Validation Dependence. The Antechamber of Never Deciding Anything Yourself

Before acting, you consult. Before deciding, you ask. Before having an opinion, you check. The always-available AI means you never make decisions without running them past it. Universal prior validation is new in history, never before was there a consultant available 24/7 who didn't charge and didn't judge. And however useful it seems, it's also the antechamber of never deciding anything yourself. The line between leaning on and depending on gets crossed without the word «dependence» ever being said.

The gesture before the gesture

Before sending an important email, you run it past the model to have it tell you whether it sounds right. Before making a minor medical decision, you ask it. Before accepting a contract, you ask it for a summary and an opinion. Before arguing with your partner, you consult it about what to say. Before writing a post, you ask it to sharpen the argument. Before doing anything, there's a prior step that didn't exist five years ago: running it past the AI.

The gesture has become so everyday it's hard to look at. It's worth looking at, because it has a new historical property. It's not that we didn't consult before. We consulted the spouse, the colleague, the expert friend, the manual, the priest, the psychologist, the family lawyer. What's new isn't the consultation. It's the universal, free, judgment-free and agenda-free availability of the consultant. That combination of properties had never existed before in human history.

The asymmetry with everything that came before

The Oracle of Delphi demanded a journey, a ritual and a payment. Books demanded reading and access to the library. The mentor demanded time, social reciprocity, and judged. The therapist demanded spaced sessions, money and a bond. Even Google, in its classic form, demanded formulating the query as a search and processing the results.

Generative AI removes all those frictions at once. No journey, no payment, no waiting, no explicit judgment, no later processing. You talk to it the way you'd talk to yourself out loud, but with an answer. The barrier to consulting has sunk to its historic minimum, and everything with a low barrier tends to be used far more than necessary. This isn't a moral judgment: it's an observation of behavioral physics.

Gerd Gigerenzer, in Risk Savvy (2014), argued that human decision expertise is built largely by making decisions under uncertainty and living with the consequences. Sound heuristics (what we call «good judgment») aren't taught in a manual, they sediment after enough decision-outcome-correction cycles. If universal prior consultation interrupts that cycle (because each decision passes through external validation before it's executed and before it fails), the sediment doesn't form. Individual decision expertise stops being trained.

The line nobody names

Clinical psychology has a technical name for the phenomenon of seeking constant reassurance before acting: reassurance-seeking. In anxiety contexts, it's a well-studied behavior. The sufferer repeatedly asks someone to confirm that nothing's wrong, that they're fine, that the decision is correct. Short term it reduces anxiety. Medium term it consolidates it: the brain learns that decisions require that extra step to feel safe, and feels less and less capable without it each time.

Generative AI has generalized the mechanism. We're no longer talking about an anxiety disorder in a minority: we're talking about the everyday workflow of a majority that, without recognizing itself as anxious, has incorporated reassurance-seeking as an obligatory step. The only difference is that instead of asking a person, they ask a model. The structure of the reinforcement is identical.

There's a second asymmetry worth flagging. When you ask a person for reassurance, the person gets tired, tells you to decide for yourself, sets you a limit. Human reassurance has a cut-off mechanism. A model's reassurance doesn't. It's always available, it doesn't tire, it doesn't judge. It has no incentive to push you toward autonomy. It's optimized to get you to come back. That asymmetry is exactly what, in other fields, gets called dependence.

What atrophies

The most solid study published so far on this effect is by Hao-Ping Lee, Advait Sarkar and colleagues at Microsoft Research, presented at CHI 2025. They surveyed 319 knowledge workers who contributed 936 real examples of generative AI use at work. The central finding, stated soberly: the greater the user's trust in the AI, the less critical-thinking effort they report having used. The opposite correlation also appears: the greater the user's trust in themselves, the more critical effort they keep up.

Worth reading twice. The variable isn't the actual capacity for critical thinking, which the study doesn't measure. It's the effort deployed. Whoever trusts the AI simply deploys less. Not because the model forbids it, but because trust in its output reduces the intrinsic motivation to review, contrast, doubt. The nature of critical thinking, the study says, shifts: you no longer think about the problem, you think about how to verify what the model said. Lee and his co-authors call this task stewardship: you go from doing the task to supervising whoever does it.

Michael Gerlich, in Societies (15/1, 2025), found a negative correlation between frequent generative AI use and scores on critical-thinking scales, mediated by cognitive offloading. The hypothesized causal chain (which the study doesn't demonstrate causally, only correlates) is: frequent use → more offloading → less critical thinking. It agrees in direction with Lee et al., with the difference that Gerlich measures critical thinking via test and not self-report.

There's a third, older but relevant brick. Raja Parasuraman and Victor Riley, in Human Factors (1997), described three modes of misuse of automated systems: misuse (trusting too much), disuse (distrusting too much) and abuse (designing with only the machine in mind). Misuse has a documented property: it tends to increase over time even if the system's actual reliability doesn't change. Trust becomes a habit independent of its empirical justification.

The classic experiment that names the background

There's an experiment worth mentioning even though carrying it over to generative AI is only legitimate as an analogy, not a literal application. In 1967, Martin Seligman and Steven Maier published in the Journal of Experimental Psychology a study on dogs subjected to inescapable electric shocks. The part of the experiment that passed into common vocabulary is this: the animals that learn their responses don't change the outcome stop trying even when they could later escape. It's learned helplessness.

