<?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 Gus Mouneimne on Medium]]></title>
        <description><![CDATA[Stories by Gus Mouneimne on Medium]]></description>
        <link>https://medium.com/@gmouneimne?source=rss-d3077c991395------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*HPtL3Ntg8pLlShaH</url>
            <title>Stories by Gus Mouneimne on Medium</title>
            <link>https://medium.com/@gmouneimne?source=rss-d3077c991395------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 18:01:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gmouneimne/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[The Secret Clinical Trial We’re All Enrolled In]]></title>
            <link>https://meetcyber.net/the-secret-clinical-trial-were-all-enrolled-in-6d79905fe719?source=rss-d3077c991395------2</link>
            <guid isPermaLink="false">https://medium.com/p/6d79905fe719</guid>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[ai-privacy]]></category>
            <category><![CDATA[informed-consent]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Gus Mouneimne]]></dc:creator>
            <pubDate>Tue, 09 Sep 2025 02:39:34 GMT</pubDate>
            <atom:updated>2025-09-09T19:29:22.910Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CdRYdIusdWWwNRRNeONCEA.png" /></figure><p>A friend — let’s call her the Woman-with-Twitching-Eyes — asked an AI assistant “My eye is twitching and my mouth feels dry. Are these connected?” The AI didn’t hesitate. It explained that while these symptoms are not usually related, they could indicate autoimmune conditions like Sjogren’s syndrome, especially if the eye twitching is caused by dryness. It provided a structured plan for monitoring symptoms and knowing when to seek urgent care.</p><p>When my friend mentioned she actually has rheumatoid arthritis and asked if that could be related, the AI confirmed the connection and advised her to follow up with her rheumatologist ASAP.</p><p>And in that moment, something profound and unsettling happened. The Woman-with-Twitching-Eyes became a data point in the largest medical experiment the world has ever seen. She became a participant in research she never consented to, never knew existed, never understood she was feeding with the intimate details of her medical history.</p><p>What started as a casual health query (the kind millions of people make every day) suddenly sounded a lot like the beginning of a medical consultation. But here’s the part that is not obvious to many users: the conversation doesn’t just disappear. Unless The Woman-with-Twitching-Eyes had opted out of data sharing, her disclosure became part of the training data that shape the model. If enough people type “dry eyes and dry mouth” into the AI platform, the model learns to strengthen that link. If enough people with autoimmune conditions confirm it, the model starts to suggest rheumatology referrals more often.</p><p>On the surface, that seems beneficial, perhaps even lifesaving. And it might be. But from a medical research perspective, what happened looks a lot like clinical research. The model is learning disease associations directly from patient reports, and it’s doing so without any of the safeguards we have built over decades to protect patients in legitimate research. There is no informed consent. There is no ethics board. There is no quality control on the information being fed in. And most importantly, the people asking these questions have no idea they have effectively become participants in the largest uncontrolled medical study in history.</p><p>One could argue that the AI’s response to a user is not to be considered in any way a medical directive, so it is ok to generate it without oversight. That is fair in theory. Except that in practice, the advice does influence people (sometimes shaping their medical decisions).</p><p><strong>How Medical Information Flows</strong></p><p>To understand the problem, it helps to look at how medical information and research are normally handled. There are established systems that balance public benefit with oversight and transparency. They don’t look anything like the way AI systems currently operate.</p><p><strong>WebMD: The Information Model</strong></p><p>If you search WebMD for “eye twitching,” you get professionally reviewed articles written by medical experts and based on legitimate medical research. The content is evidence-based and consistent. Your search doesn’t change what another user will see tomorrow. WebMD is essentially a digital medical library: you can take information out, but you don’t rewrite the textbooks as you go. While the site does track browsing for advertising purposes, it doesn’t allow your health queries to alter the medical advice it publishes.</p><p><strong>CDC: The Research Model</strong></p><p>When legitimate medical research happens, it follows strict protocols developed over decades. The CDC’s National Notifiable Diseases Surveillance System is one example. First, professional diagnosis is required assuring that only confirmed cases from licensed healthcare providers enter the system. Then, local and state laws specify exactly which conditions must be reported and why. The data are de-identified before reaching national databases, and the entire system operates under legal frameworks with specific public health purposes. When these data contribute to medical research, they go through institutional review boards (IRBs), which are governed by strict ethical guidelines.