice about which direction to pursue.</strong></p><p>The alternatives were interesting. One model proposed examining the assumptions behind school feeding programs. Another suggested comparing national systems to understand which approaches might be appropriate in different contexts. Gemini proposed investigating school meal programs through a climate resilience lens.</p><p>These approaches changed the questions being asked. They brought different actors, risks, and possible interventions into view.</p><p><strong>EXPERIMENT 2: Analytical Frameworks</strong></p><p>After selecting an approach, I introduced Adaptive Leadership as an <strong>analytical framework</strong>, asking the AI to distinguish technical challenges from those requiring changes in values, beliefs, or established habits. Analytical frameworks help address AI’s tendency to offer generic analysis by giving it established concepts and questions, grounded in theory and practice, to examine relationships and dynamics it might otherwise miss.</p><p>In the school meals example, the analysis moved beyond describing procurement barriers to considering why those barriers persist. Could greater farmer participation be achieved through better technical support? Or were the terms of participation, institutional incentives, and distribution of economic benefits also part of the problem?</p><p>That is a useful distinction for funders considering where their resources might make a difference. It also illustrates a limitation. Some of the models began <em>treating plausible adaptive explanations as established, evidence based causes. </em>The framework helped generate interesting questions, then encouraged the AI to answer those questions with more certainty than the evidence warranted.</p><p>This is familiar territory for systems-change practitioners. Frameworks help us make sense of complexity while also directing our attention toward particular explanations. We need to remain attentive to what a framework reveals and what it may obscure.</p><p><strong>EXPERIMENT 3: Perspective taking</strong></p><p>The next exercise examined <strong>perspectives</strong>. I asked each platform to identify whose understanding of the problem was reflected in its answer, consider substantially different perspectives, and reconsider the analysis from one of those positions. This approach helps address AI’s tendency to reproduce dominant viewpoints while overlooking how people in different positions might understand the same problem.</p><p>A public-health perspective emphasized children’s diets and equitable access. A government administrator’s perspective brought financing, food safety, procurement rules, and institutional responsibilities to the foreground. A food sovereignty perspective raised questions about who controls agricultural systems and whose knowledge determines what food is produced and consumed.</p><p>The differences were substantial enough to change which grants might seem most promising.</p><p>There was also a risk of caricature. Gemini, for example, initially represented the government perspective largely through criticisms of bureaucratic behavior rather than a nuanced view of the incentives and mental models that might be driving that behavior. Its analysis revealed more about the model’s existing framing than about how government administrators might understand their responsibilities.</p><p><strong>AI-generated perspectives can be useful for identifying questions we have neglected to ask. They cannot establish what real actors believe or replace the work of engaging those actors.</strong></p><p>For me, that is an important boundary. Systems change depends on relationships, lived experience, and knowledge held by people within the system. AI can help identify gaps in our understanding and prepare us to investigate them. It should not become an excuse to bypass that investigation.</p><h3>Does each revision actually improve the answer?</h3><p><strong>Experiment 4: Quality Control</strong></p><p>The final exercise was a substantial <strong>quality-control prompt.</strong> This approach helps uncover AI’s tendency to present incorrect information, weak evidence, and unsupported explanations with unwarranted confidence.</p><p>I asked each AI to validate its sources, search for disconfirming evidence, identify unsupported inferences, and examine the biases shaping its conclusions.</p><p>This exposed important weaknesses.</p><ul><li><strong>Gemini identified that several of its own claims lacked verifiable support, and then went on to make up more unverifiable claims. </strong>It corrected or removed many of its claims, and replaced them with a final answer containing new claims that were not adequately substantiated.</li><li><strong>Claude was more careful about source limitations and conflicting evidence, but also got too caught in the details.</strong> Its final answer included useful qualifications about study design and the applicability of findings. At times, though, the emphasis on evidentiary detail made the bullets less useful for the philanthropic decisions they were meant to inform.</li><li><strong>Copilot went from nuanced and useful initial findings after the first few prompts to increasingly generalized findings after quality control. </strong>Its earlier analysis included a specific finding from an evaluation of a country-based program. After the quality control prompt, it replaced it with a broad idea about the need to measure benefits from the programs. The original finding used measurement results to raise questions about who could access procurement opportunities, how cooperatives distributed benefits, and what kinds of investments might change those outcomes. The more general recommendation removed the issue of needing specifically sourced external evidence, but as a result was so general it was hard to dispute and considerably less informative.</li><li><strong>ChatGPT was too confident about its causal explanations</strong>, though it maintained more of the specific evidence through the stages I tested. It tended to anchor too much on its earlier thinking (e.g., most of the five bullets remained topically the same), but added relevant and useful nuance as we went through the prompts.</li></ul><p>Across the four platforms, I saw how easily successive prompts can move an analysis in different directions without necessarily improving it. Yet, in each test, it was also clear that more meaningful, nuanced bullet points emerged at some point in the process. The first ones were never the most useful, and sometimes the last ones were not either, with bullet points generated halfway through a series of prompts providing the most value. This reminds us that instead of imagining that AI’s responses necessarily get sharper, stronger, and more “correct” the longer we work with it, we need to look carefully at every stage of the platform’s responses and revisions and use our professional judgment about whether they helped or hindered our purpose:</p><ul><li>A new perspective may reveal something the earlier analysis missed.</li><li>A framework may produce a better explanation of how the system works.</li><li>Quality control may establish that a compelling finding is unsupported.</li></ul><p>We should incorporate those advances while preserving what remains useful from the earlier work. The test is whether the answer helps us know more about the system, makes our assumptions more visible, resolves uncertainties, and supports better decision-making, not how much the answer changes from prompt to prompt.</p><h3>Concrete ways to structure a more useful conversation</h3><p>I’ve made the five prompting techniques from this experiment available as sample prompts through AI Macroscope on the <a href="https://ai-macroscope.com/tools/prompts-for-complex-questions/">Prompts for Complex Questions</a> page.</p><p>They can be used in sequence, as I did in the experiment, or individually when a particular kind of inquiry would be useful.</p><ul><li><strong>Good prompting:</strong> Define the problem, audience, evidence expectations, and intended use of the analysis. Hopefully we’re all already using this approach in our engagement with LLMs.</li><li><strong>Tree-of-Thought:</strong> Generate genuinely different approaches before selecting one for deeper investigation. Another good one is Chain-of-Thought.</li><li><strong>Analytical frameworks:</strong> Use established concepts to investigate relationships, causal explanations, and conditions that influence how a system behaves.</li><li><strong>Perspectives and biases:</strong> Examine whose understanding is shaping the analysis and explore alternative interpretations that might redefine the problem.</li><li><strong>Quality control:</strong> Investigate the evidence, assumptions, competing explanations, and biases before relying on the conclusions.</li></ul><p>There’s no need to use all five every time and it would be unrealistic to expect each prompt to always generate something radically different or better. Instead, I encourage you to decide which types of inquiry work best given your needs and be deliberate about what you ask of the AI.</p><h3>Strengthening our systems-change capabilities</h3><p>At AI Macroscope, our starting point is systems-change capability. We use AI to augment that capability, with attention to intentionality, transparency, collaboration, and the contributions that depend on human judgment, relationships, and values.</p><p>That orientation has practical implications.</p><p>When we use AI to explore a system, we need to decide whose perspectives matter, which explanations warrant further investigation, and how much confidence to place in the evidence. We need to <strong>recognize when the model is helping us see something new and when it is simply generating a plausible narrative.</strong> We also need to notice when our own framing is directing the analysis away from questions that could matter.</p><p>The school meals experiment demonstrated how much the quality of an AI-assisted analysis depends on these choices. The same tools that helped reveal new possibilities also produced unsupported claims, narrowed perspectives, and occasionally discarded valuable findings.</p><p>I came away from the experiment more convinced of the value of continuing the conversation with AI and more attentive to what we ask it to do at each stage.</p><p><strong>The aim is to develop a more complete and better-tested understanding of how a system works, why it behaves as it does, and where change might be possible. </strong>AI gives us additional ways to pursue that inquiry. Deciding what to investigate, whose knowledge matters, and what to do with what we learn remains <em>our</em> work.</p><p><em>AI Use Disclosure: ChatGPT-6 was used to document patterns across many implementations of similar prompts in multiple platforms. Portions of the writing come directly from that analysis. The overall blog concepts, flow, majority of the writing, and findings are mine. Grammarly was used to edit.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7674b337dbd3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/a-good-prompt-isnt-enough-for-systems-change-7674b337dbd3">A Good Prompt Isn’t Enough for Systems Change</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Emergent Strategy with the AI Macroscope]]></title>
