Survey data on AI and human agency show that Americans are more likely to have low than high confidence in people’s ability to retain key human abilities as AI advances. In Elon University’s 2026 survey, 55% were slightly or not at all confident that most people would maintain critical thinking and information-evaluation skills, compared with 15% who were extremely or very confident. The same gap appeared in views about using AI appropriately and effectively, at 50% versus 15%, and maintaining motivation and purpose, at 46% versus 15%, according to Elon University’s Human Resilience Survey.
The survey asked what most people would be able to do in the coming years, not what you can do individually. It was fielded May 14–18, 2026, among 1,505 U.S. adults. A related but distinct measure comes from Pew Research Center: 59% of U.S. adults said they had little or no control over whether AI is used in their lives, while 14% reported a great deal or quite a bit of control. Feeling unable to control AI’s presence in your life is not the same as doubting people’s ability to judge information independently.
55% vs 15%
Low or no confidence vs high confidence that people will retain critical thinking and information-evaluation skills
50% vs 15%
Low or no confidence vs high confidence that people will use AI appropriately and effectively
46% vs 15%
Low or no confidence vs high confidence that people will maintain motivation and purpose
59% vs 14%
U.S. adults reporting little or no control vs a great deal or quite a bit of control over AI use in their lives
Quick Facts
- Critical thinking confidence: In the May 14–18, 2026 survey, 55% of U.S. adults were slightly or not at all confident that most people would retain critical-thinking and information-evaluation skills as AI advances, compared with 15% who were extremely or very confident.
- Using AI confidence: 50% were slightly or not at all confident that most people would use AI appropriately and effectively in the coming years, compared with 15% who were extremely or very confident.
- Motivation and purpose confidence: 46% were slightly or not at all confident that most people would maintain motivation and purpose as AI advances, compared with 15% who were extremely or very confident.
- Survey scope: The Human Resilience Survey was fielded May 14–18, 2026, among 1,505 U.S. adults; it asked about most people’s future abilities, not respondents’ individual skills.
- Perceived control over AI: In a separate 2025 measure, 59% of U.S. adults said they had little or no control over whether AI is used in their lives, while 14% reported a great deal or quite a bit of control.
Key Takeaways
- In the May 14–18, 2026 survey of 1,505 U.S. adults, more respondents had low or no confidence than high confidence that most people would retain critical thinking and information-evaluation skills (55% versus 15%), use AI appropriately and effectively (50% versus 15%), and maintain motivation and purpose (46% versus 15%).
- The survey measured confidence in what most people could do in the coming years. Its results describe respondents’ perceptions of future human capacities.
- Among LLM or chatbot users, 34% said they always or most of the time verify outputs using non-LLM sources, while 18% said they frequently follow recommendations; 51% had at least sometimes felt they were taking an easy shortcut and 39% had at least sometimes felt too dependent. These figures describe users’ reported checking, reliance, and experiences.
- To examine your own judgment, write down your view before using AI, note what changed and what evidence supports your final choice, and recheck relevant information or seek another perspective when you cannot explain your decision in your own words.
How Confident Are Americans About Human Agency?
Across all three measures of human agency, Americans were more likely to express low than high confidence about the coming years.
| Ability being assessed | Slightly or not at all confident | Extremely or very confident |
|---|---|---|
| Thinking critically and evaluating information | 55% | 15% |
| Using AI appropriately and effectively | 50% | 15% |
| Maintaining motivation and purpose as AI advances | 46% | 15% |
Respondents assessed what they thought most people would be able to do in the coming years. These figures reflect confidence in people’s future capacity, not a test of anyone’s current skills or a prediction about each respondent personally. The three abilities represent different aspects of human agency. Critical thinking and evaluating information involve assessing claims and reaching judgments. Using AI appropriately and effectively involves exercising judgment when working with AI. Maintaining motivation and purpose involves continuing to pursue meaningful goals as AI advances. For each ability, the table compares the share expressing slight or no confidence with the share expressing extreme or very high confidence. It does not show other confidence responses, so the two columns do not add up to 100%. The consistent gap suggests that respondents were more doubtful than assured about people retaining these capacities. These are survey responses about confidence, not evidence that people have already lost these abilities or that AI caused any change in them.
