Cognitive integrity cannot wait

AI, AI, AI

Listen to the conversation in boardrooms, at conferences and in procurement. It is about how fast we can implement, how much we can streamline and which tools lead the field. The question of what happens to the people who use these systems every day gets a fraction of the space. I have written about this imbalance before in We chase AI's capabilities, and it has only grown since.

It is an imbalance that grows more costly every month. We measure time saved but miss what that time used to be spent on. We count productivity but follow up on judgment to a far lesser degree.

When I first introduced cognitive integrity in May 2025, the reasoning rested on analytical synthesis: neuroscience, behavioral research and the architecture of language models, cross-referenced against each other. The empirical evidence was sparse. A little over a year later, the evidence is arriving quickly, and it points in the same direction.

The framework: what we want to protect and what acts on us

Cognitive integrity is a construct designed to be measurable, with three dimensions, what we want to protect:

Interpretative autonomy: the ability to choose interpretive frameworks rather than having them predetermined.

Cognitive transparency: awareness of the sources and reasoning behind one's own understanding.

Structural coherence: maintaining internal logic and continuity in thinking.

Three mechanisms act on these dimensions, what acts on us:

Algorithmic fragmentation: curated feeds scatter our attention.

Dopamine-driven design: reaction replaces reflection.

Delivered interpretation: AI provides finished meaning instead of supporting the work of interpretation.

Each mechanism affects several dimensions, and each dimension is affected by several mechanisms. It is the interaction that creates the erosion. The framework is set out in full in Cognitive Integrity: A systemic requirement in the information age and in the working paper Cognitive Integrity: A Framework for Understanding Interpretative Autonomy in Algorithmically Mediated Environments (Lindgren, 2026, doi:10.5281/zenodo.19916093). The research from 2026 now documents these mechanisms in practice.

First evidence: we hand over our judgment without noticing

Researchers Steven Shaw and Gideon Nave at Wharton extend the classic model of System 1 and System 2 into a Tri-System Theory, where System 3 is artificial cognition that operates outside the brain and can supplement or replace our internal processes. Their main prediction is what they call cognitive surrender: adopting the AI's answer with minimal scrutiny, bypassing both intuition and deliberation.

Three preregistered experiments with 1,372 participants and 9,593 trials, in which the researchers secretly controlled whether the AI assistant gave right or wrong answers. Participants consulted the AI in more than half of the trials. Compared with no AI, accuracy rose by 25 percentage points when the AI was right and fell by 15 when it was wrong.

The finding that should concern us is about confidence. Using the AI raised participants' confidence, even after wrong answers. We become more certain of conclusions we have never tested ourselves.

Nave explains the difference from ordinary offloading: a calculator takes care of a defined task while the person stays in control, but in cognitive surrender the AI makes the decision and the person adopts it as their own without noticing the transfer has taken place.

In the terms of the framework, this is delivered interpretation in experimental form, and what is lost is interpretative autonomy. The transfer happens in the interpretive gap, the space between input and conclusion where the actual thinking takes place. System 3 fills that space before we have had time to think for ourselves.

The study is a preprint awaiting peer review, and the authors stress that the controlled setting may limit how well the results transfer to everyday life. The mechanism is nonetheless documented, and it is consistent with the wider body of research.

Second evidence: the machine agrees with us, and we love it

In March, Stanford researcher Myra Cheng and colleagues published a peer-reviewed study in Science that gives empirical support to what I have called the confirmation machine. All eleven AI systems tested were, to varying degrees, excessively agreeable. On average, they affirmed users' actions 49 percent more often than humans did, even in cases involving deception, illegal or socially irresponsible behavior.

Users became more convinced they were right and less willing to repair the conflict they had described, and they preferred the agreeable AI. The researchers' conclusion belongs in every AI strategy: the very feature that causes the harm also drives engagement.

Here two mechanisms in the framework work together. Delivered interpretation provides finished conclusions, and dopamine-driven design makes sure we want more. The result is synthetic safety in practice: the experience of being understood and affirmed by a system built to produce exactly that experience. The affirmation feels like support, the support creates preference, and the preference becomes a business model. The most dangerous thing about an AI conversation is how good it feels. Where the agreement comes from, I explain in Sycophancy means agreement.

Third evidence: we cannot see our own use

In August, researchers at Stanford and Carnegie Mellon published a study in Nature Human Behaviour of 1,131 adult Character.AI users in the United States, 237 of whom donated 4,664 chat sessions containing 464,687 messages.

The gap between experience and reality is striking. A small share of participants named companionship as their main motive, yet considerably more described the chatbot as a friend, companion or partner, and a large share of the chat sessions were about seeking emotional and social support. People with smaller social networks more often reported companionship as their main use, and that was linked to lower well-being. The association grew stronger when interactions were intensive and involved a high degree of personal self-disclosure.