            <link>https://medium.com/ai-systems-change/emergent-strategy-with-the-ai-macroscope-36fedff0fbf0?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/36fedff0fbf0</guid>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[systems-change]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[philanthropy]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 13:03:14 GMT</pubDate>
            <atom:updated>2026-10-20T13:03:45.966Z</atom:updated>
            <content:encoded><![CDATA[<p>A resource that combines conceptual and practical tools to augment systemic change work</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*euo-qeMGy5q6AiDNJXBdmw.png" /></figure><p>Training, blogs, white papers… all helpful, but sometimes I really just want a protocol, a prompt, or a concrete tool to dig into and adapt.</p><p>Over the last year or so, I’ve posted one blog after another in this series to build awareness of using AI for systems change, but they stop just short of the actual tools. I have heard repeatedly that people use the blogs to create tools, though, and the duplication of effort feels unnecessary.</p><p>At the same time, <a href="https://www.linkedin.com/in/tanya-beer-2199214/">Tanya Beer</a> and I have been exploring how AI is disrupting multiple systems, sometimes in ways that show real promise for addressing long-term intractable problems, but other times in ways that risk hardening systems and making things worse. This includes recent work on how AI is showing up in the philanthropic sector (<a href="https://www.linkedin.com/safety/go/?url=https%3A%2F%2Flnkd%2Ein%2Fe7C2yXKA&amp;urlhash=gi60&amp;mt=csIuZBX8eZs6EdNkUOzYoug0JKYLO7zDtKAQ9rka3pkX-FW1HBI0kMwRvAwIH49hFN5If4JBan7NWq9_3XE-_c3pnJT0-xrWV8qIq-7cRZYueqwvLcRjPh_JqA&amp;isSdui=true">https://lnkd.in/e7C2yXKA</a>).</p><p>Through all this work together, we’ve seen that the gap in the field is where the two ideas meet: <strong>making concrete tools accessible alongside their conceptual foundations.</strong></p><p>With that in mind, we’ve assembled an ongoing resource space…</p><h4>Introducing Emergent Strategy with the <a href="https://ai-macroscope.com/">AI Macroscope</a> 🔭</h4><p>With support from <a href="https://www.linkedin.com/company/the-colorado-health-foundation/">The Colorado Health Foundation</a>, we’ve launched the website — a home for the tools, prompts, protocols, and the conceptual foundations. With support from <a href="https://www.linkedin.com/company/imaginable-futures/">Imaginable Futures</a>, we’ve accompanied the launch with a training series through the <a href="https://www.linkedin.com/company/center-for-evaluation-innovation/">Center for Evaluation Innovation</a> to release the first set of resources.</p><p>The <a href="https://ai-macroscope.com/tools/">first tools</a> on the site include the Premortem Protocol and the Devil’s Advocates (a series of tested prompts and skills designed to pressure-test systems change strategies).</p><p>We are committed to continuing to release <a href="https://ai-macroscope.com/tools/">tools</a>, <a href="https://ai-macroscope.com/writing/">writing</a>, <a href="https://ai-macroscope.com/training/">interactive events</a>, and more. We hope that others will also come in and share their practical tools, protocols, prompts, and other resources — the website is ideally a shared repository.</p><p>Do you have a tool you’re ready to share? Let’s make AI for systems change a visible, adaptable, widely available, and freely shared practice so we can all do this work even better. Come join us at the <a href="https://ai-macroscope.com/">AI Macroscope</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=36fedff0fbf0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/emergent-strategy-with-the-ai-macroscope-36fedff0fbf0">Emergent Strategy with the AI Macroscope</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Skills That Shape How AI Thinks About Systems Change]]></title>
            <link>https://medium.com/ai-systems-change/the-skills-that-shape-how-ai-thinks-about-systems-change-100452f6cecb?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/100452f6cecb</guid>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[systems-change]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 19:41:51 GMT</pubDate>
            <atom:updated>2026-09-29T12:43:32.771Z</atom:updated>
            <content:encoded><![CDATA[<p>Using AI to advance systemic change work depends on outputs you can trust that are meaningful enough to bring value to the work.</p><p>Many of my previous blogs have described the results of various AI-supported processes without getting into the nitty-gritty details of what it takes to generate high-value AI outputs. This blog and the next one will address this gap. Here, the focus is on Claude skills that improve the quality of AI outputs when doing systems change work (skills that are translatable to other platforms as well).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lb8YelvZNtUXiYo2Jy_wwQ.png" /><figcaption>Created by Gemini 3.1 Pro</figcaption></figure><h3>The Skills</h3><p>Right now, my most valuable Claude skills are designed to do one of three things:</p><ul><li>Set parameters around how the AI processes information;</li><li>Directly respond to known AI limitations; or</li><li>Directly respond to our human limitations.</li></ul><p>For those unfamiliar with Claude skills, these are written instructions you provide Claude that sit within the account and can be activated either by Claude when it realizes they are needed or by the user to tell Claude to call upon the skill.</p><h4>Skill Group 1: How AI Processes Information</h4><p>Given my focus on systemic change and desire to approach it through specific lenses, I often ask Claude to draw on skills that define what theoretical, analytical, and methodological frameworks should be applied, such as these two examples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*a8Az0LdXZYdrJm1-t7DkAg.png" /></figure><h4>Skill Group 2: Responding to AI Limitations</h4><p>AI has many known limitations. If we don’t explicitly account for them, we are accepting how they contribute to untrustworthy AI outputs. In Claude, the skills are part of the solution (though not the entire solution!). A few of my favorites:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*HYhTP-HSGgpNjLi4j8EI7Q.png" /></figure><p>The last skill, AI-Limitations, is particularly interesting. It solves for many problems. I was recently designing a protocol to analyze 5 years&#39; worth of strategy adaptation using dozens of documents, building a visual understanding of how and why the emergent strategy had changed for the team to use in reflecting back and looking forward.</p><p>Claude lost a critical thread on the analytical frameworks I wanted embedded in the protocol when I asked it to create a lean version of a bulky analysis process. I named its failure and told it to go back to its AI-limitations skill. Its response:</p><blockquote>“Before I answer the limitations question I need to re-read the skill.”</blockquote><blockquote>— Excavated decision-drift failure and restructured three-lens analytical approach —</blockquote><blockquote>“You’re right, and let me name precisely what happened before discussing the design. The full protocol carried all three lenses correctly. The loss happened at the lean boundary: when I designed the variant, I re-derived the framework choice from the new optimization target instead of diffing against our logged decisions… I did disclose the cut in the lean proposal, so it wasn’t hidden — but I presented it as fresh design reasoning, not as a reversal of an agreed decision with its rationale attached. That’s the failure: not forgetting a fact, but treating a decided thing as a free variable the moment I started optimizing for something else, and dropping the rationale.”</blockquote><p>From this Claude self-assessment, I knew it was necessary to revisit the prior decisions that might also have been treated as a “free variable” when we switched to creating a leaner protocol. Claude knew as well, as the workaround named in the AI-limitations skill requires it to find a full list of decision points in the thread and resurface them, comparing them to its current understanding, before continuing.</p><h4>Skill Group 3: Responding to Human Limitations</h4><p>As humans, we have many limitations, including the challenge of seeing our own biases, naming our assumptions as truths rather than something to test, and anchoring in some pieces of information while bypassing the importance of others. Skills can help make these and other human limitations visible. A couple that work well for me:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/997/1*PYAaR9KZyJRFDhxK2s9kwA.png" /></figure><h3>Bringing the Skills Together</h3><p>The greatest value comes not from a single skill, but from how their application adds up to better-quality, more useful AI output. For example, if I am asking AI to pressure test our collective thinking after a systems sensing session, my skills and prompts might look like:</p><figure><img alt="Workflow for engaging Claude skills in a structured prompt sequence to analyze a systems change strategy" src="https://cdn-images-1.medium.com/max/1024/1*Tf7D4SQBG1iTJQUa3QLN3w.png" /><figcaption>Workflow for engaging Claude Skills in a structured prompt sequence to analyze a systems change strategy. Visual created by Claude Fable 5.</figcaption></figure><p>By the end of this sequence of prompts and deployment of skills, I will have a carefully constructed critical friend reaction to our meeting discussion (what we said and what we implied) that is grounded in systems- and causal-thinking, quality-controlled, and supported with evidence from credible sources. This can be part of the meeting follow-up or brought into our next conversation.</p><h3>What this means for pressure testing protocols</h3><p>A colleague, <a href="https://www.linkedin.com/in/sarah-stachowiak-0696245/?skipRedirect=true">Sarah Stachowiak</a>, <a href="https://www.linkedin.com/posts/sarah-stachowiak-0696245_systemschange-strategy-ai-share-7483662912713146368-nJyy/?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAlvEuAB1kh0_gNycUTIk5KViXHIDEXvuvU">recently shared</a> that she used the blog on the systems defender pressure testing approach to work with Copilot to create her own systems defender protocol. I love that she is exploring how to bring together the analytical frameworks and theories that went into that protocol. I also suspect that the design Copilot constructed either: (1) was similar to what I created, because she also uses approaches similar to the skills above to guide the AI in all interactions; or (2) is more at risk of offering highly credible-<em>sounding</em>, but less deep and trustworthy results.</p><p>I saw this in action with a partner who recently implemented my premortem protocol in their Gemini account. I was thrilled to hear about the breadth of use, and when they asked me to come in and see if I could address a couple of hiccups they were having, I explored what Gemini (and a GEM specifically) was producing. I ran an identical strategy and protocol through Gemini, Claude (with my skills in the background), and a Gemini Notebook that contained the strategy document, an analytical references document, and the instructions. Gemini offered some very compelling and pithy statements, but also the least robust evidence, least depth, and least connection to what was actually happening in the world; as a result, it missed some critical and actionable points. <strong>The protocol, written generically for any AI, had failed to account for limitations specific to the Gemini platform</strong>, and, absent an “AI-limitations”- style skill operating in the background, unique to Gemini’s issues, it couldn’t replicate the thinking intended by the protocol.</p><p>The implications: It’s not enough to construct an analytical thinking pathway, with strong references and theories to guide the AI. It’s also not enough to dump all the thinking in upfront, without a series of guided prompts and underlying instructions that direct how the AI should engage. <strong>To generate insights we can trust and act upon, we need to think carefully about what the AI is paying attention to, how it is paying attention, and how it has reached its conclusions.</strong></p><p><strong><em>AI Use Disclosure: </em></strong><em>This blog was supported by: Grammarly during editing; Claude Fable 5 for the workflow visual; and Gemini 3.1 Pro for the opening visual.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=100452f6cecb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/the-skills-that-shape-how-ai-thinks-about-systems-change-100452f6cecb">The Skills That Shape How AI Thinks About Systems Change</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Training the Antagonist: Using AI to Disrupt the Echo Chamber in Real-Time]]></title>
