Human Behaviour and Perception in the Age of AI

Routine adoption alongside public ambivalence

What has happened is less dramatic, and perhaps more consequential, than a sudden replacement of human intelligence. AI has become woven into ordinary searching, advice, writing and conversation. A 2026 Pew survey reported that 44% of U.S. adults had used ChatGPT, while 24% used AI chatbots daily; 51% did not use them. [c1]

Source 1

And that routine use sits beside real unease. In a separate Pew survey, half of Americans said they were more concerned than excited about increased AI use in daily life, while 57% rated the societal risks as high. So adoption and apprehension are happening together, not one after the other. [c2]

Source 2

That is the useful starting point. The evidence does not say AI makes everyone smarter, or everyone less intelligent. My interpretation is that it shifts where people put effort, trust, attention and emotional investment. The result depends on the task, the user and the design around the system. [c23][c24]

Source 3Source 7Source 8Source 9Source 10Source 13Source 14Source 15Source 16Source 18Source 17

So the question is not simply, “Is AI good or bad for people?” It is: what are people now doing themselves, what are they delegating, and what do they believe while they are doing it?

Exactly. And public expectation is already part of behaviour. Pew found that 53% thought AI would worsen creative thinking, and 50% thought it would worsen the ability to form meaningful relationships. But those are beliefs about effects, not proof that the effects are occurring. [c3][c4]

Source 2

That distinction matters because fear can alter behaviour too. If someone assumes a machine is making them passive, they may check less carefully—or avoid a useful tool altogether. Both overtrust and rejection can change the outcome.

Trust, disclosure and persuasion

Let’s begin with trust, because perception often comes before verification.

People do not always judge advice by its content alone. Experimental research reported that participants sometimes rated AI-generated advice as more effective, higher quality and more authentic than human-generated advice. But that advantage weakened when participants were told it came from ChatGPT. [c5]

Source 3

That is a striking reversal. The same words can be received differently once the source label changes. In health advice, a preregistered survey experiment reported that licensed clinicians retained a credibility advantage, but advice attributed to specialist AI or ChatGPT was still often seen as competent and legitimate. [c6]

Source 4

So disclosure does not produce one universal reaction. Sometimes knowing the source reduces confidence; sometimes the system still appears credible. People are judging both the message and the imagined competence of the author.

There is a matching danger in persuasion. Research reported that large language model messages could persuade people on policy issues, generally by roughly two to four points on a 101-point policy-support scale. That is not mind control, but it is measurable movement. [c7]

Source 5

And a Stanford HAI policy brief reported that human-selected and human-edited GPT-3 propaganda articles were, on average, as persuasive as or more persuasive than the real-world propaganda articles used for comparison. Human editing was part of that result, which is important. [c8]

Source 6

Yes. The story is not merely a machine speaking to a passive public. It is people selecting, refining and distributing material at scale. That makes accountability harder: who chose the argument, who checked it, and who is responsible for its effect?

And trust can fail in the opposite direction. Research on ethical advice found that people may discount advice after learning it was generated by AI, even when expert evaluators rate the advice highly. That is algorithm aversion: rejecting a source because of its category. [c9]

Source 7

So a person can overtrust an apparently polished answer or undertrust a sound one. The practical problem is calibration. The label “AI” is neither a guarantee nor a disqualification.

Which brings us to thinking itself. What happens when the machine supplies the first draft, the explanation or the route through a problem?

One research perspective describes generative AI as extending cognition through external resources. That can expand what people are able to do, but it also increases the importance of metacognitive judgement: knowing when to rely on a suggestion, and how much. [c10]

Source 8

The phrase “how much” is doing a lot of work. Asking for a list of possibilities is different from handing over the decision. Outsourcing a repetitive step may preserve attention for a harder one; outsourcing the whole chain may leave the user unable to inspect the result.

Cognitive offloading and the jagged frontier

A Microsoft Research survey of knowledge workers reported that respondents generally used less critical-thinking effort when they felt confident in AI, and more when they felt responsible for checking or correcting its output. Confidence changed the amount of human effort. [c11]

Source 9

That helps explain why a fluent answer can be risky. Fluency makes checking feel unnecessary, particularly when the user believes the system is reliable. Responsibility, by contrast, seems to pull attention back toward scrutiny.

The work itself also matters. A field experiment with knowledge workers found that AI improved performance on some consulting tasks but worsened performance on others. The researchers described a task-dependent, or “jagged,” capability frontier. [c12]

Source 10

So “AI improves productivity” is too blunt. It may help on one side of a job and impair another, especially where the user needs to recognise an unusual case. The average result can conceal that unevenness.

There is intriguing but preliminary evidence from essay writing too. The MIT research reported differences in EEG patterns, recall and perceived ownership among people using ChatGPT, a search engine or no external assistance. [c13]

Source 11Source 12

But this is exactly where restraint is needed. The supplied research describes that work as a small-sample preprint involving a particular task. It does not establish that ordinary ChatGPT use causes lasting cognitive decline. [c14]

Source 11Source 12

That question remains open. We can say cognitive offloading changes the immediate process, and that effects vary by workflow. We cannot responsibly turn one preliminary essay study into a verdict on the human brain.