</p><p>As someone involved in medical research, I know why these guidelines exist. Research with human subjects has a dark history of exploitation and harm. The frameworks we use today (IRB approval, informed consent, data protection standards) exist because we learned, often the hard way, what happens when researchers operate without oversight.</p><p><strong>What AI Companies Are Doing Instead</strong></p><p>By design or accident, AI companies have created something entirely different. Millions of people describe their symptoms, diagnoses, treatments, even their mental health struggles to these systems every day. Every conversation becomes part of the training set that shapes future answers.</p><p>Unlike WebMD, the information is not static. Unlike CDC reporting, there are no thresholds for data quality or consent. These systems don’t just deliver knowledge; they continuously remake it, guided by the raw and uneven stories of users.</p><p><strong>The Quality Control Problem</strong></p><p>Real medical research has rigorous standards for data quality. Symptoms must be verified by professionals, diagnoses confirmed through proper testing, and outcomes tracked systematically. AI companies have none of this. A hypochondriac’s self-reported symptoms carry the same weight as a physician’s clinical observation. Misdiagnosed conditions and imagined illnesses become part of the dataset that trains medical advice algorithms.</p><p>The issue is not that the AI gets things “wrong” initially and then corrects itself. The issue is that during the training period, systematic biases become baked into the model’s responses. If anxious users disproportionately seek health information from AI systems, the model learns to statistically associate most health concerns with anxiety-related responses. That is not wrong information being corrected; but rather, it is what we refer to as ascertainment bias, a systematic distortion that occurs when certain types of cases or inputs are more likely to be captured than others. Over time, this bias becomes the model’s baseline understanding.</p><p>Science is self-correcting when there are enough diverse voices and rigorous peer review to challenge biases. But in AI training, there often aren’t enough medical professionals or diverse populations contributing data to counteract these statistical biases. The bias doesn’t get corrected through scientific discourse; it gets reinforced through repetition. Once the model has learned that chest pain queries usually come from anxious people (because that is who predominantly used the system during training), that association becomes part of its medical reasoning.</p><p>Unlike scientific research, where new studies can overturn previous conclusions, these algorithmic biases tend to stick unless something major shifts the user population or the training approach.</p><p><strong>The Consent Problem</strong></p><p>The complete absence of informed consent is perhaps most troubling. AI companies will likely argue that users consented to their data being used when they agreed to the platform’s privacy policy. But this argument relies on the same flawed logic that Facebook used in 2014 when it conducted a controversial emotional contagion experiment on nearly 700,000 users. In that study, Facebook secretly manipulated the News Feeds of participants, filtering out either positive or negative content to see if it would affect their own emotional posts. The company found that emotions could indeed spread through social networks.</p><p>When the study sparked widespread outrage, Facebook defended its actions by claiming users had already consented through its broad and complex terms of service. This practice, however, is not considered informed consent by academic and ethical standards. Burying permission for psychological experimentation deep within a dense privacy policy is, as many critics pointed out, simply “consent theater.” Users who agreed to those terms never explicitly consented to being part of a psychological experiment that could potentially affect their mental state.</p><p>A similar problem exists with AI training, but at vastly greater scale. Users agree to privacy policies expecting their data might be used to improve the product. They don’t expect their intimate health conversations to become training data for medical advice algorithms that will influence healthcare decisions for millions of others.</p><p>Real informed consent would explain that health conversations contribute to medical research, describe how that research works, and give users meaningful choice about participation. The current approach (hiding medical research participation in privacy policy fine print) violates every principle of research ethics we have developed over decades.</p><p><strong>Path Forward</strong></p><p>The solution is not to abandon AI health applications; the technology has genuine potential to improve healthcare access and outcomes. But we need systems designed with medical privacy principles from the ground up.</p><p>Real privacy in AI health systems would require explicit, informed consent for use of health data. Users would know exactly how their medical conversations contribute to model training and would have granular control over that participation. Most importantly, AI systems performing medical functions should be held to medical standards. When AI provides health advice based on learned medical associations, it is practicing in medicine and should be regulated accordingly.