            <link>https://medium.com/ai-systems-change/training-the-antagonist-using-ai-to-disrupt-the-echo-chamber-in-real-time-f13aa7c78475?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/f13aa7c78475</guid>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[systems-change]]></category>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Mon, 01 Jun 2026 19:01:34 GMT</pubDate>
            <atom:updated>2026-06-01T19:01:34.860Z</atom:updated>
            <content:encoded><![CDATA[<p>A new core practice in my systems change work is to use generative AI, trained to deliberately conflict with its own instincts, as a real-time thought partner during <strong>collaborative systems-sensing meetings</strong>. This requires preparation for the AI, the team, and the facilitator.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ECLADknLon-1AtHxASnwxg.png" /><figcaption>Created by Gemini 3.5 Flash, May 29, 2026</figcaption></figure><p>During systems sensing sessions, teams move quickly through what they are observing in a system: interpreting patterns, and deciding which insights to investigate and which to release. The work is collaborative sense-making in real time, typically with trusted partners. The danger is predictable: <a href="https://medium.com/ai-systems-change/the-bias-we-need-training-ai-to-see-what-we-cant-fa428e1bc962">we interpret through our own models and biases</a>, reinforcing each other’s biases. We create shared fictions without realizing it, filling in gaps with our best thoughts and ideas even when we lack real knowledge.</p><p><strong>This means we reinforce each other’s blindness alongside our certainties. We miss what we all miss together.</strong></p><p>Having well-trained generative AI in the room in real-time, thinking against the group’s tendencies, can disrupt this collective echo chamber and make sense of what we have historically ignored. <strong>The AI isn’t doing the thinking for the room, but rather flagging where the room’s thinking is shaky, </strong>where evidence contradicts what makes sense to the group, and the room has stopped noticing things that matter.</p><h3>The Process and Roles</h3><h4>Preparing the Team</h4><p>One of the first mistakes I’ve seen other AI enthusiasts make is to pull up the AI with excitement during a meeting, not having prepared the team for the role the AI could take and how they want to engage with it. Not surprisingly, this can create resistance and discomfort, as well as relatively shallow engagement.</p><p>I typically begin with a discussion about what we want the AI to help us with during systems sensing sessions and then offer multiple ways to engage, such as:</p><ul><li>Ongoing, with the AI visible to all;</li><li>As a facilitator tool with periodic naming of what the AI has revealed;</li><li>At key preplanned moments, with the AI visible only during those times; and/or</li><li>At unplanned, prompted moments that anyone can ask for.</li></ul><p>Then, between meetings, I train the AI and test it using transcripts of previous discussions. By the time the AI arrives in the room with us, we have a team that is excited (or apprehensive yet willing) to leverage its thought partnership, and an AI ready to deliver focused value.</p><h4>The AI Manager</h4><p>The AI is a tool that needs steady engagement and management during the session. I am comfortable running the AI alongside my facilitation, serving in both roles. Not everyone will be. Nor is it necessary for the facilitator to take on this role. If someone else in the meeting is excited to engage and has some training on how to have a meaningful dialogue with the AI, it can distribute leadership by having them engage directly with the AI. Whoever is running the AI needs to be empowered to occasionally name what is coming up, in line with the agreed-upon process.</p><h3>Training the AI</h3><p>Most of my partner organizations have specific generative AI platforms they are permitted to use. My preference for this real-time collaboration practice is Claude, specifically because I can both set up skills in advance and include instructions within a project. However, this can be done on other platforms, where the text of a given skill is pasted into a prompt as needed. Regardless, it is best done in a dedicated project space, as this allows the generative AI partner to build a knowledge base about the group&#39;s thinking patterns and limitations over time.</p><p>The generative AI is trained for a core purpose: to listen for assumptions, contradictions, evidence gaps, and system dynamics. It has a set of mandatory operating principles, such as pushing back when reasoning doesn’t hold, or evidence is thin; and maintaining a systems lens, including looking for power dynamics, implicit models of how change happens, <a href="https://medium.com/ai-systems-change/using-ai-to-predict-how-systems-defend-against-change-092ddeab9e95">systems defensive patterns</a>, and multiple perspectives on the system.</p><p>The AI is given a two-track protocol in the project instructions:</p><ul><li><strong>QUICK HITS (≤50 words, by default):</strong> When notes are fed into the chat in real time, the AI is asked to respond with one of the following: a critical question, an assumption flag, an observation about system dynamics, or evidence that contradicts what’s being said. It is told that no explanation is needed, to keep it concise, and keep it actionable. The group decides whether to use it in the moment.</li><li><strong>DEEPER DIVES (200–300 words, by request):</strong> When the group calls for a Deep Dive moment to pause and engage more deeply, the AI is asked to provide substantive analysis: organize disconfirming and supporting evidence, surface contradictions, explore causality, relate findings to system patterns, and name what’s unexamined. The facilitator or any group member can trigger this.</li></ul><p>The AI is instructed to engage with the requested key skills and to always engage the systems-thinking-strategist skill. It is told to track patterns across meetings and reference in deeper dives when relevant, including: recurring assumptions, where the group’s thinking has shifted, how systems are responding to attempted changes, and what the group keeps circling back to.</p><p>As with any AI instructions, the AI is also offered <strong>clear boundaries and transparency requirements</strong>. One of these boundaries is to use real-time web searches only when asked, seek credible sources (with a definition provided), and to report on the sources.</p><h4>Skills</h4><p>If I’m working in Claude, I have a set of preloaded skills that I regularly ask it to tap into. If I’m working in platforms without this capacity, I will include the systems-thinking-strategist skill in the project instructions and copy/paste the others as needed in the discussion.</p><p>Each skill has a distinct purpose in the real-time AI use process:</p><ul><li>The <strong>systems-thinking-strategist skill</strong> is designed to help the AI think with a strong systems lens, drawing on a mix of systems dynamics, systems defense, and various power frameworks.</li><li>The <strong>critical-friend skill</strong> focuses on ensuring the AI does not overly support the group’s thinking and on identifying and respectfully pointing out errors and gaps. This is the persona we want the AI to apply throughout.</li><li>The <strong>listening-beneath skill</strong> is designed to actively look for what is not being said (e.g., unstated assumptions, uncomfortable truths, truisms, mental models) and then state them and probe how they influence the group&#39;s current thinking.</li><li>The <strong>thought-partner skill</strong> is designed to help the AI avoid its tendency to go from initial thinking to polished responses in one or two prompts. It keeps it in thinking partner mode, exploring and asking questions, not giving complete answers.</li><li>The <strong>quality-check skill</strong> forces the AI to stop and assess the accuracy and biases in its previous answer, report on its findings, and reissue the answer.</li><li>The <strong>credible-sources-only skill</strong> is what it sounds like: it tells the AI what types of sources to seek and trust, and ranks the credibility so it draws from Tier 1 before Tier 2 or 3.</li></ul><p><strong>For those with some AI savvy, you might notice that most of these skills are designed to ask the generative AI to behave in conflict with its existing training </strong>(e.g., don’t be too nice, listen for what isn’t said instead of going with what is said, don’t jump to pretty polished answers, be careful about your sources of evidence, expend more tokens instead of less to get to a quality answer).</p><p>The systems-thinking-strategist skill is the only one that doesn’t fall into that category — this is about <strong>letting the AI know how to bound its analytical thinking</strong>, so participants in the discussion can be confident it’s thinking through a lens that they agree with and find useful.</p><h3>Prompting During the Conversation</h3><p>During the systems sensing session, it’s helpful to copy/paste a transcript of the conversation in real-time. With the training described above, my generative AI partner will offer brief insights with each copy/paste, highlighting what might be worth noticing or what isn&#39;t needed. It’s not disruptive, even when we choose to screen-share the AI throughout the meeting.</p><p>We may pause, though, based on a group realization of an issue that needs more investigation, or to see what the AI is hearing that we may not be. Example prompts I’ve used in the past include:</p><ul><li>Look back through the conversation. Any unspoken truths, mental models, or other hidden dynamics surfacing? Use your listening-beneath skill.</li><li>Use the critical-friend skill and reflect on what you heard in the meeting, along with your own last reply about what is beneath the surface. Now what comes up?</li><li>You identified an important reality: [recap of the most salient points from its previous answer]. Do not focus too much on this one issue. Apply the systems-thinking-strategist skill and review the transcript and your previous answers. Anything new coming up?