In the first Anthropic Public Record, a nationally representative survey of 51,993 Americans, 56% said they were worried about cognitive dependency—defined in the study as AI integration leaving people unable to think for themselves.

Only job loss, at 64%, ranked higher. Cognitive dependency was more widely reported than concern about misinformation, criminal use, surveillance or AI systems going rogue.

This is a significant shift in the public conversation. Concern about the human effects of AI is no longer confined to specialist research, education or technology policy. It has become a mainstream question.

But Anthropic’s findings also reveal something more complicated: the people most worried about cognitive dependency were not necessarily those who expected the greatest disruption if AI disappeared.

That difference between what people fear, how they perceive their own use and what their behaviour may reveal deserves closer attention.

Cognitive dependency is now a mainstream concern

Anthropic commissioned YouGov to conduct the survey in November and December 2025. The sample was weighted to US Census benchmarks and included internet users aged 16 and over across all 50 states, the District of Columbia and Puerto Rico.

The study found broad agreement across political, geographic and educational groups. Americans were hopeful about AI’s potential to cure disease and support people with disabilities, while also expressing concern about job displacement, cognitive dependency and misinformation.

The prominence of cognitive dependency is especially notable because it is personal. Job loss and misinformation are often discussed as social or economic risks. Cognitive dependency asks a more intimate question:

If AI becomes embedded in everyday thinking, could people gradually become less able—or less willing—to think without it?

Cognitive dependency is part of a wider question about human capability in the age of AI: whether repeated delegation changes how much critical thinking, human judgement, verification and independent reasoning people continue to practise for themselves.

Anthropic’s separate qualitative research points in the same direction. In an international study of more than 80,000 Claude users, 21.9% raised concerns related to autonomy and agency, while 16.3% raised concerns about cognitive atrophy, including over-reliance, declining critical thinking and reduced practice of human skills.

These studies use different methods and their percentages are not directly comparable. Together, however, they show that concern about AI’s influence on independent thought is appearing across both the general public and existing AI users.

Concern is not the same as dependency

The most revealing part of the Public Record may be the gap Anthropic found between concern and anticipated disruption.

To explore whether people might already be experiencing dependency, respondents were asked how disrupted they would feel if AI became unavailable the next day.

Among the 56% who were worried about cognitive dependency, only roughly one-fifth said losing AI would cause significant disruption. Among those who were not worried about dependency, approximately one-third expected significant disruption.

Anthropic concluded that cognitive dependency currently appears to be mostly an anticipatory fear. That is a reasonable interpretation of the measure used: many people fear a future loss of independent thought without reporting that they presently need AI to function.

But the result also illustrates an important measurement problem. Concern about dependency and behavioural reliance are not the same thing.

Someone can be highly alert to the risk while using AI selectively. Another person can feel entirely comfortable with their use while AI has already become the default starting point for research, writing, problem-solving or decisions.

Expected disruption is not itself proof of cognitive dependency. AI may be deeply useful without weakening human capability, just as losing access to search engines, email or a smartphone would disrupt modern work. Nevertheless, the mismatch shows why asking people whether they are worried—or even whether they consider themselves dependent—cannot provide a complete view of their behaviour.

Why our own AI behaviour can be difficult to see

AI use is often measured through frequency: whether someone uses it occasionally, weekly or every day. Frequency is useful, but it cannot reveal what is happening inside the interaction.

Two people can use AI every day and develop very different patterns.

One may form an initial view, use AI to challenge it, verify important claims and retain clear authorship of the final judgement. Another may routinely begin with AI, allow it to frame the problem, accept plausible answers with little checking and gradually practise less independent reasoning.

Both are daily users. Their frequency is identical; their behaviour is not.

The Anthropic survey found that concern about cognitive dependency fell as AI use increased. Among Americans who never used AI at work, 62% were worried about dependency. Among daily workplace users, the figure was 46%.

There are several plausible explanations. Experience may help people understand AI’s limitations and use it more deliberately. People who are comfortable with new technology may also begin with fewer concerns. But familiarity can make a behaviour feel normal, and useful routines can become difficult to examine from inside the routine itself.

The survey does not establish which explanation is responsible. It does show why low concern should not automatically be interpreted as low reliance—and why high concern should not be treated as evidence that someone is already dependent.