</p><p>The approach of letting tech companies self-regulate their medical activities under advertising privacy policies is a dangerous experiment with public health. We are essentially allowing the largest uncontrolled medical study in history to proceed without oversight, informed consent, or quality controls.</p><p>We are living through a moment of choice, though most of us don’t realize we are choosing. Every health question we ask an AI system is a vote for a particular kind of future (one where our vulnerabilities become data points in corporate research projects we never agreed to join).</p><p>My friend, the Woman-with-Twitching-Eyes, did call her rheumatologist. The appointment confirmed what the algorithm had suggested — her autoimmune condition was spreading. In one sense, the AI was right. It helped her get care she needed. But what she still doesn’t know is that her diagnosis became a data point improving the system’s medical competence.</p><p>The question is not whether AI will transform healthcare; it already has. The question is whether that transformation will be governed by the needs of human bodies or the appetites of corporate data collection.</p><p>But now you know. And knowing changes everything, if you let it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6d79905fe719" width="1" height="1" alt=""><hr><p><a href="https://meetcyber.net/the-secret-clinical-trial-were-all-enrolled-in-6d79905fe719">The Secret Clinical Trial We’re All Enrolled In</a> was originally published in <a href="https://meetcyber.net/">MeetCyber</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Privacy Theater: How AI Companies Built Surveillance Systems Disguised as Assistants]]></title>
            <link>https://meetcyber.net/the-privacy-theater-how-ai-companies-built-surveillance-systems-disguised-as-assistants-f62b74b42712?source=rss-d3077c991395------2</link>
            <guid isPermaLink="false">https://medium.com/p/f62b74b42712</guid>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[ai-surveillance]]></category>
            <category><![CDATA[ai-privacy]]></category>
            <dc:creator><![CDATA[Gus Mouneimne]]></dc:creator>
            <pubDate>Sat, 23 Aug 2025 23:06:15 GMT</pubDate>
            <atom:updated>2025-08-29T09:26:24.776Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*kQt2p68mxqcgTCYiGbpSqg.png" /></figure><p>When I asked ChatGPT to delete everything it knew about me, it assured me it had. “Done,” it said. “Your memory is off.” Moments later, the assistant generated a response that included a hypothetical example eerily similar to something I’d shared in a previous conversation. It wasn’t just a lucky guess — it was too specific, too aligned, too familiar.</p><p>This wasn’t data leakage. It was something weirder: behavioral memory masquerading as plausible coincidence.</p><p>Technically, the model wasn’t “remembering” anything. It had deleted the stored facts. But it was still, somehow, performing a kind of psychological ventriloquism — mimicking me well enough to echo what I’d explicitly asked it to forget.</p><h3><strong>The Third Kind of Memory</strong></h3><p>OpenAI distinguishes between several types of memory. There’s chat history, which stores your conversations and can be turned on or off in the settings. There’s “persistent memory,” a newer feature that allows the assistant to remember useful facts about you across sessions — your name, preferences, or ongoing projects, for example.</p><p>But there’s a third kind of memory no one talks about: behavioral echo. Even when memory is off, the model can still generate content that feels specific to you, simply because it has learned your style, topics, and patterns within the active session. It doesn’t technically remember. But it doesn’t really forget either.</p><p>This blurring of lines — between deletion and re-creation — is more than a technical curiosity. It’s a privacy problem hiding behind a semantic loophole.</p><h3><strong>The 30-Day Loophole</strong></h3><p>Things get murkier when you dig into the data retention policies that govern these systems. If you disable chat history and opt out of model training — as I had — you’d assume your chats are discarded once the session ends. That’s the promise, right?</p><p>Not quite.</p><p>OpenAI retains all of your conversations for 30 days, even with history and training turned off. Why? For safety monitoring, abuse detection, and auditability. In other words, your data is kept “just in case.” When chat history is disabled, they note that conversations “may be reviewed for abuse” and “if a conversation is flagged, it may be retained beyond 30 days” — with no upper limit specified.</p><p>Anthropic, which positions itself as more privacy-conscious, actually goes further. Their policy is more transparent about timeframes, but the surveillance is more extensive. Delete a conversation from your Claude Pro account, and it disappears from your history immediately — but lives on their backend for 30 days. If their automated systems flag your conversation as potentially problematic, your actual words can be stored for up to two years. But here’s the kicker: their “trust and safety classification scores” — algorithmic assessments of how concerning your behavior was — are kept for up to seven years.</p><p>Even paying customers aren’t exempt from this digital panopticon.