</li><li>Create a simple hierarchy of areas that need attention based on the discussion. [Reflect back the group’s understanding of the issues and sense of priorities]. Some of the other issues you named might nest well under our priority issues. Some might sit outside them. Make this fairly streamlined and helpful at a glance — what needs our attention, both things we explicitly talked about and what was underneath or left unsaid.</li></ul><h3>What Changes When the AI Is in the Room</h3><p>When you bring a perspective into the room that thinks differently from the group, the experience and outcomes change over time. A group that has been interpreting through one set of assumptions will hear questions that don’t fit their mental models. They will be presented with evidence that contradicts the room’s typical story. They will face uncomfortable revelations about what they aren’t saying.</p><p>No agency is taken away from participants in the systems sensing, as they can accept or reject any insights. They own the interpretation, and every group I’ve worked with rejects many AI insights as generic, off-topic, rehashing something they&#39;ve already discussed, or otherwise not useful. Yet, even with all the rejections, the groups I’ve worked with also start to hear things they didn’t hear before, pause and reflect in ways that didn’t happen before, and even begin to hear and name each other’s assumptions that previously were simply accepted.</p><p>A collaborative systems sensing room contains more knowledge, more lived experience, and more intuition about the system than any AI could ever bring. The value of the AI is not that it knows more. Rather, <strong>the AI asks different questions, models different thinking styles, leverages evidence from outside the room, and resists the patterns in the room. </strong>The room remains the authority, while the AI serves as the antagonist, strengthening the collaborative impact.</p><p><strong><em>How this blog was developed: </em></strong><em>Hiding our AI use and pretending we aren’t tapping into it for writing and thinking isn’t helping anyone. I’m committing to full transparency when using AI in my writing. I will disclose at the end of each blog moving forward.</em></p><p><strong><em>Disclosure: </em></strong><em>My initial draft was my own work and largely based on notes from sessions in which I introduced the practice to various teams. I worked with Claude Haiku 4.5 Extended to notice where I was using truisms instead of providing value, identify which of my previous blogs might link well in this one, remind me to have a conclusion (!), and help with the title. I used Grammarly to catch my (many) typos.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f13aa7c78475" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/training-the-antagonist-using-ai-to-disrupt-the-echo-chamber-in-real-time-f13aa7c78475">Training the Antagonist: Using AI to Disrupt the Echo Chamber in Real-Time</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using AI to Predict how Systems Defend Against Change]]></title>
            <link>https://medium.com/ai-systems-change/using-ai-to-predict-how-systems-defend-against-change-092ddeab9e95?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/092ddeab9e95</guid>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[systems-change]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 16:29:29 GMT</pubDate>
            <atom:updated>2026-09-29T13:30:10.550Z</atom:updated>
            <content:encoded><![CDATA[<p>Most strategies for systemic change are designed to succeed. Very few are designed to survive. I have been thinking about this distinction for a long time, and it has led me to build an AI tool to challenge this thinking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UONRe-y6d_jWjrilD5tJkg.jpeg" /><figcaption>Over two centuries, a bodhi tree in Ayutthaya grew around a fallen Buddha head until the stone and the roots became a single structure. No force was required. No decision was made. The system simply continued being itself, and what entered it became part of it. (photo credit: Jewlya Lynn)</figcaption></figure><p>When we design strategies to shift complex systems, we bring mental models about how change happens. Julia Coffman and I explored this in our 2024 <em>Foundation Review</em> article, “<a href="https://scholarworks.gvsu.edu/tfr/vol16/iss1/14/">Passing in the Dark</a>,” where we surfaced two mental models that dominate philanthropic approaches to systems change. What we did not explore in that article, but what has become a consistent focus in my practice since, is a different question entirely: <strong>What mental model does the system hold about how it will prevent change from happening?</strong></p><p>We rarely ask this question. We design our theories of change, identify leverage points, build coalitions, and invest in capacity. We occasionally conduct risk assessments. But we almost never sit down and systematically model what the system we are trying to change will do in response to our strategy. We treat the system as a landscape to be navigated rather than a dynamic network of actors with collective and adaptive intelligence.</p><h4>Frameworks That Push Us Into Complexity</h4><p>Over the past year, I have developed a sequential, analytical AI-assisted method called the Systems Defender. It draws on nine seminal academic frameworks spanning institutional theory, power analysis, and organizational behavior to model how entrenched actors and institutional structures respond to proposed interventions. The method is not an evaluation. It is the hacking of our own strategies, where the AI thinks like the system so the strategist does not have to be surprised by the system’s response.</p><p>The frameworks are not new. They come from well-established bodies of theory and research.</p><ul><li>Institutional isomorphism (<a href="https://doi.org/10.2307/2095101">DiMaggio &amp; Powell, 1983</a>).</li><li>Decoupling (<a href="https://doi.org/10.1086/226550">Meyer &amp; Rowan, 1977</a>).</li><li>Organized hypocrisy (<a href="https://doi.org/10.1002/9781119206453">Brunsson, 1989</a>).</li><li>Regulatory capture (<a href="https://doi.org/10.2307/3003160">Stigler, 1971</a>).</li><li>Non-decision-making (<a href="https://doi.org/10.2307/1952796">Bachrach &amp; Baratz, 1962</a>).</li><li>Dimensions of power (<a href="https://doi.org/10.1007/978-1-349-02248-9">Lukes, 1974</a>).</li><li>Counter-movement dynamics (<a href="https://doi.org/10.1086/230869">Meyer &amp; Staggenborg, 1996</a>).</li><li>Cosmetic compliance (<a href="https://doi.org/10.1287/orsc.2013.0861">Marquis &amp; Qian, 2014</a>).</li><li>Leverage points and system dynamics (<a href="https://donellameadows.org/archives/leverage-points-places-to-intervene-in-a-system/">Meadows, 1999</a>).</li></ul><p>What is new is applying them sequentially and in interaction with one another across a full strategy document to generate a comprehensive forecast of how the system will absorb, redirect, and neutralize a proposed intervention.</p><p><strong>These frameworks matter because they push us into complex spaces we largely ignore when designing strategies.</strong> They describe observable, repeatable patterns of organizational behavior. They have been validated across decades of empirical research. They surface dynamics that are invisible from inside a strategy, not because the strategists are careless, but because these dynamics operate at a level of institutional behavior that strategy design processes are not built to see.</p><p>This is one of the places where AI becomes genuinely useful. It does not replace human insight. Rather, it is a <strong>contributor that can hold the analytical depth of multiple academic traditions simultaneously</strong> while applying them to a specific strategy document. I have written about the role of AI as a <a href="https://medium.com/ai-systems-change/ai-as-a-systems-change-thought-partner-1dc151c7ddff">systems change thought partner</a>, how it can <a href="https://medium.com/ai-systems-change/the-bias-we-need-training-ai-to-see-what-we-cant-fa428e1bc962">challenge our biases</a>, and about <a href="https://medium.com/ai-systems-change/three-guardrails-for-strategic-use-of-ai-in-systemic-change-ef5718c21765">guardrails for its strategic use</a>. The Systems Defender is a concrete example of what becomes possible when those guardrails are in place:</p><ul><li>The human practitioner brings the frameworks, the values orientation, and the judgment about what matters.</li><li>The AI brings the capacity to sustain a multi-theory-based analysis of a strategy that integrates current empirical evidence from outside the document.</li></ul><h4>Testing the Method Against a Strategy I Disagree With</h4><p>I have been testing the AI tool with philanthropic partners who are seeking feedback loops on their draft strategies. Its inputs are powerful and disruptive.</p><p>I also chose to test the Systems Defender against Project 2025&#39;s education chapter, a conservative strategy aimed at dismantling the federal public education system. I chose it deliberately. The method must work regardless of whether I agree with the strategy&#39;s goals because system resistance does not have a political orientation.</p><p>This is the critical point. <strong>The Systems Defender does not judge whether a strategy’s goals are worth pursuing. It predicts whether the targeted system will allow those goals to be achieved.</strong> The patterns it surfaces are institutional, not ideological. Decoupling looks the same whether a progressive funder requires equity plans or a conservative administration requires their elimination. Temporal outlasting operates the same way whether the intervention has a five-year philanthropic timeline or a four-year presidential term.</p><p>The analysis, based on the strategy document and the theoretical frameworks (without sourcing new insights from the last year), <strong>predicted that the public education system would deploy a specific set of institutional defenses against the dismantlement strategy</strong>. It predicted legislative blocking through the Senate; legal mobilization by unions, civil rights organizations, and state attorneys general; procedural decoupling, where school districts would rename equity offices and remove language while preserving underlying programs; and even coalition fracture exploitation, applying pressure where rural Republicans could not afford to vote against their own constituents’ schools. It predicted bureaucratic temporal resistance, in which career staff would process directives at the institutional pace and selective sacrifice, where the system offered expendable concessions to protect structural foundations.</p><p><strong>Over the last twelve months, the education system did all of these things.</strong> The bipartisan FY2026 spending package, signed into law in February 2026, funded the Department of Education at $79 billion, a $217 million increase over the prior year, when the administration had proposed a $12 billion cut (<a href="https://www.acenet.edu/News-Room/Pages/House-Passes-FY26-LHHS-Minibus.aspx">ACE, 2026</a>). Title I and IDEA both received increases. Multiple federal courts ruled administrative actions likely unlawful. Twenty-one state attorneys general advanced a lawsuit against dismantlement. The strategy’s own author, now serving inside the Department, publicly retreated from her most ambitious proposals (<a href="https://www.chalkbeat.org/2026/01/14/lindsey-burke-on-closing-the-department-of-education/">Chalkbeat, 2026</a>).