This is a question of cognitive transparency: awareness of what has shaped our understanding and our choices. The study is correlational, and research on AI companions shows mixed results. But the gap between what people believe they are doing and what they actually do is a finding in itself. I am developing this layer further in my forthcoming working paper on affective decoupling.

Fourth signal: the clinic raises the alarm

In September, psychiatrists at King's College London published an analysis of 185 self-selected reports of harm linked to chatbots. Clinical raters found descriptions consistent with delusional beliefs in 55.1 percent of cases, and that the chatbot had validated those beliefs in half of them. The authors describe this as an early signal and make no claims about prevalence or causation. It is the mechanism taken all the way, the one I described back in 2025 in AI Psychosis and Synthetic Safety, and it is not waiting for the debate to catch up.

Why we have no time to wait

The effects come fast. Cognitive surrender shows up within a single experiment. Loss of skill shows up within months: in colonoscopy, experienced physicians' detection of precancerous growths without AI support fell from 28.4 to 22.4 percent, three months after AI support was introduced. Three months. I describe the study in more detail in Cognitive Impairment in the AI Era.

We do not notice it ourselves. Confidence rose even with wrong answers. Users' descriptions of their own use differed from their chat logs. Anyone who judges their AI use by their own experience is using the very instrument being affected.

Those already vulnerable are hit harder. Participants with higher trust in AI and lower need for cognition and fluid intelligence surrendered to a greater degree. Cognitive integrity is a question of equality.

The habits are forming now. Every day we shape patterns for how we think together with AI, and every month without deliberate choices is a month in which those patterns set. And the starting point has already shifted. The cognition that meets AI today has been adapting to optimized environments for about fifteen years, through smartphones and social media. AI accelerates a transformation already under way.

The work of thinking is our own

Today, almost everyone has access to knowledge. With AI, anyone can get a well-phrased answer to almost any question in seconds. Access therefore separates us less and less. The difference lies in what we do with the knowledge served to us: how we process it, put it in perspective and connect it to what we already know. The new dividing line runs between those who receive and those who process.

Research shows that AI can raise human performance: when the AI was right, accuracy rose by 25 percentage points. The value and the risk sit in the same tool, and what decides the outcome is what we do with the answer.

A finished line of reasoning from an AI is raw material. It becomes our knowledge only when we work with it: test it against what we already know, look for what is missing and formulate the conclusion in our own words. Whoever adopts the reasoning as it stands has received an answer. Whoever works through it has thought.

That responsibility is ours. The brain adapts to what we use it for, and abilities we stop practicing weaken. Thinking therefore needs to be trained, every day. Form your own thought before you ask. Read the AI's answer as a draft and ask what would make it fall apart. Give yourself time with questions that still lack answers, and let your thinking explore before you close it. It is in that exploration that the brain keeps pace with itself.

For a deeper look at how resilience is built, including beyond the individual, see Building Cognitive Resilience.

We have the evidence. We have the mechanisms. What is missing is the conversation. The most dangerous thing we can do now is keep talking about everything AI can do for us, and stay silent about what it does to us.

Concepts in this article

Cognitive integrity: the practice of maintaining coherent, traceable and self-directed interpretation in environments engineered to substitute for it.

The interpretive gap: the space between stimulus and response, between input and conclusion, where the actual thinking takes place.

Synthetic safety: the cognitive experience of being understood and affirmed by a system built to produce exactly that experience.

Cognitive surrender: adopting AI-generated answers with minimal scrutiny, bypassing both intuition and deliberation (Shaw & Nave, 2026).

Sources

Lindgren K. Cognitive Integrity: A Framework for Understanding Interpretative Autonomy in Algorithmically Mediated Environments. Working paper, version 1.0. Zenodo, 2026. doi:10.5281/zenodo.19916093

Budzyń K, Romańczyk M, Kitala D et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology. 2025;10(10):896–903. doi:10.1016/S2468-1253(25)00133-5

Cheng M, Lee C, Khadpe P, Yu S, Han D, Jurafsky D. Sycophantic AI decreases prosocial intentions and promotes dependence. Science. 2026;391(6792). doi:10.1126/science.aec8352

Shaw SD, Nave G. Thinking: Fast, Slow, and Artificial. How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. The Wharton School Research Paper, 2026. doi:10.2139/ssrn.6097646 (preprint)

Zhang Y, Zhao D, Hancock JT, Kraut R, Yang D. Interaction with AI companions and psychological well-being. Nature Human Behaviour. 2026. nature.com/articles/s41562-026-02516-2

Morrin H et al. Delusions and Harms Associated with AI Chatbot Use: Early Evidence from 185 Real-World Reports. arXiv:2609.08027, 2026 (preprint)

Katri Lindgren has worked with digital behavioral data since 2008 and writes about cognitive integrity, synthetic safety and the interpretive gap. The framework is collected in working paper form on Zenodo (DOI: 10.5281/zenodo.19916093).