Ego Depletion by Design: How Chat Interfaces Systematically Erode Human Oversight
A few weeks ago, I caught myself doing something I should have known better than to do.
An AI assistant had just completed a set of tasks on my behalf. It summarised what it had done. I read the summary the way you read a “your order has been placed” confirmation email: eyes moving, brain not really engaged. I approved. I moved on. It was only later, when something felt slightly off, that I went back to check what had actually happened. And what had actually happened was not exactly what I had assumed.
I am a product designer. I work on AI tools professionally. I have written about the importance of designing for meaningful human oversight. I know what automation bias is. I know what decision fatigue is. And I still did it.
I want to be clear about what that means. It does not mean I was being careless. It means a design pattern did exactly what it was built to do: it trained me, through repeated low-consequence interactions, to treat AI outputs as default-accept. My brain had learned “reviewing this wastes cognitive resources and nothing bad has come from not reviewing” and it acted accordingly. That is not a personal failure. That is behavioral science, operating precisely as the research describes.
If you are designing AI products, what just happened to me is happening to your users. And the design choices your team made are the reason.
The science behind the behaviour
In the 1990s, the social psychologist Roy Baumeister proposed a theory he called ego depletion. The core argument: self-regulation, including the capacity to make deliberate, effortful decisions, draws on a finite cognitive resource. Each act of willpower, each conscious choice, each moment of critical engagement depletes that resource incrementally. As it depletes, the quality of subsequent decisions degrades. You do not make worse decisions because you become less intelligent. You make worse decisions because your brain shifts toward the lowest-effort strategy available.
Baumeister and colleagues demonstrated this across a striking range of behaviours: people who suppressed their emotions showed less persistence on subsequent physical tasks; people who made a series of product choices became less likely to engage critically with a later unrelated decision. The depletion transferred across domains. The resource was general, not specific to any single type of decision.
The most cited real-world illustration of this comes from a 2011 study of Israeli judges making parole decisions. Judges who had recently eaten a meal granted parole at roughly 65% of cases. As the session wore on without a break, that rate dropped toward zero, where the default outcome (denial) became the near-universal response. The judges were not becoming more punitive. They were running low on the cognitive fuel required to deviate from the default, and so they stopped deviating. Later analyses of this specific study have added nuance, but the broader pattern it illustrates (decision quality degrades with volume and time, and defaults absorb the slack) is supported across many research streams and is not seriously contested.
Alongside ego depletion sits habituation, a related but distinct mechanism. Habituation is the brain’s efficiency response to repeated stimuli: when something happens consistently and the consequences are neutral or positive, the brain gradually stops allocating attentional resources to it. It is not laziness. It is the perceptual system doing exactly what it evolved to do, filtering out what it has learned does not require conscious processing.
These two mechanisms working together produce what I observed in myself: a trained pattern of approval that felt like efficiency but was in fact the systematic dismantling of critical engagement. I had been approving AI outputs repeatedly. Most of them were fine. My brain filed “reviewing AI outputs” as a low-priority process and began bypassing it. By the time the output that was not fine arrived, the bypassing was already habituated.
We have watched this happen before
The design community has seen this pattern play out at scale. Twice in recent memory it produced policy-level crises, and both times the root cause was the same: a consent or review mechanism deployed at a volume and friction level that made meaningful engagement behaviorally impossible.
The first is the Terms of Service problem. In a landmark 2008 study, Lorrie Faith Cranor and Aleecia McDonald estimated that if a typical internet user read every terms of service agreement they were asked to accept in a year, it would take 76 working days. Nobody reads them. The consent mechanism became legal theatre. Designers and lawyers knew this and continued designing it, because the mechanism served compliance purposes, not comprehension purposes. The user’s signature was collected. Whether it reflected genuine understanding was not the product’s problem.