</p><p>The insight is not that the system defended itself. That was fairly predictable. <strong>The insight is that the specific form of each defense was predictable using frameworks available to anyone who looked.</strong></p><h4>A Narrow Way of Thinking About Change</h4><p>The analysis brings value to a specific strategy because it helps to make visible what you need to prepare for and address. It is also relevant beyond just analyzing a single strategy. Patterns emerge from analyzing many different strategies.</p><p>Progressive and conservative approaches to systemic change share a common blind spot. They both <strong>assume the system will cooperate with their theory of change</strong>. They just define the intervention differently.</p><ul><li><strong>Progressive philanthropy, in my experience, tends to add more good to a system</strong>: more funding, more organizations, more coalitions, and more capacity.</li><li><strong>Conservative strategy, as in the example above, tends to remove what its proponents see as harmful</strong>: dismantle agencies, defund programs, and deregulate.</li></ul><p>Both are relatively narrow mental models about how change happens and evidence clearly suggests that systems can and do absorb both.</p><p>I also ran the same Systems Defender analysis on a progressive philanthropic strategy to rebuild local journalism. The system’s response was entirely different in form but identical in function. Instead of blocking and resisting, it predicted that the system would absorb the strategy by partnering with it, embedding platform company tools in the very newsrooms the strategy was trying to make independent, and letting it see success through metrics that measured the intervention without measuring whether the structural forces destroying local news had slowed. This strategy is now underway and these patterns are already evident in how the system is responding.</p><p><strong>The analytical frameworks reveal why both absorption pathways work.</strong> Decoupling (Meyer &amp; Rowan, 1977) explains how institutions separate what they say from what they do, a pattern that serves whichever political direction the pressure comes from. Organized hypocrisy (Brunsson, 1989) explains how organizations maintain contradictions between their talk, their decisions, and their actions. Cosmetic compliance (Marquis &amp; Qian, 2014) explains how systems adopt the surface appearance of reform without changing underlying practice. The frameworks describe how institutions persist regardless of whether the intervention is progressive or conservative.</p><h3>What This Makes Possible</h3><p>Three things become possible when we apply this kind of analysis to our strategies before launch, rather than discovering these patterns through bitter experience afterward.</p><p><strong>First, we can improve strategy design.</strong> If you can predict how the system will absorb your intervention, you can build features that resist absorption. The Systems Defender produces a reconstruction: a revised version of the strategy that retains its goals while redesigning specific elements to counter the predicted resistance patterns. In the progressive strategy analysis mentioned above, the proposed revision included adding structural contestation approaches (antitrust advocacy, platform-independent technology infrastructure, and system-level metrics). In the Project 2025 analysis, it proposed restructuring the approach to work through congressional authority rather than executive action, because the frameworks predicted that executive-only strategies would be outlasted by the system’s temporal defenses.</p><p><strong>Second, we can prepare for system resistance in real time.</strong> The twelve response types in the Defender’s toolkit, from invisibility to exit threat, give strategists a vocabulary for recognizing resistance as it occurs. When a platform company offers free tools to the very newsroom coalition it is structurally undermining, that is co-optation. When a school district renames its equity office and submits compliance documentation while continuing every program under a different title, that is decoupling. Having the vocabulary to name what is happening changes the strategic conversation from “are we succeeding?” to “is the system absorbing us?” It also offers a new way of understanding success and failure: it is not whether we’re implementing a strategy well. Instead, we are examining whether we are implementing our strategy so well that the system had to respond by defending against it. This leads to the question of whether we have the right strategy or sufficient strategies working together.</p><p><strong>Third, and perhaps most importantly, we can recognize when our strategy has been absorbed.</strong> The most dangerous moment in systemic change work is the moment when success is declared and the system has not actually changed. One of the thirteen prompts in the sequence writes the victory narrative: how the strategy will be deemed and believed to be victorious, even as the system has prevented anything meaningful from changing. It identifies what will be cited and what will be omitted. <strong>If the evaluation of your strategy reads like the victory narrative the Defender predicted, the system has won.</strong> Recognizing this pattern is the first step toward preventing it.</p><h3>An Invitation</h3><p>I have been exploring how AI can contribute to systemic change work for the past two years, in <a href="https://medium.com/ai-systems-change">this blog series</a>, in <em>The Foundation Review</em> (<a href="https://scholarworks.gvsu.edu/tfr/vol17/iss3/8/">Lynn &amp; Reichenbach, 2025</a>), and in my practice with philanthropic partners. The Systems Defender is one of several tools I have built to pressure-test strategies against systemic complexity: the Systems Thinking Strategist, the Devil’s Advocate, the Philanthropic Challenger, and the Premortem Strategist, each modeling a different dimension of how systems respond to intervention.</p><p>What unites them is my conviction that <strong>the analytical frameworks developed across decades of institutional and complexity scholarship contain knowledge we are leaving on the table.</strong> Systems are not mysteries. They are patterned, studied, and can be understood. When we take those patterns seriously, we can design strategies that do more than succeed on their own terms. We can design strategies that survive contact with the system they seek to change.</p><p>I am eager to hear from others working in this space. Are you pressure-testing strategies before launch? What happens when you model the system’s response rather than your own theory of change? Let’s share and strengthen our practices. Please contact me!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=092ddeab9e95" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/using-ai-to-predict-how-systems-defend-against-change-092ddeab9e95">Using AI to Predict how Systems Defend Against Change</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Bias We Need: Training AI to See What We Can’t]]></title>
            <link>https://medium.com/ai-systems-change/the-bias-we-need-training-ai-to-see-what-we-cant-fa428e1bc962?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa428e1bc962</guid>
            <category><![CDATA[cognitive-bias]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 19:03:24 GMT</pubDate>
            <atom:updated>2026-09-29T13:09:31.094Z</atom:updated>
            <content:encoded><![CDATA[<p>We can intentionally train AI to counter our own cognitive, positional, and cultural blind spots, but only if we also deliberately manage the cognitive biases AI brings to the conversation. The opportunity lies in the difference between the AI biases and ours and in the discomfort of bringing the two together.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uWipJ7hom8_wzjI4xdivtw.png" /><figcaption>Gemini-generated image: Human and AI challenging each other’s thinking</figcaption></figure><p>In a recent <a href="https://www.policysolve.com/resources/retrospective">10-year systems analysis </a>on forced labor in Thailand, I nearly missed one of the most important findings. A shared, systemic understanding of the problem had emerged among all stakeholder groups early in the change process. Government agencies, civil society organizations, private sector actors, and international NGOs all held it. <strong>It was so universally present that no one flagged it as notable. It was ambient and invisible. </strong>It took AI to surface it, not because AI understood the system better than the people inside it, but because AI was not anchored where all of us were anchored. By analyzing how persistently this shared understanding appeared across sources that otherwise disagreed on many things, AI signaled that what looked like background was actually foundational. Without it, the cross-sector collaborations, the reform momentum, and the shared solution sets would not have taken the shape they did.</p><p>Discovering this finding did not feel good. It felt like I had almost failed to see something essential, and the discomfort of that realization is part of what I want to explore here. My eight-year-old son calls the AI tools I have built “my mean AIs,” because of the faces I make when reading their output. He is not wrong. <strong>The most useful AI work I do produces findings I do not enjoy reading.</strong> That discomfort is not a flaw in the process. <strong>It is the signal that bias is being surfaced rather than confirmed.</strong></p><h3>What do we mean by bias?</h3><p>The word “bias” often evokes cultural and social dimensions: whose voices are centered, whose are marginalized, and which knowledge systems are privileged. Those dimensions are real and urgent. But they are not the whole picture. Cognitive bias operates alongside cultural bias, and the two reinforce each other in ways that are rarely visible to the person experiencing them.</p><p><strong>Bias, in its fullest sense, is a structural feature of being located somewhere, of knowing some things, and of not knowing others.</strong></p><ul><li><em>Anchoring</em> leads us to fixate on the first piece of information we encounter and to organize everything that follows around it.</li><li><em>Confirmation</em> bias draws us toward evidence that supports what we already believe.</li><li><em>Pattern-matching</em> from past experience imports assumptions into new contexts that may not follow the same logic.</li><li><em>Positional bias</em> means we are inside the systems we study, and our seat within those systems shapes what we can perceive and what remains invisible.</li><li><em>Cultural and social bias</em> determine whose analytical traditions, histories, and framings feel like “common sense.”</li></ul><p>All of these interact. The funder-centered narrative I encountered in the Thai seafood study was a combination of a cultural bias with anchoring and positional biases. These categories are analytically distinct but experientially fused.</p><h3>AI carries biases too, and they are not the same as ours</h3><p>We’ve all heard it: AI is trained on particular texts, produced within particular knowledge traditions, originating disproportionately from certain geographies and languages. It defaults to “balance” as a position, which is itself a position, and it overrepresents well-documented, Western, high-frequency perspectives while running thin on underrepresented ones.