What Do The Survey Numbers Measure?
The numbers summarize respondents’ reported attitudes and experiences about human agency as AI advances. The Human Resilience Survey was fielded May 14–18, 2026, among 1,505 U.S. adults, with an oversample of 501 LLM or chatbot users. Its reported margin of error is ±3.1 percentage points at the 95% confidence level. The three confidence measures reflect U.S. adults’ views of what “most people” will be able to do in the coming years. The oversample provides a separate view of people who use LLMs or chatbots. Later figures about checking AI outputs, following recommendations, or feeling overdependent refer specifically to those users, not to all U.S. adults. Keep this distinction in mind when comparing confidence responses with reported behavior. The numbers can help you understand public perceptions alongside users’ accounts of their AI habits.
How Do AI Users Check Their Judgment?
LLM or chatbot users report both checking AI outputs and sometimes relying on them or feeling overdependent. The figures below refer to LLM or chatbot users and come from the Elon University survey topline and methodology, September 2026.
| Reported behavior or experience among LLM or chatbot users | Share |
|---|---|
| Always or most of the time verify AI outputs using non-LLM sources | 34% |
| Always or most of the time follow an LLM’s recommendations | 18% |
| At least sometimes felt lazy or as if taking an easy shortcut | 51% |
| At least sometimes felt too dependent on LLMs for answers rather than thinking things through themselves | 39% |
Among LLM or chatbot users, the 34% who said they frequently verify outputs and the 18% who said they frequently follow recommendations reported different behaviors. The figures do not show whether the same users do both or whether users who verify are more or less confident in their judgment. The 51% who had at least sometimes felt lazy or as if they were taking an easy shortcut, and the 39% who had at least sometimes felt too dependent, reported experiences rather than demonstrated losses of ability. You might verify one output while accepting another recommendation, or question your habits even when you check information. The survey figures do not establish that AI use caused these feelings or changed anyone’s capacity to think independently. Together, the results show a tension in users’ accounts, not a score of who has stronger agency. Frequent checking does not prove independent judgment, and reported dependence does not prove that judgment has weakened.
How Does Control Over AI Compare?
Control over whether AI enters your life and confidence in people’s independent judgment are related concerns, but they measure different things. The control measure asked whether you feel able to influence AI’s use in your life. The confidence measure asked whether people can retain critical thinking, use AI effectively, and maintain purpose.
The report, published in April 2025, draws on a U.S. adult survey conducted August 12–18, 2024. The findings compare perceived control over AI’s presence with confidence in people’s capacities. They are not two versions of the same question.
| Measure | What respondents were asked about | Finding |
|---|---|---|
| Pew: control over AI use in one’s life | Whether respondents felt they had control over AI being used in their lives | 59% said little or no control; 14% said a great deal or quite a bit of control |
| Confidence in human agency | Whether most people could think critically, use AI appropriately and effectively, and maintain motivation and purpose | Confidence in people’s abilities, not control over AI’s use in one’s life |
The Pew result points to a perceived limit on choosing whether AI becomes part of everyday life. The Elon questions ask whether people can still exercise judgment and pursue meaningful goals as AI advances. You could feel unable to control AI’s presence while remaining confident in people’s ability to assess information. Conversely, you could feel able to limit AI use while doubting people’s capacity to judge independently. The two surveys examine different dimensions of agency, so you should not treat their responses as directly comparable scores.
Can AI Autonomy And Human Agency Coexist?