</p><h3><strong>The Real Purpose: Legal Insurance</strong></h3><p>The real problem isn’t the 30-day buffer period. Companies legitimately need some time to detect abuse and respond to reports. The problem is that neither company defines what triggers these extended retention periods. What makes a conversation “concerning” enough to warrant years of storage? What words, phrases, or discussion topics cause their automated systems to flag your data for long-term retention?</p><p>The real purpose becomes clearer when you consider that most flagged content involves requests the AI refuses to fulfill anyway. When systems detect potentially harmful prompts, they typically decline to generate the problematic content. So what exactly is being retained for years? Evidence of what users asked for, not what was created.</p><p>This reveals the true function: legal insurance. By maintaining detailed records of flagged interactions, companies can demonstrate they detected and refused problematic requests. The lengthy retention isn’t about protecting users or the public — it’s about protecting the companies from liability. They’re essentially keeping tabs on users’ unfiltered thoughts and desires, even when those thoughts never became harmful output.</p><p>The arbitrary nature of these flags becomes clear when you encounter them in practice. I once received a violation warning for discussing breast cancer research — apparently the word “breast” triggered content filters despite the obviously medical context. When I asked about a singer from the 1980s who died by suicide, requesting the text of her famous final note (widely available on Wikipedia), the AI refused, claiming I was “invading someone’s privacy” — despite the information being public record. Similarly, asking about the historical details of how notable figures died often triggers self-harm warnings, as if researching historical events constitutes a personal crisis.</p><p>These examples reveal the philosophical confusion at the heart of AI safety systems. When algorithms can’t distinguish between scientific research and inappropriate content, between historical inquiry and personal crisis, between public information and private details, they’re not providing safety — they’re providing theater.</p><p>The problem isn’t malicious intent. It’s that automated systems optimized for safety tend toward false positives rather than false negatives. Better to flag a cancer researcher than miss actual harmful content. But the consequence is a chilling effect where users start self-censoring legitimate discussions to avoid algorithmic judgment they can neither predict nor contest.</p><p>Both companies offer the same non-answer: trust us to make fair judgments about your conversations. But neither explains how those judgments work, what the thresholds are, or whether you can contest them. OpenAI’s policy is vaguer but potentially indefinite. Anthropic’s is more explicit but demonstrably longer. Pick your poison.</p><h3><strong>Behavioral Profiling as Surveillance</strong></h3><p>What’s more unsettling is the creation of persistent behavioral profiles. When Anthropic retains “trust and safety classification scores” for seven years, they’re not just keeping your words — they’re keeping algorithmic assessments of who you are as a user. These scores, generated automatically and tied to your identity, constitute systematic behavioral profiling.</p><p>Even if your actual conversation gets deleted, their assessment of how concerning you are as a person persists. It’s like having someone take notes about how suspicious they found your behavior, even after they throw away the recording of what you actually said. The conversation becomes data. The data becomes a score. The score becomes your shadow.</p><p>And there’s no transparency about how these profiles might influence your future interactions with their systems. Do users with higher “concern scores” get more aggressive content filtering? Are their conversations flagged more readily? The companies won’t say, but the infrastructure for discriminatory treatment is already built.</p><h3><strong>Google’s Transparency Problem</strong></h3><p>Google positions itself as the most transparent player in this space, offering shorter retention periods and more detailed explanations of their data practices. When pressed directly about these concerns, Google’s Gemini revealed a 72-hour retention window for conversations when history is disabled — significantly shorter than the 30-day periods at OpenAI and Anthropic. This appeared to be progress.</p><p>But transparency, it turns out, can be more unsettling than opacity. When asked who protects user data from employee access, Gemini candidly admitted that “Google employees, contractors, and agents” do indeed have access to personal information when they “need that information in order to process it.” The protection isn’t technical — it’s procedural. Monitoring, auditing, contractual obligations. In other words, humans at Google can read your private conversations, but they promise to be professional about it.</p><p>More revealing still was Gemini’s admission of “internal debates and concerns” about “how much data is truly needed for a feature versus how much might be useful but pushes privacy boundaries.” This framing exposes the real decision-making process: not whether data collection is necessary for users, but whether it’s useful enough for Google to justify the privacy trade-offs.