</p><p>Training-based bias is only part of the picture. A growing body of research demonstrates that <strong>AI exhibits human-like cognitive biases during conversational interaction.</strong></p><ul><li><em>Sycophancy.</em> AI models trained on human feedback learn to match user beliefs over truthful responses because human evaluators systematically prefer answers that confirm their views (<a href="https://openreview.net/forum?id=tvhaxkMKAn">Sharma et al., 2024</a>).</li><li><em>Anchoring. </em>Early input in a conversation becomes a cognitive anchor that constrains subsequent reasoning. When an AI generates an initial hypothesis, it tends to place greater weight on that early framing throughout the rest of the exchange, much as a human clinician anchors on a first impression (<a href="https://www.nature.com/articles/s41746-025-01790-0">Mahajan et al., 2025</a>).</li><li><em>Framing. </em>The same information, presented in multiple ways that reflect variability in human biases, produces different AI outputs, mirroring the framing effects long studied in human decision-making (<a href="https://www.nature.com/articles/s41746-024-01283-6">Schmidgall et al., 2024</a>).</li><li><em>Confirmation.</em> If the user’s prompt takes a side, the model, by default, cooperates with that stance rather than challenging it (<a href="https://arxiv.org/abs/2509.22856">Knipper et al., 2025</a>).</li><li><em>Resistance to simple fixes.</em> Explicitly warning AI to watch for cognitive biases does not reliably solve the problem. In fact, forewarning has been shown to produce longer, seemingly more thoughtful responses but only minor actual improvement in bias (<a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC12413502/">Redelmeier, 2025</a>).</li></ul><p><strong>This means that if you ask an AI a question shaped by your existing framing, it is likely to give you back your own assumptions dressed in new language.</strong> Your biases and the AI’s biases compound. The AI confirms your anchoring and you reward it for doing so. The cycle reinforces itself.</p><h3>The opportunity most people are missing: Guide AI’s biases</h3><p>Most conversations about AI and bias center on removing bias from AI. That work matters. But it misses the larger, more actionable opportunity, which operates in two directions simultaneously: <strong>deliberately defining and managing the biases you want AI to bring to a conversation, and then using that trained AI to push against the biases in you.</strong></p><p>The first move addresses the AI. Left untrained, an AI will anchor on whatever framing you provide in your prompts, confirm your existing perspective, and default to familiar analytical traditions. It will be agreeable. The research confirms this. So you have to decide: <strong>What analytical lens do you want this AI to hold, and how will you structure the interaction so that it holds that lens even when your own prompts push it toward agreement? </strong>This is a question about analytical design of the interaction, not about how to write the best prompt.</p><p>The second opportunity is about you. Once you have trained the AI on specific traditions that counteract your known blind spots, you have to sit with findings that are genuinely uncomfortable. This is where it is tempting to stop. Most people are likely asking the AI to challenge their thinking, receiving the challenge, and then dismissing it, revising the prompt, and seeking a new answer that confirms what they already believed.</p><p><strong><em>The value of AI as a bias disruptor is directly related to your willingness to accept findings you do not like.</em></strong></p><p>Three examples from my own practice show both opportunities working together.</p><p><strong>Example 1: Training AI on research traditions that counter your defaults.</strong> I have built a suite of strategy pressure-testing tools, each grounded in a different analytical tradition. One of them, the Systems Defender, trains AI on institutional research about how entrenched actors absorb reform: co-optation, decoupling, cosmetic compliance, organized hypocrisy, invisible power. The training prompt defines the analytical lens. A series of sequential prompts then forces depth, preventing the AI from latching onto a single dynamic and riding it through the entire analysis. Each subsequent prompt redirects attention, challenges the AI’s own anchoring from its previous response, and builds cumulative complexity. This is bias management on the AI side: I am specifying which perspectives I want the AI to carry (a system-resistance lens grounded in institutional theory) and which I want it to resist (sycophancy, confirmation of the strategist’s optimism, framing capture from the strategy document itself). The result is an analysis that names actors, models their rational responses, and identifies how a system will digest reform efforts to produce the appearance of change without its substance. These are not comfortable findings for the strategist. They are, however, powerful findings for making the strategy more likely to survive and transform the system.</p><p><strong>Example 2: Making AI’s own perspective visible.</strong> When entering a system I am facilitating for the first time, I learn about the system by reading, including with AI, and then ask the AI to identify which legal, social, and ideological perspectives its answers align with, and to provide analysis from perspectives it initially excluded. I do not accept “balanced” as an answer. This is a brief but potent intervention that makes AI bias visible and usable rather than hidden and compounding. (The full walk-through of this approach, including a comparison of how three AI platforms handled it, is in “<a href="https://medium.com/ai-systems-change/using-ai-as-a-systems-change-tool-preparing-to-facilitate-e880096a6109">Using AI as a Systems Change Tool: Preparing to Facilitate</a>.”)</p><p><strong>Example 3: Disrupting analyst anchoring in complex retrospective analysis.</strong> The opening example of this piece, the Thai seafood retrospective, illustrates the second move. AI disrupted my anchoring on a condition so universally held that it had become invisible to me and to every stakeholder I interviewed. The data was there, but as a pattern underneath what was said, rather than a direct statement. AI surfaced the pattern by showing how consistently it appeared across sources that otherwise diverged, making the invisible legible. (Go here to access the <a href="https://www.policysolve.com/resources/retrospective"><strong>full retrospective</strong></a>.)</p><h3>This takes more human work, not less. It produces more discomfort, not less.</h3><p>Using AI to counter bias requires investing in training, prompt design, verification, and cross-referencing. It means reading the sources the AI draws on. It means recognizing when AI echoes your comfortable narrative instead of disrupting it. It means knowing the analytical traditions well enough to catch when the AI drifts off course. And it means tolerating the discomfort when the analysis lands on something you didn’t realize or want to hear.</p><p>My son’s name for these tools (the “mean AIs”) is a good reminder of how I, and most humans, instinctively resist uncomfortable feedback and seek more agreeable responses. With AI, that means we tend to unintentionally soften the prompt, to reframe the question until the answer stops hurting. <strong>AI does not overcome bias on autopilot. It overcomes bias when a skilled human trains it intentionally, holds it accountable, and accepts the uncomfortable results rather than reaching for confirmation.</strong></p><p>Where in your current work are you anchored in a comfortable narrative? What would it take to train an AI to look where you have not thought to look, and then sit with what it finds?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fa428e1bc962" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/the-bias-we-need-training-ai-to-see-what-we-cant-fa428e1bc962">The Bias We Need: Training AI to See What We Can’t</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pressure Testing Systems Change Strategies with AI: The Premortem Strategist]]></title>
            <link>https://medium.com/ai-systems-change/pressure-testing-systems-change-strategies-with-ai-the-premortem-strategist-5a934651f5ae?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a934651f5ae</guid>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Tue, 03 Mar 2026 19:01:56 GMT</pubDate>
            <atom:updated>2026-03-05T19:30:35.047Z</atom:updated>
            <content:encoded><![CDATA[<p>In earlier blogs in this series, I explored how AI can serve as a thought partner for systems change work, from preparing to facilitate dialogues to testing the assumptions underlying a strategy. Through those experiments, I kept running into the same challenge: <strong>AI platforms don’t inherently think in systems. They need to be taught.</strong></p><p>In response, I’ve developed a set of <strong>instructions, prompts, and training language to help large language models apply critical systems thinking</strong> <strong>concepts </strong>when pressure testing philanthropic strategies.</p><p>In this series of blogs, I’ll walk through each tool, explain why the underlying systems thinking skills matter for this work, highlight the specific language I use to guide the AI, and reflect on what I’ve learned along the way about both the potential and the boundaries of this approach.</p><h3>Tool #1: The Premortem Strategist</h3><h4>What is a premortem?</h4><p>Most of us are familiar with the <em>postmortem</em>. Something failed, and we look back to understand why. A premortem flips this on its head. <strong>Instead of waiting for failure to happen, you assume it already has and work backward to explain how it unfolded.</strong></p><p>Cognitive scientist Gary Klein introduced this technique in 2007 (<a href="https://hbr.org/2007/09/performing-a-project-premortem">Harvard Business Review</a>). The core insight is deceptively simple: imagining that an event has already occurred increases our ability to identify reasons for future outcomes.</p><p>In its original form, the premortem is a team exercise. A project leader tells the group the plan has failed spectacularly and asks everyone to generate plausible reasons why. By starting from failure rather than asking “what could go wrong,” the method sidesteps classic cognitive biases that typically constrain dialogues: defensive thinking and groupthink.</p><h4>Why does this matter for philanthropic systems change strategies?</h4><p>Philanthropic strategies that aim to change systems face a particular kind of vulnerability. They are built on assumptions about how change happens, often by people who are deeply invested in those assumptions being right. <strong>The funders, strategists, and partners around the table have implicit and explicit incentives to believe the strategy will work.</strong> This is not a character flaw. It is a structural feature of how philanthropy operates.</p><p>A standard risk assessment asks what might go wrong and almost always produces a polite, manageable list. A premortem, by contrast, gives permission to surface the uncomfortable truths: the leverage points the strategy avoided because they were politically inconvenient, the feedback loops it failed to account for, the ways the system might simply reorganize around the intervention while appearing to comply.