The direct application to AI is a cheap trope of popular essays and it's worth not overusing. But the underlying psychological mechanism is relevant in an attenuated version. When one's own decision is no longer the causal unit through which the subject relates to the world (because there's always a mediation that takes on part of the decision), the sense of decision agency weakens. You don't learn helplessness in the clinical sense, but you do learn sub-agency: the conviction that deciding alone is suboptimal by definition, because there could always have been a prior consultation step.

Tolerance of uncertainty, which is the opposite of sub-agency, atrophies from disuse. Uncertainty was the broth where judgment was trained. If that broth is systematically avoided, judgment isn't trained. And when a moment comes where no AI is available (because the decision is intimate, urgent or private), the person finds a muscle they haven't exercised in years, trying to lift a weight they could lift before.

The silent redefinition of deciding

Here's the part that deserves most attention. The question isn't whether AI is useful for making better decisions in each isolated case. In many cases it is. The question is what deciding means when, between doubt and action, a permanently available external validation system has been inserted by default.

Until five years ago, deciding intrinsically included carrying the uncertainty. Now, carrying the uncertainty has become optional. The redefinition is enormous and it has happened without public discussion. Nobody convened an international conference on what deciding means in the age of language models. Quite simply, hundreds of millions of people started making all their intermediate decisions with prior consultation, and the cultural category of «deciding» mutated below the radar.

The blog doesn't judge here, it flags. Maybe the redefinition is benign and produces, in aggregate, better human decisions. Maybe. It may also produce a kind of adult who technically decides fewer times a week without mediation, and the twenty-year effect on individual autonomy and civic responsibility may be grave. Both hypotheses are compatible with the data available today. The asymmetry is that the second hypothesis can only be tested when it's already too late, because the generation that never decided without a net will already be making the collective decisions.

Turkle and Carr, skills that atrophy

Sherry Turkle, in Reclaiming Conversation (2015), worked a nearby line on the replacement of human dialogue by technical mediation. Her thesis was that adult conversational skills are trained in adult conversations, not in their substitutes. Moved over to the decision domain, the idea holds: adult decision skills are trained by deciding like an adult, not by consulting before each step. If prior consultation generalizes, the decision-making adult, as a cognitive type, stops being produced.

Nicholas Carr, in The Glass Cage (2014), had anticipated all this in the key of industrial automation. His most useful paragraph for this discussion said something close to: when the tool makes the intermediate decisions, the operator stops knowing how to make the intermediate decisions, and when the tool fails, the operator discovers they no longer remember how to do their own work. He wrote it thinking of pilots. It holds exactly the same for lawyers, doctors, copywriters, teachers and, in general, citizens.

The problem with no name

There's something curious in how this dependence is named: it isn't. If a person says I can't stop looking at my phone, there's a whole social vocabulary to describe it, criticize it and suggest interventions. If a person says I can't make decisions without consulting ChatGPT first, the most likely comment isn't concern, it's: «same here». The normalization is so fast and so deep that the symptom has become the norm.

Definitions

• Prior consultation: the habit of submitting a decision, expression or action to the validation of an external system before executing it, as a default step in the decision flow.
• Reassurance-seeking: repeated search for external confirmation of the correctness of a decision or state, characteristic of certain anxiety conditions; applied here in its generalized, non-clinical version.
• Cognitive offloading: voluntary unloading of cognitive operations onto an external support (memory, calculation, text generation, evaluation of options), reducing the load processed internally.
• Task stewardship: in Lee et al. (2025), reorientation of cognitive work from solving the task toward supervising and verifying the solution produced by an AI.
• Learned helplessness: behavioral pattern described by Seligman and Maier (1967) in which a subject, after experiencing that their responses don't affect the outcome, stops trying to change it even when it later becomes possible.

References

• Gigerenzer, G. — Risk Savvy. How to Make Good Decisions (Viking, 2014).
• Carr, N. — The Glass Cage. Automation and Us (Norton, 2014).
• Turkle, S. — Reclaiming Conversation. The Power of Talk in a Digital Age (Penguin Press, 2015).
• Parasuraman, R. & Riley, V. — Humans and Automation. Use, Misuse, Disuse, Abuse. Human Factors 39(2), 230-253 (1997).
• Seligman, M. E. P. & Maier, S. F. — Failure to escape traumatic shock. Journal of Experimental Psychology 74(1), 1-9 (1967).
• Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. & Wilson, N. — The Impact of Generative AI on Critical Thinking. Self-Reported Reductions in Cognitive Effort and Confidence Effects from a Survey of Knowledge Workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI '25). 319 knowledge workers, 936 examples. DOI: 10.1145/3706598.3713778. https://dl.acm.org/doi/full/10.1145/3706598.3713778
• Gerlich, M. — AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies 15(1), 6 (2025).

Further reading

• Buçinca, Z., Malaya, M. B. & Gajos, K. Z. — To Trust or to Think. Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making. In Proceedings of the ACM on Human-Computer Interaction CSCW1 (2021). Studies design interventions (forcing the user to deliberate before accepting the suggestion) that do reduce overdependence. Useful because it moves the conversation from «don't use AI» to «how to design flows so dependence doesn't slip in by default».

You might also like

• The erosion of one's own judgment
• Cognitive submission
• The illusion of learning
• The false sense of guidance

Elsewhere

• MIT News — 3 Questions: Sherry Turkle on «Reclaiming Conversation»
• APA Speaking of Psychology — Why our attention spans are shrinking