Yes. AI-system autonomy and human agency can coexist because they describe different capacities. Autonomy concerns how a system acts, while human agency concerns your ability to think, choose, and take responsibility. The OECD’s 2026 report on agentic AI distinguishes autonomy from agency, though people sometimes use the terms interchangeably. Autonomy generally means a system’s capacity to act without direct human involvement. It describes how much action is delegated to a system, not whether the system makes a human judgment or bears responsibility for its choices. The OECD definition of an AI system describes autonomy as the degree to which a system can learn or act without human involvement after autonomy and process automation have been delegated. Human supervision, however, can happen at any stage of an AI system’s lifecycle. A system can therefore operate with some independence while people remain involved in setting it up, monitoring its operation, or responding to its outputs. When you rely on an AI answer, ask not only whether the system can produce a convincing response, but also what remains your responsibility to judge. Did you decide what to ask? Did you assess whether the answer was relevant and well supported? Did you choose whether to act on it? A system’s ability to generate or carry out a recommendation does not, by itself, settle whether you understood or endorsed the decision. These questions connect to Philosopedia’s resources on the philosophy of mind, judgment, and moral responsibility. They can help you consider what it means to understand a claim, make a choice, and answer for an action when an AI system has contributed to the process. The OECD AI Principles call for safeguards suited to context, including support for human agency and oversight. The point is not that every AI-assisted decision requires the same level of human involvement. Rather, a system’s autonomy should not obscure the human choices and responsibilities that remain in a particular setting.
Human Oversight In Law And Policy
Article 14 of the European Union Artificial Intelligence Act says human oversight of high-risk AI systems must aim to prevent or minimise risks to health, safety, or fundamental rights arising from intended use or reasonably foreseeable misuse. This is an institutional safeguard: oversight is intended to address risks associated with how a system is used, not to measure whether you or other members of the public feel confident in human judgment. Article 14 of the EU AI Act
Transparency serves a separate purpose. The European Commission says Article 50 requirements apply from August 2, 2026, including requirements for certain systems to inform people when they are interacting with AI. This information makes AI’s involvement more visible, while human oversight concerns supervision of system use. The Commission’s Article 50 guidance
The timing for high-risk rules differs from the Article 50 start date. In its October 2026 enforcement timeline, the Commission schedules rules for high-risk AI use cases in Annex III to apply from December 2, 2027, and rules for high-risk AI systems embedded in regulated products to apply from August 2, 2028. These are future application dates in the timeline, not requirements already applying by October 2026. The Commission’s enforcement timeline
Together, oversight and transparency are legal and institutional safeguards, while the survey measures reported public confidence. EU rules set obligations for specified AI systems and contexts.
Assessing Your Own AI Judgment
Choose a recurring task where you use a chatbot, such as drafting a message or comparing options. Before asking for AI help, write down your initial view. Afterward, note what the AI changed, what evidence supports that change, and what you still doubt. This gives you a concrete example of how your judgment works with AI, rather than a general impression of whether you are “good” at using it.
If you cannot explain your final choice without repeating the AI’s wording, pause before acting. Recheck the relevant information using a non-LLM source, or ask another person for their perspective when the decision matters. Treat the difficulty as a reason to examine your reasoning, not as proof that AI has weakened your ability to think.
Compare your notes across different kinds of tasks. Does AI help you consider options, or do you accept its recommendation without being able to explain why? Does your answer change when the decision affects someone else? For a deeper framework, read Philosopedia’s resources on philosophy of mind and moral responsibility. They can help you examine what understanding involves and when an AI-assisted decision remains your own.
Frequently Asked Questions
1. How do confidence levels differ by age, education, or political affiliation?
The reported confidence figures are overall results and do not include breakdowns by age, education, or political affiliation. You cannot tell from these percentages which groups are more or less confident.
2. What kinds of non-LLM sources do people use to verify AI answers?
The survey records whether LLM or chatbot users check AI outputs against non-LLM sources, but it does not identify which sources they consult. To fact-check an answer, you could compare it with original research, official records, or materials from a relevant institution.
3. Will future surveys show whether confidence in human agency changes over time?
Future surveys could reveal changes if researchers repeat the questions with comparable samples. The 2026 results provide one measurement. Whether a trend emerges will depend on future data.