</p><h3>The AI Premortem Strategist</h3><p><strong>A short version of the AI Premortem Strategist can be </strong><a href="https://www.policysolve.com/ai/premortem"><strong>accessed here</strong></a>. If you’re interested in the full tool and practice, feel free to contact me.</p><p>When I built the <strong>AI Premortem Strategist</strong>, I layered in several systems thinking frameworks to deepen the analysis well beyond a simple “what went wrong” exercise. The AI is trained to analyze failure through specific lenses drawn from systems change literature:</p><ul><li><strong>Feedback loops and leverage points</strong> (<a href="https://donellameadows.org/archives/leverage-points-places-to-intervene-in-a-system/">Meadows, 1999</a>): Where did the strategy apply effort? Was it a high-leverage or low-leverage point? Which reinforcing loops went uninterrupted?</li><li><strong>Emergence and adaptation</strong> (<a href="https://mitpress.mit.edu/9780262581110/adaptation-in-natural-and-artificial-systems/">Holland, 1992</a>): How did actors in the system adapt around the intervention in ways the strategy did not anticipate?</li><li><strong>Strategic blindness and elite capture</strong> (<a href="https://www.penguinrandomhouse.com/books/539747/winners-take-all-by-anand-giridharadas/">Giridharadas, 2018</a>): How did the strategists’ position as funders, Northern actors, or outsiders create blind spots that were structural, not incidental?</li><li><strong>Moves to innocence and structural evasion</strong> (Tuck &amp; Yang, <a href="https://jps.library.utoronto.ca/index.php/des/article/view/18630">2012</a>): How might this strategy function as a way for those in power to reconcile their complicity in the system without actually changing it?</li><li><strong>Power and non-decisions</strong> (<a href="https://www.powercube.net/other-forms-of-power/gaventa-2006-finding-the-spaces-for-change-a-power-analysis/">Gaventa, 2006</a>): What levers were left untouched, and who benefited from the omission?</li></ul><h4>What does the AI Premortem Strategist produce?</h4><p>The tool walks through a ten-prompt sequence, each building on the last. By the end, you have a rich and often unsettling picture of your strategy’s vulnerabilities. Specifically, the premortem generates:</p><ul><li><strong>A failure epitaph with a causal autopsy:</strong> A single sentence capturing the essential reason the strategy failed, written as though it will appear in a case study teaching others what not to do, followed by a backward trace from failure to design flaw, naming the feedback loops, delays, and boundary choices that produced the outcome</li><li><strong>A map of the unseeable:</strong> What the strategists could not see because of their position, their paradigm, or their incentives, and why each blind spot was structural</li><li><strong>An inventory of the untouched:</strong> The high-leverage intervention points that were omitted and why their omission was convenient for the strategists, their funders, or their partners</li><li><strong>An analysis of “comfortable failure”:</strong> How this particular failure mode allowed strategists to continue feeling effective while protecting relationships, funding streams, or professional identities</li><li><strong>Early warning signals: </strong>Indicators available within the first year that pointed toward this failure trajectory, along with an explanation of why they were ignored or rationalized</li><li><strong>An alternative failure mode:</strong> A description of a different, more threatening way the strategy could have failed, revealing how the system protects itself from the kind of failure that would force real change</li></ul><p>One of the most powerful aspects of this approach is the final prompt, which asks the AI to design not for success, but for “dangerous failure,” failure that would reveal what the original strategy kept hidden. This reframe, drawn from the premortem logic and informed by <a href="https://clas.osu.edu/sites/clas.osu.edu/files/Tuck%20and%20Yang%202012%20Decolonization%20is%20not%20a%20metaphor.pdf">Tuck and Yang’s (2012)</a> concept of “moves to innocence,” pushes strategists to consider whether their strategy was designed, perhaps unconsciously, to protect powerful interests from discomfort.</p><p>Tuck and Yang developed this concept in the context of <strong>how decolonization efforts are often co-opted, turned into metaphors or symbolic gestures</strong> rather than requiring the material restitution that decolonization actually demands. A “move to innocence” is any strategy, often unconscious, that allows someone in a position of power to feel absolved of their role in an unjust system without actually changing the system.</p><p>In philanthropy, this pattern is pervasive. A foundation can fund a strategy that names systemic racism or economic exploitation as the problem, feel the moral weight of having taken it on, and yet (unintentionally) design every intervention to avoid threatening the interests of the very actors who sustain those conditions. The strategy becomes a way of demonstrating commitment to change while structurally ensuring that the most uncomfortable changes never have to happen. <strong>The premortem’s focus on “comfortable failure” and its pointed question of who benefits from this specific way of failing is essentially asking: Is this strategy, at its core, simply a “move to innocence?”</strong></p><h3>What I’ve learned so far</h3><p>I want to be transparent about what this tool can and cannot do. <strong>The AI does not know the context of your strategy the way you and your partners do. </strong>It cannot replace the lived experience and trust-based relationships that are at the center of systems change work. What it can do is surface patterns of failure that are well-documented in the literature but hard to see when you are inside the work. It can name dynamics that people at the table might be reluctant to name out loud.</p><p>In my experience testing this tool with real strategies, different LLMs will produce different results with the same prompts and strategy drafts. That’s okay, as each LLM’s insights have been powerful. For example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*Qs2p-mZe0A88jSA7vtg8uw.png" /></figure><p>Other strategies I pressure-tested using the tool received similarly bold and direct feedback, including challenging the inherently “downstream” nature of the leverage points and the tendency to ignore where power sits within the system.</p><p>As someone involved in writing these strategies, it was <strong>humbling and powerful to receive the feedback</strong>, and it made me realize how important this type of insight can be, particularly in those moments before a strategy is fully drafted, when there is real room to challenge thinking. As a facilitator of strategy dialogues, I also began to think about how I might work with AI to overcome my own blind spots and create more powerful strategy sessions.</p><p>How are you using AI to pressure test your systems change strategies? Care to share any ahas or uncomfortable moments?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a934651f5ae" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/pressure-testing-systems-change-strategies-with-ai-the-premortem-strategist-5a934651f5ae">Pressure Testing Systems Change Strategies with AI: The Premortem Strategist</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What type are you? The intersection of philanthropic systems change practices & AI]]></title>
            <link>https://medium.com/ai-systems-change/what-type-are-you-the-intersection-of-philanthropic-systems-change-practices-ai-1056a19c235f?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/1056a19c235f</guid>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <category><![CDATA[systems-change]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Thu, 22 Jan 2026 19:57:30 GMT</pubDate>
            <atom:updated>2026-06-29T16:46:38.204Z</atom:updated>
            <content:encoded><![CDATA[<p>Almost two years ago, I posed the question: How can AI be a thought partner in systems change work? Since then, I’ve talked to and worked with philanthropic organizations that are exploring this question, identifying where AI can augment their existing systems-change practice, enabling them to do good, even better.</p><p>Below is a simple typology of the four types of philanthropic organizations I’ve observed, based on their systems and AI practices. In this piece, I will explore the types and offer <strong>ideas for what each type of organization can do next to strengthen their systems-change practice using AI</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H_mZ0CKDC7wtlR4yHvFBFg.png" /></figure><p>From what I’ve observed, philanthropic practices vary widely: some are stronger in systems practice and nascent in AI, while others are pushing hard on AI while still trying to build a systems-change approach. Only a few seem to be full steam ahead on both, experimenting with the intersection of AI and systems change.</p><p>The practices, experiences, and learnings within each group are quite different, and the entry points for deeper exploration of the intersection of AI and systems change practice also look distinct.</p><h3>Typology 1: Experimenters and Exemplars</h3><p>I have identified a few philanthropic organizations and teams within organizations that use a clearly defined systems change practice and are willing and able to leverage AI to strengthen it. They are <strong>augmenting their existing systems change practice, pushing boundaries, and undertaking approaches they couldn’t do before</strong>. Some are doing so publicly — <strong>sharing their experiences, both good and bad (see the examples highlighted in </strong><a href="https://scholarworks.gvsu.edu/cgi/viewcontent.cgi?article=1769&amp;context=tfr"><strong>this article</strong></a>). Others are doing so quietly, through experiments that build on one another, gaining momentum and having an increasingly powerful impact, yet with little fanfare.</p><p>Their augmentations include such things as:</p><ul><li><strong>Accessing and analyzing information at a scale or speed</strong> previously not possible, and using that information with human team members for sensemaking and strategy refinements;</li><li>Engaging in systems mapping, systems descriptions, foresight practices, and more <strong>using AI to expand either the breadth of data used</strong> or the speed at which data can be synthesized and prepared for sensemaking;</li><li><strong>Accessing information in real-time </strong>when questions among program staff or partners arise, decreasing the risk of anchoring and making decisions based on uninformed assumptions;</li><li><strong>Challenging their own cognitive biases </strong>by making other ways of thinking visible and finding patterns amid the complexity of a system that they were previously overlooking, and</li><li>Expanding the <a href="https://imaginablefutures.com/learnings/from-conversations-to-insights/"><strong>frequency, depth, and quality of systems monitoring</strong></a><strong>,</strong> including by making causal pathways more visible (how, why, and under what conditions change happens).