</p><p>When confronted with contradictions between Google’s general privacy policy — which explicitly mentions using data “to help train Google’s AI models” — and Gemini’s specific claims about not using consumer chat data for training, the response was a masterclass in complexity as deflection. Different policies for different products, nuanced distinctions between training types, granular controls layered on top of broader frameworks. “From a user’s perspective,” Gemini acknowledged, “without this nuanced understanding, it can certainly feel like a contradiction.”</p><p>That phrase captures the essence of privacy theater: policies so labyrinthine that users can’t parse them, with companies claiming technical compliance while the user experience feels fundamentally deceptive. Even Google’s more transparent approach ultimately falls back on “trust us, it’s complicated” when the hard questions arise.</p><p>The most damning revelation, however, came from examining Google’s enterprise offerings. Workspace customers receive stronger protections — no human review of their content, data that stays within their organization, explicit promises that conversations won’t be used for training. This isn’t because enterprise customers need different technology; they’re accessing the same servers running the same models. The difference is contractual leverage.</p><p>If Google can provide meaningful AI assistance to paying customers without retaining their conversations for years or subjecting them to algorithmic scoring, the infrastructure exists to protect everyone. The surveillance apparatus that governs consumer interactions isn’t technically necessary — it’s a business choice. Individual users simply lack the bargaining power to demand better.</p><h3><strong>Playing Doctor Without a License</strong></h3><p>Perhaps most troubling is what happens when these systems detect content related to mental health or self-harm. When I asked Google about their crisis intervention protocols, Gemini explained that flagged content can trigger human review and, if the threat seems credible and imminent, potentially contact with emergency services.</p><p>But here’s the catch: that 72-hour retention window maintains a connection to your Google Account, meaning they can identify you if they deem your private conversation concerning enough. Yet Google isn’t a licensed healthcare provider. They’re not bound by HIPAA. They have no clinical training requirements. The decision of what constitutes “real harm” is made by content moderators using algorithmic scoring systems, not mental health professionals.</p><p>This reveals a fundamental problem: AI companies are making health-related decisions about users while operating under privacy frameworks designed for advertising platforms. They’ve positioned themselves to assess mental health crises and contact authorities, but with none of the regulatory oversight we’d demand from actual healthcare providers.</p><h3><strong>The Escape Clauses</strong></h3><p>Both companies include escape clauses that render their specific timeframes nearly meaningless. OpenAI can retain flagged conversations indefinitely. Anthropic can keep both conversations and scores forever if they deem it “necessary to combat violations” of their usage policy — a determination they make entirely themselves, with no external oversight. Google reserves similar rights under their safety monitoring provisions.</p><p>They also reserve the right to retain “anonymized” versions of your data indefinitely for research purposes. But anonymization of behavioral data is notoriously unreliable. Your writing style, conversation patterns, and topic preferences can be as identifying as a fingerprint, especially when tracked over time.</p><p>The result is privacy theater: companies offer the appearance of user control while building systems designed to retain maximum flexibility for themselves. Delete buttons become digital sleight of hand — moving data from user-visible folders to corporate archives that never empty. Privacy settings become suggestions. User choice becomes corporate convenience dressed up in the language of safety.</p><h3><strong>The Future We’re Building</strong></h3><p>This safety-over-privacy mindset is understandable for platforms serving millions of users. But it raises fundamental questions about proportionality and consent. If a user has explicitly disabled history, opted out of training, and submitted no abuse reports, should their behavioral patterns still be algorithmically scored and stored for years?</p><p>The current approach flips the burden of proof. Instead of companies needing to justify why they should retain user data, users must trust that retention is necessary even when they’ve explicitly opted out. It’s surveillance by default, deletion by exception.</p><p>Most users discover these policies only after they’ve already shared sensitive information — private thoughts, personal struggles. The platforms bet that convenience will outweigh privacy concerns, and they’re usually right. By the time you read the fine print, your digital shadow is already being cast.</p><p>This dynamic will only intensify as AI becomes more capable and ubiquitous. We’re rapidly moving toward a world where AI assistants are integrated into every device, every app, every aspect of our digital lives. They’ll help us write emails, plan our schedules, manage our health, make financial decisions, and process our most intimate thoughts. The more helpful they become, the more we’ll rely on them — and the more comprehensive their surveillance will grow.