Source 8Source 11Source 12

Preliminary evidence versus sweeping conclusions

A useful interpretation is that practice may move. If the system performs a foundational step repeatedly, the user may get fewer opportunities to exercise that step—but whether that matters depends on the task, the person and whether they still practise or verify.

Source 6Source 7Source 8Source 9Source 10Source 3Source 13Source 14Source 15Source 16Source 18

Now consider the social side, where the interface is not just a tool but a conversational partner.

A meta-analysis reported that human-like cues in text-based conversational agents had a moderate effect on perceived humanness, small-to-moderate effects on rapport, trust and positive affect, and smaller effects on attitudes and behaviour. The cues do change social responses. [c15]

Source 13

That does not mean the system is a friend in the human sense. It means people respond socially to signals such as human-like language. The interface can invite emotional investment even when the underlying relationship is asymmetric.

Source 13Source 8Source 9Source 10Source 14Source 17Source 18

The evidence on consequences is more mixed. A four-week randomized controlled study involving nearly 1,000 participants examined voice style, conversation type, loneliness, real-world social interaction, emotional dependence and problematic use. Its supplied research cautions against generalising the English-speaking U.S. sample across cultures or languages. [c16]

Source 14

A longitudinal randomized study also examined anthropomorphism and possible social effects of companion-chatbot use, but the meaning of “social impact” varies and needs cautious interpretation. That makes broad claims about companionship premature. [c17]

Source 15

Human-like chatbots and emotional investment

There is, however, a strong association worth taking seriously. A 2026 cross-national survey of 7,027 people in Germany, China, South Africa and the United States reported a strong correlation between emotional attachment to chatbots and user dependence. Correlation does not establish causation. [c18]

Source 16

Right. Dependence might encourage attachment, attachment might encourage dependence, or another factor might influence both. We have evidence of a relationship, not a settled explanation.

Source 16

And the open question is long-term: does chatbot companionship supplement human contact, substitute for it, or do different patterns occur for different people? The supplied research does not yet answer that at population level.

Source 14Source 15Source 16

What we can say is that design and user state matter. A human-like voice, a lonely moment, an always-available system and a low-friction interface may combine differently from a brief practical interaction. That is an interpretation consistent with the mixed findings, not a universal law.

Source 11Source 12Source 13Source 14Source 15Source 16Source 3Source 7Source 8Source 9Source 10Source 18

The final shift is perceptual: not only trusting what AI says, but recognising what AI has made.

Pew reported that 76% of Americans considered it very or extremely important to know whether text, images or video were made by AI or humans. Yet 53% were not too or not at all confident in their ability to identify AI-generated content. [c19]

Source 2

Recognising synthetic media

And exposure is widespread. A Pew browsing-data study of about 2.5 million webpage visits from 900 U.S. adults during March 2025 reported that 58% encountered an AI-generated summary during at least one search, while 13% visited an AI-tool website directly. [c20]

Source 17

That last comparison is revealing. People may encounter synthetic or machine-produced material without deliberately seeking an AI tool. The boundary between using AI and simply moving through an AI-shaped information environment is becoming less visible.

Source 2Source 17

Human detection is not hopeless, but it is not fixed either. A systematic review reported that humans and automated deepfake detectors often rely on different cues, suggesting collaboration could outperform either humans or detectors alone. [c21]

Source 18

And research on AI-synthesised faces reported that training can improve human detection, although performance depends on the synthetic content and the training strategy. So education may help, but a single checklist will not solve the problem. [c22]

Source 19

That leaves another moving target: synthetic systems improve, while detection skills and expectations try to catch up. The supplied evidence does not establish how well human detection will generalise as those systems change.

Source 18Source 19

Which returns us to accountability. If verification is costly, users may accept a plausible answer. If disclosure is absent, they may misjudge the source. If responsibility is unclear, the person least able to check may carry the greatest burden.

Source 3Source 4Source 5Source 7Source 9Source 2Source 17Source 8Source 10Source 13Source 14Source 18

Verification and accountability

We can now answer one question with some confidence: AI does not have one uniform effect on human behaviour. Results vary by task, population, user behaviour, disclosure, interface and context. That is the strongest general conclusion available in this evidence. [c23]

Source 3Source 7Source 8Source 9Source 10Source 13Source 14Source 15Source 16Source 18

But several important questions remain unanswered. We do not know whether routine AI use causes lasting, general cognitive decline. We do not know the long-term population effects of companion chatbots on loneliness, relationships or dependence. [c14][c16][c17]

Source 11Source 12Source 14Source 15

We also do not know a universal level of trust or disclosure that produces the best reliance across domains. The right arrangement for medical advice may not be the right arrangement for creative work, education or ordinary conversation.

Source 3Source 4Source 7Source 8Source 9Source 10Source 13Source 14Source 15Source 16Source 18

What remains open

So what happened? People gained systems that can answer, persuade, imitate social presence and alter the information they encounter. Human behaviour adjusted around them—not uniformly, but through changing effort, trust and attention.

Source 1Source 5Source 13Source 17Source 8Source 9Source 10Source 14Source 18

And what remains open is whether institutions and habits can keep responsibility attached to the human decision-maker while the machine becomes more convincing. The future question is not simply what AI can produce, but what people will still know, notice and choose to do themselves.

Source 6Source 8Source 9Source 18Source 3Source 7Source 10Source 13Source 14Source 15Source 16Source 17