The second is GDPR cookie consent banners. In 2018, a regulation designed to give Europeans meaningful control over their personal data produced one of the most widely studied failures of informed consent in digital history. Within months of rollout, users were clicking “Accept All” in under a second, without reading. Researchers documented the emergence of what they called consent fatigue: repeated exposure to low-stakes consent requests, combined with the friction asymmetry between accepting and declining, trained users to dismiss without engaging. The more important the decision being made, the less meaningful the process through which it was being made.
User-centered design did not ignore these patterns. It responded to them with a set of principles that took years to develop and refine: progressive disclosure, which limits decision volume by surfacing complexity only when needed; sensible defaults, which reduce the number of decisions that require conscious engagement; and graduated friction for consequential actions. That last principle was the hardest-won. Destructive actions got red buttons and confirmation dialogs. Irreversible actions required the user to type a confirming phrase. Settings with significant downstream consequences were separated architecturally from routine settings. The design community learned, through iteration and research, that not all decisions should feel the same. Important decisions need interface weight that signals their importance.
That principle is now being abandoned by the AI industry, at speed, and the reason is adoption mechanics.
Why chat was the wrong choice for this
The decision to use conversational chat as the primary interaction paradigm for AI was not a decision about decision quality. It was a decision about adoption. Chat is familiar. It requires no onboarding. It feels approachable and human. It gets users into the product faster and with less resistance than any structured interface would. These are real advantages, and for the task of getting millions of people to try AI tools for the first time, they were probably the right optimisation.
The problem is that the properties that make chat a good adoption mechanism are precisely what make it dangerous for consequential interactions.
When you engage with something in natural language, your brain processes it using the same social cognition machinery it uses with humans. In social interactions, agreement is the default. Disagreeing with a conversational partner requires more cognitive effort and social override than agreeing. You have been practicing social agreement your entire life. It is one of the most deeply habituated patterns of human behaviour. Chat AI imports that habituated tendency and applies it to a system that is not making conversation. It is proposing actions.
There is another layer. AI-generated text is extremely fluent. It is smooth, grammatically correct, confidently expressed, and easy to read. Processing fluency, the cognitive ease with which a statement is read and understood, has been shown in research to correlate with perceived truth and competence. Text that is hard to read triggers slower, more critical processing. Text that is easy to read accelerates agreement. AI responses are among the most fluent text most users will encounter in any interface. The very quality that makes them pleasant to read is reducing the critical scrutiny they receive.
Get Abdul-Rashid Lansah Adam’s stories in your inbox
Join Medium for free to get updates from this writer.
And then there is the confirmation momentum that builds over time. The foot-in-the-door effect, documented in social psychology, shows that small prior commitments prime people for larger subsequent ones. In AI chat, every low-stakes approval (“yes, file that email,” “yes, reschedule that meeting,” “yes, send that summary”) is building a behavioural model of “this system is usually right and agreeing with it is the efficient path.” By the time a higher-stakes action arrives in the same conversational thread, the momentum toward approval is already established.
The buried action problem
There is a structural design failure at the heart of chat-based AI that goes beyond any individual interaction: there is no persistent state visible to the user.
In a chat thread of forty messages, an AI might have taken fifteen actions. Those actions are not collected anywhere. There is no summary screen. There is no view of “here is the current state of the world this AI has created.” Each action was embedded in the conversational moment where it was proposed and approved, surrounded by the context of that exchange, and then it scrolled away. The conversation moved forward. The actions became part of the past.
This is the exact opposite of what UCD established for consequential interactions. A bank transfer shows you a confirmation screen with the complete transaction before you commit. A form submission shows you what you are changing before the change is made. A settings screen shows you the current state alongside what will be different. These patterns exist because decades of research showed that surfacing complete state at the moment of decision improves decision quality. The review moment needs a clear representation of what is being decided, not a conversational reconstruction of it buried in a thread.
Chat moves fast. That speed is part of its appeal. But speed is the enemy of deliberation, and deliberation is what makes consequential decisions consequential. When an AI can take an action in the same conversational beat as answering a trivia question, and both interactions look identical in the interface, the interface is communicating that they are the same kind of thing. They are not.