AI dependency cannot be understood through one question

The phrase cognitive dependency can describe several different changes, many of which involve some form of cognitive offloading—shifting parts of thinking, remembering, evaluating or deciding from the person to an external system:

  • turning to AI before attempting a problem independently;
  • becoming less confident in personal judgement when AI disagrees;
  • transferring more decisions to AI-generated recommendations;
  • reducing verification as trust and familiarity increase;
  • relying on AI to regulate uncertainty, discomfort or emotional strain;
  • practising particular cognitive or creative skills less often;
  • experiencing significant difficulty when AI is unavailable.

These behaviours can interact, but they are not interchangeable. Frequent reliance does not necessarily mean low agency. High trust can coexist with strong verification. A close thinking partnership with AI can extend human capability, or it can quietly reduce independent effort, depending on how that relationship operates.

This is why HCI examines AI behaviour across connected dimensions—including Reliance, Trust, Verification, Decision Delegation, Human Agency and Thought Partnership—rather than attempting to classify a person as simply dependent or not dependent. HCI’s AI Reliance Benchmark explores this distinction in greater depth.

The purpose is not to turn a complex relationship into a clinical label. It is to make the pattern underneath everyday use easier to see.

What behavioural comparison can add

Self-reflection remains valuable, but it lacks a reference point.

A person may know they consult AI before important decisions but have no idea whether this is typical among other frequent users. They may believe they verify AI regularly without recognising that their checking has declined as the technology has become more familiar. They may use AI intensively while retaining unusually strong control over final decisions.

Behavioural comparison cannot determine one universally correct way to use AI. Nor does an uncommon behaviour automatically mean that it is harmful. What comparison can provide is context:

  • Is this behaviour broadly typical or unusual among other participants?
  • Does reported reliance align with the way the person perceives their own use?
  • Is high AI involvement accompanied by continued verification and decision ownership?
  • Where does AI appear to extend human capability, and where might it be replacing human practice?

These are more useful questions than asking whether someone simply uses AI “too much.”

Independent human measurement is becoming more important

The Public Record also found that only 15% of Americans trusted AI companies to make decisions about how AI is developed and used. Independent experts received the highest trust of any institution measured, at 43%.

Anthropic itself concluded that the direction of AI should not be determined only by the companies building it.

That principle applies to measurement as well. AI companies hold valuable information about model performance and platform use. But society also needs independent research focused on the person interacting with the technology: how people adapt, what behaviours are becoming normal and which human capabilities remain active as AI becomes more capable.

Technical capability tells us what AI can do. Usage data tells us what people ask it to do. Human behavioural research helps examine what may be changing around the interaction.

All three perspectives matter.

From public concern to personal clarity

Anthropic’s findings establish that cognitive dependency is not a fringe fear. More than half of the Americans surveyed expressed concern about the possibility of AI leaving people less able to think for themselves.

But the findings also show why concern alone is insufficient. People who worry about dependency may not rely heavily on AI. People who feel unconcerned may still anticipate substantial disruption without it. And two people with the same frequency of use may differ fundamentally in how much they verify, delegate, question and retain ownership of their thinking.

The next question is therefore not only whether society should be concerned about cognitive dependency.

It is whether each of us can clearly recognise the pattern forming around our own use.

How does your AI behaviour compare?

The Human Clarity Institute’s AI Identity & Behaviour Assessment examines 39 reported behaviours across nine dimensions, including Trust, Reliance, Verification, Decision Delegation, Human Agency and Thought Partnership.

It compares your responses with HCI participant benchmark data to show where your pattern appears broadly typical, where it differs and which behaviours may be difficult to recognise through instinct alone.

Approximately five minutes · Free personalised results · No account required

Compare your AI behaviour →

The HCI assessment is designed for personal insight and behavioural comparison. It is not a psychological, medical or mental health assessment, diagnosis or clinical dependence test.

Source and methodology note

The principal external source for this article is Anthropic’s Results from the first Anthropic Public Record, published 12 June 2026. The survey was conducted online by YouGov between 1 November and 11 December 2025 among 51,993 US residents aged 16 and over. Results were weighted by state, age, gender, education and race or ethnicity. Anthropic reports a national margin of sampling error of ±0.6 percentage points at the 95% confidence level. The findings describe reported attitudes and anticipated disruption; they do not establish the prevalence of clinical or behavioural dependency.

Additional context is drawn from Anthropic’s qualitative study, What 81,000 people want from AI. That research used voluntary interviews with Claude users and should not be interpreted as nationally or globally representative.