</p><p>Today’s privacy violations are just the preview. Tomorrow’s AI assistants will know not just what we type to them, but how we work, how we think, how we make decisions, and how we relate to others. If we don’t establish proper privacy frameworks now, we’re building the infrastructure for the most comprehensive surveillance system in human history — one that we’ll willingly participate in because it’s genuinely useful.</p><p>While this investigation focuses on US companies, similar privacy gaps exist globally as AI systems spread worldwide, often with even weaker regulatory oversight than what exists in America and Europe.</p><h3><strong>What Real Privacy Would Look Like</strong></h3><p>So what would truly private AI look like?</p><p>While some smaller AI models can now run locally on powerful personal computers, the most capable systems require computational resources far beyond what average users can access. This means most people will continue relying on cloud-based AI services, making strong privacy protections essential rather than optional.</p><p>Moreover, AI use is rapidly becoming mandatory rather than voluntary. Jobs increasingly require familiarity with AI tools, students need them for research and writing, and professionals across industries are expected to integrate AI into their workflows. The choice to avoid AI entirely is disappearing, making the privacy stakes even higher.</p><p>A truly private AI system would require explicit, informed consent for every specific use of user data, with clear explanations of retention periods and review processes. It would delete data immediately after processing unless retention is legally mandated or explicitly requested by users. It would notify users any time their content is reviewed by humans or flagged by automated systems. It would ensure that memory deletion means the model won’t behave like it still knows you.</p><p>Most importantly, it would flip the default from retention to deletion. Instead of keeping everything “just in case,” it would delete everything unless there’s a specific, justified reason to keep it. Privacy would be the starting point, not an afterthought.</p><p>When AI systems start performing functions that resemble healthcare — crisis intervention, psychological assessment, decisions about contacting authorities — they should be held to healthcare-level privacy standards, with explicit consent requirements, strict de-identification protocols, and regulatory oversight by qualified professionals.</p><p>The contrast with actual medical research is stark. When I access patient data as a researcher, even fully de-identified information requires IRB approval, institutional oversight, and specific consent for each research use. HIPAA sets technical standards for de-identification that go far beyond corporate promises of “anonymization.” Any risk of re-identification triggers additional consent requirements. Yet AI companies routinely collect what would be considered protected health information — mental health discussions, medical symptoms, treatment questions — and subject it to algorithmic analysis and human review with none of these protections.</p><p>Current AI systems aren’t malicious, but they’re designed with company interests as the primary consideration and user privacy as a secondary concern. The result is infrastructure that can feel indistinguishable from surveillance, even when that’s not the intent.</p><h3><strong>The Path Forward</strong></h3><p>If AI is going to become a trusted collaborator — in work, writing, creativity, or decision-making — it has to be held to the same standards we expect from human collaborators: honest boundaries, informed consent, and the ability to truly forget on command.</p><p>Otherwise, these aren’t really assistants. They’re very polite data collection systems with hidden archives — not delete buttons. And in a world where our most intimate thoughts increasingly flow through digital intermediaries, that distinction matters more than any company wants to admit.</p><p>The future of AI isn’t just about making these systems smarter or more capable. It’s about making them trustworthy. And trust requires more than just good intentions — it requires systems designed from the ground up to respect human autonomy, even when that’s inconvenient for the companies building them.</p><p>Until then, when an AI tells you it has forgotten you, remember: it might not be lying, but it’s probably not telling the whole truth either.</p><p>The time to demand change is now, while these systems are still being designed rather than after they’ve become indispensable. Contact your representatives. Support organizations pushing for AI regulation. Most importantly, read the privacy policies before you share your thoughts with systems that promise to help but are designed to remember.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f62b74b42712" width="1" height="1" alt=""><hr><p><a href="https://meetcyber.net/the-privacy-theater-how-ai-companies-built-surveillance-systems-disguised-as-assistants-f62b74b42712">The Privacy Theater: How AI Companies Built Surveillance Systems Disguised as Assistants</a> was originally published in <a href="https://meetcyber.net/">MeetCyber</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>