Where this becomes dangerous
The failure modes described above are recoverable in low-stakes contexts. An email filed incorrectly can be moved. A calendar invitation sent to the wrong time can be corrected. The costs are friction and embarrassment, not harm.
Banking and healthcare are different. These are domains that evolved their interaction design specifically because the cost of decision errors is asymmetric and, in many cases, irreversible. SWIFT transfers have explicit confirmation steps with full transaction detail. Prescription entry systems have override warnings that require documented acknowledgment. Insurance claim approvals carry audit trails visible to the user. These are not bureaucratic friction for its own sake. They are UCD responding, over decades, to the documented cost of what happens when consequential decisions are made carelessly.
An AI banking assistant operating through a chat interface is asking a user to approve a transfer the same way they would approve a restaurant suggestion. The interface provides no signal that these are categorically different kinds of decisions. And because the pattern of approval-without-scrutiny has been trained through dozens of low-stakes interactions (find this transaction, summarise my spending, categorise this receipt), it is already in place when the consequential action arrives.
Healthcare is more insidious because the temporal gap between error and consequence is longer. An AI health assistant that books the wrong appointment, files the wrong insurance claim, or surfaces a medication interaction warning that a user has been habituated to dismiss as just another message in the thread: the cost of that interaction may not become visible for weeks or months. The behavioral pattern was trained in a context where consequences were immediate and mild. It gets deployed in a context where consequences are delayed and serious. The user has no way of knowing the threshold has changed. The interface gave them no signal.
What this means for design
The core argument is straightforward to state even if it is costly to build: the interaction model for an AI action should be calibrated to the consequence of that action, not to the convenience of the chat paradigm.
This does not mean abandoning chat. It means refusing to let chat absorb everything. For exploratory, reversible, low-stakes interactions, chat is appropriate. The AI suggests, the user engages, the friction is low and the stakes match. But when an action crosses a consequence threshold — when it is irreversible, when it involves financial or personal data, when it has downstream effects the user cannot easily undo, the AI should leave the chat paradigm entirely and surface that action in a form that matches its weight.
That means a dedicated decision screen, not a chat bubble. Complete state visible before commitment. Explicit acknowledgment rather than conversational continuation. And, critically, a persistent record of consequential actions that is not buried in the scroll of conversation but is accessible, reviewable, and surfaced proactively.
The behavioral logic is the same principle UCD already learned: graduated friction for graduated consequence. The AI industry needs to apply that principle to its own paradigm, which it has so far declined to examine critically because the adoption numbers are good and the harms are not yet catastrophic enough to be visible.
The uncomfortable close
Roy Baumeister’s ego depletion research was conducted in laboratories, with participants who knew they were in a study and were making decisions about things that did not matter much. The depletion he documented was real and significant even in that context.
We are now running the same experiment on hundreds of millions of people, with AI systems that are taking real actions in their lives, at volumes and speeds that would have been impossible to study in a lab. The habituation is building. The approval patterns are forming. The depleted-default is being trained in.
Most of the designers building these products know about decision fatigue. Many of them have written about it. Some of them, like me, have experienced it in their own use of the tools they build.
And the chat interface is still shipping.
We designed consent fatigue once, with cookie banners, and it took a regulatory intervention to force the design community to reckon with what it had built. We designed ToS fatigue, and nobody reckoned with it at all. We are now designing AI approval fatigue, at a scale and with stakes that exceed both of those. The difference this time is that the system on the other side of the approval is not just collecting your data. It is taking actions in the world on your behalf.
The moment that most matters to get the design right is not when the product is easy to use. It is when the decision is hard to reverse.
Abdul-Rashid Lansah Adam is a Senior Product Designer specialising in AI-native products and behavioural design. He writes about what AI is breaking and what designers should build next at uxlansah.com. Read his related piece: a UX framework for human-AI oversight that actually works.