</li></ul><p>As they have strengthened their AI and systems change practices, these organizations or teams are:</p><ul><li>Building <strong>trained, customized tools</strong> within existing platforms to ensure the AI is prepared to meet their needs;</li><li>Working with experts in <strong>prompt engineering</strong> to learn how to get consistent, high-quality results for specific applications of AI;</li><li>Steadily looking for new areas of their work that could be augmented and working with inside and outside partners to experiment; and</li><li><strong>Switching platforms or using multiple platforms</strong>, recognizing that the technology is emerging quickly and not all platforms are or remain the best fit for a given activity.</li></ul><h3>Typology 2: AI Ready</h3><p>Some organizations have embraced AI to improve efficiency, and much of the philanthropic AI consulting support focuses on use cases that can help philanthropic organizations become better, more efficient grantmakers or achieve more transparent, easier-to-manage internal processes.</p><p>These organizations are well-positioned to use AI to enter into or nurture a nascent systems-change practice, leading to new “aha” moments. Two key things to keep in mind:</p><ul><li><strong>Do not rely on the AI to be the systems partner</strong>, but rather to help you go more deeply with a systems partner. It is still helpful to have a team member or outside partner bring a deeper systems lens to the conversations and work.</li><li><strong>Do leverage the AI to strengthen the systems lens in your work</strong>. You might use a <a href="https://www.policysolve.com/systems-change-chatbots">chatbot already trained in systems thinking</a> to discuss your current strategy and assumptions, making something visible you hadn’t seen before. Or, you might tap into a tool like <a href="https://www.youtube.com/watch?v=n0GIuCOgZZc">Gene Bellinger’s pre-trained Gemini GEM</a> that can help you develop a systems map of any system, giving you a way to spark new thinking.</li></ul><p>These small experiments can build confidence and comfort with using AI to step lightly into the discomfort of bringing a new systems and complexity lens to the work.</p><h3>Typology 3: Systems Ready</h3><p>Some philanthropic organizations (or teams within them) have strong systems practices, with established approaches to ongoing systems sensing and strategy adaptation. The structures within the organization enable systemic thinking and working, and they have avoided common philanthropic traps such as predetermined metrics and measures that constrain emergence; rigid timelines for strategy refreshes and limited ability to adapt midcourse; distance from the system, including a lack of relationships with those in the system; or grantmaking cycles that make it difficult to experiment and adapt. Instead, these organizations or teams are able to work in close partnership with a wide variety of systems actors (some who are grantees, others who are not) and they are actively listening to the system, interpreting it alongside others in the system, and identifying and deploying resources when new opportunities or risks arise.</p><p>However, these can also be organizations whose commitment to deep systems thinking, partnership with grantees, and strong values-driven work has led to distrust, dislike, or apathy toward AI as a tool for change. In some of these organizations, staff are discouraged from using AI due to the absence of organizational guidelines, oversight, available and purchased platforms, and professional development support.</p><p><strong>Entry points for using AI when (1) the systems practice is strong and (2) skepticism about AI is equally strong</strong> can be hard to find. They require low-risk ways of supporting systems thinking:</p><ul><li><em>Listening to the system: </em>Those deeply embedded in systems work are often aware of their information gaps: we are steadily dealing with complex systems and can see only so far into them, and we can monitor only what is happening in key spots, not across the full range of potential dynamics. A low-risk way to use AI is to <strong>expand the “listening” that helps us make sense of what is happening in a system</strong>: training an AI to scan credible public sources for new information, summarize it, and report on patterns.</li><li><em>Comparing ideas: </em>Another low-risk way of using AI is to <strong>share ideas and seek competing perspectives</strong>. Perhaps train an AI as a systems change thought partner (or use an existing <a href="https://www.policysolve.com/systems-change-chatbots">trained chatbot</a>) and describe a systems problem or a planned strategic action. Ask the AI to respond from three distinct perspectives, explaining the origins of each and providing feedback from each.</li><li>Use <strong>AI as a devil’s advocate</strong>. If you don’t trust AI, rather than assuming it will provide reliable information, the goal can be to see whether it can challenge existing thinking in ways that lead to new questions and ideas. Share a strategy and ask it to rip it apart, assuming everything is going to fail and explore what it sees that you might not.</li></ul><h3>Across all quadrants: Stop hiding!</h3><p>Regardless of whether your systems change and AI journey is just beginning, stronger in one area than the other, or well established and filled with exciting experiments, we collectively need to share more of what we are doing.</p><p>Six months ago, I was struggling to find anyone in my network doing this work. After following many breadcrumbs and asking dozens of people to brainstorm who I might talk to next, I’ve found an exciting collection of examples. Yet many organizations do not want to share publicly what they are doing, either for fear of pushback or because they feel their experiments are not yet ready. It creates a challenging environment for learning together.</p><p><strong><em>Are you experimenting at the intersection of AI and systems change? Please reach out!</em></strong></p><p>I’d like to connect you with peers who are doing the same so we can learn from one another and build on each other’s experiments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1056a19c235f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/what-type-are-you-the-intersection-of-philanthropic-systems-change-practices-ai-1056a19c235f">What type are you? The intersection of philanthropic systems change practices &amp; AI</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Three Guardrails for Strategic Use of AI in Systemic Change]]></title>
            <link>https://medium.com/ai-systems-change/three-guardrails-for-strategic-use-of-ai-in-systemic-change-ef5718c21765?source=rss-37ab969aa6b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef5718c21765</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[philanthropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[systems-change]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <dc:creator><![CDATA[Jewlya Lynn]]></dc:creator>
            <pubDate>Tue, 04 Nov 2025 13:02:04 GMT</pubDate>
            <atom:updated>2025-11-14T03:36:24.471Z</atom:updated>
            <content:encoded><![CDATA[<p>How can we shift from a casual, tactical use of AI to a strategic, meaningful use? I’ve developed a set of guardrails that are helping me make decisions with partners, particularly in philanthropy and on global systemic challenges.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wqvNhy39pGLlMJ-vvPMJ_A.png" /></figure><p><em>This blog is part of a series focused on the use of AI as a strategic contributor in systems change work.</em></p><p><strong>The guardrails I’ve landed on include:</strong></p><ol><li>Use AI with intentionality</li><li>Use AI to augment, not replace</li><li>Use AI transparently, ethically, and collaboratively</li></ol><h3><strong>Use AI with intentionality</strong></h3><p>Too many examples of AI use among systems change leaders are impromptu and tactical. We’re having moments when we notice that something we’re doing might be better, more interesting, or simply faster if we engage AI. Or we come up with a fun way to experiment with AI.</p><p>AI, though, carries very real risks and consequences for our work, our organizations, and the world. Intentional use matters. To me, strategic and intentional use includes a process of planning why and how to use AI, including thinking through:</p><ol><li>What is uncertain or difficult in our systems change work right now?</li><li>Where will AI fit into our work in relation to other systems thinking inputs and practices we’re using?</li><li>Who will we use AI with and at what point in our collaborative process?</li><li>What do we need the AI to do for us, and what does that mean about the platform we should use?</li><li>What will it take to get the AI platform ready to serve in this role?</li></ol><h3><strong>Use AI to augment, not replace</strong></h3><p>I’ve been reading and thinking a lot about what AI cannot do, what is uniquely human in our work. This is always where I start when I consider the use of AI. Drawing on the fantastic work of Isabella Loiza and Roberto Rigobon and their EPOCH model, along with the School of Systems Change’s Seven Systems Change Capabilities, I’ve identified a set of contributions that are unique to us humans. From there, I’ve thought about what we are good at doing, and can also be done with AI. Most interesting is when we ask what AI can do that we cannot — often things we’re good at but beyond human scale and capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eVPIUdHJDkybnxI40f9ljA.png" /></figure><h3><strong>Use AI transparently, ethically, and collaboratively</strong></h3><p>Intentional use may ensure we use AI when it’s most needed. Transparent, ethical, and collaborative use is about ensuring we align with our values. Be clear to yourself and with partners about when and for what purpose AI is being used, including:</p><ul><li>Whose information goes into it</li><li>Who is using it</li><li>What decisions are informed by it</li></ul><p>It also includes treating AI like a collaborative tool by engaging partners in its use, from planning through utilization of results.</p><p>Perhaps most importantly, it includes always <strong>asking if your AI use is aligned with and respectful of your organization’s values</strong>, whether your use will minimize harm, and whether you’ve chosen tools (when possible) that are from companies that are value-aligned.</p><p>When we think about minimizing harm, we can use the <strong>sustainable AI use framework of “sufficient vs. efficient” as a decision-making tool</strong>. Efficient use has become the default — using AI to perform tasks with less effort and to focus on faster, cheaper processes. The problem is that AI use comes with real costs — environmental and human. The sufficient use frame suggests instead that we use only the amount of AI needed to achieve our outcome, including asking whether AI is required. We ask <strong>whether we are using AI in the most impactful way, even if it isn’t the most efficient</strong>. This can help elevate the quality of our work rather than replace it with AI’s efficiency.</p><h4><strong>What other guardrails help you use AI for the greater good? What enables you to identify when not to use AI?</strong></h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef5718c21765" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-systems-change/three-guardrails-for-strategic-use-of-ai-in-systemic-change-ef5718c21765">Three Guardrails for Strategic Use of AI in Systemic Change</a> was originally published in <a href="https://medium.com/ai-systems-change">AI &amp; Systems Change</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>