Last updated November 15, 2026.
About 3 in 4 US adults (76%) say it is extremely or very important to be able to tell whether content was made by AI or by people, yet about half (53%) are not confident they could tell the difference themselves, according to a Pew Research Center survey of 5,023 US adults from June 2025.
AI disclosure labels usually lower reported trust, but they do not reliably reduce how much people rely on the advice, and a label on its own does not help people spot unreliable AI advice. Teams that disclose AI use should pair the label with clear accountability and recommendation-level cues, then test it against decision quality: people should rely more on correct AI recommendations and less on incorrect ones.
At The Decision Lab, we run controlled experiments that change one cue at a time, such as a message's framing or its apparent source. Low trust in AI accuracy was one of five barriers we identified in a 20,000-person Fortune 500 AI adoption program, and that work is why we treat trust as several outcomes a label can move in different directions.
Compliance note. Since 2 August 2026, Article 50 of the EU AI Act has set transparency obligations for some AI systems and AI-generated content offered to people in the EU. Scope and timing vary by type of system and type of content, so confirm with legal counsel how they apply to you. Where disclosure is required, the design question that remains is how to word and place the label so it supports informed reliance.
Five outcomes a disclosure label can change
Across the studies reviewed here, AI labels commonly lower self-reported trust, especially for medical advice and news. The other four outcomes move less predictably, and far fewer studies measure them, even though decision quality and appropriate skepticism are usually what an organization most needs to know about.
| Outcome | What it captures | Example measure |
|---|---|---|
| Self-reported trust | How reliable or competent people rate the advice or its source | Rating scales for reliability and competence |
| Reliance | How far people move their decision toward the advice | Weight of advice, or the share of decisions that switch to the AI's recommendation |
| Comprehension | Whether people understand the advice and what the label means | Questions about the advice and about who they believe reviewed it |
| Decision quality | Whether final decisions are more accurate | Accuracy of final decisions against a known answer |
| Appropriate skepticism | Whether people rely more on correct advice than on incorrect advice | Reliance when the AI is right minus reliance when it is wrong |
What research says about trust, reliance and decision quality
Self-reported trust usually falls
In two preregistered experiments with 2,280 participants on AI medical advice, Moritz Reis, Florian Reis and Wilfried Kunde showed people identical medical advice and changed only its stated source. Advice labeled as coming from AI, or from a physician working with AI, was rated less reliable and less empathetic than the same advice labeled as coming from a physician alone, and participants said they were less willing to follow it.
Sacha Altay and Fabrizio Gilardi ran two preregistered experiments on AI-labeled headlines with 4,976 people in the US and UK. Labeling headlines as AI-generated lowered their perceived accuracy by 2.66 percentage points compared with unlabeled headlines, whether the headlines were true or false and whether a person or AI had written them. Labeling headlines as false lowered perceived accuracy by 9.33 points, roughly three and a half times as much.
In a conjoint experiment on task-specific AI disclosures in news with 683 people, Nicolas Mattis, Kimon Kieslich and Claes de Vreese found that every kind of AI disclosure they tested lowered the perceived trustworthiness of news. The size of the drop varied with readers' political position and their knowledge of how newsrooms use AI.
The penalty also reaches whoever discloses. Across 13 experiments on the AI transparency dilemma, Oliver Schilke and Martin Reimann found that people who disclosed using AI, from professors grading assignments to job applicants writing cover letters, were trusted less than people who said nothing, whatever framing the disclosure used. Favorable attitudes toward technology made the penalty smaller without removing it.
Reliance changes less consistently
Lower ratings do not necessarily mean lower behavioral reliance. Robert Böhm, Moritz Jörling, Leonhard Reiter and Christoph Fuchs ran four preregistered experiments on generative AI advice with 3,308 US participants. People who knew advice on societal and personal problems came from AI rated its author as less competent, but they did not rate the advice itself lower and were no less willing to follow or share it.
Sebastian Krügel, Andreas Ostermaier and Matthias Uhl asked ChatGPT the same moral dilemma several times and got opposite recommendations. In an experiment on ChatGPT's moral advice, that advice shifted participants' judgments about as much when ChatGPT was named as the source as when it was not, and in one dilemma it reversed the majority judgment even with the source disclosed. Participants also underestimated how much the advice had swayed them.
In forecasting and estimation tasks, a label can increase reliance. Across six experiments on algorithm appreciation, Jennifer Logg, Julia Minson and Don Moore found that lay participants moved their estimates further toward advice labeled as coming from an algorithm than toward the same advice labeled as coming from a person. Experienced forecasters relied less on the algorithm, and their accuracy suffered for it.
Decision quality depends on cues a label cannot provide
A system-level label typically does not vary by recommendation quality, so it can raise or lower overall reliance without telling people which recommendation is wrong. Cues that do vary by recommendation carry risks of their own. In studies of AI explanations and human-AI team performance by Gagan Bansal and colleagues, where human and AI accuracy were comparable, explanations increased acceptance of correct and incorrect AI recommendations alike and did no better than showing the AI's confidence.
Zana Buçinca, Maja Barbara Malaya and Krzysztof Gajos tested designs that push people to think before accepting AI advice with 199 participants. These designs cut overreliance compared with simple explanations but did not eliminate it, and participants rated them least favorably.
Why findings differ by domain and label wording
We see four reasons the studies above point in different directions.
What readers think the label means
In the Reis study, the label barely changed how comprehensible people found the advice. What people misread is the label. In Altay and Gilardi's second experiment, skepticism came from people assuming that "AI-generated" meant AI had chosen and written the content with no human supervision. Wording shifts that inference: in a study of AI disclosure wording published in Journalism by Patrick Merle, Mitch Krueger and Erika Schneider, messages labeled AI-generated were judged less credible than messages described as AI-assisted or AI-influenced. With half of US adults doubting they can spot AI content themselves, the label is often the only signal people have about how the advice was made.
The kind of decision
In the Logg study, lay participants leaned more on advice labeled as algorithmic when making numerical forecasts. Medical advice labeled as AI drew a penalty instead, including lower ratings of empathy. The same label can therefore help in one setting and hurt in another.
How much of the AI's output people have seen
In a three-condition experiment on AI ethical advice with 642 participants run by researchers at the Wharton School, 72.6% preferred ethical advice from a human before seeing any. Among participants who then compared AI and human advice with the sources shown, 53.2% still preferred the human, and with the sources hidden that share fell to 46.3%. These figures measure stated preference, but they suggest a label weighs most before people have seen what the AI produces.
Whether the advice or the adviser is being judged
Schilke and Reimann's penalty falls on the legitimacy of the person who disclosed, a different target from the advice. A 2026 systematic review of AI provenance and disclosure cues in journalism by Lorena Liçenji and Julian Hoxha included 47 studies and found no consistent penalty for content attributed to AI. Only 10 of those studies tested disclosure cues, with mostly null or conditional results, and the authors called for designs that separate who appears to have written a piece from whether AI use is disclosed.
How to test an AI disclosure label: 7 steps
When clients ask us whether a label will hurt or help, we design the test to answer the decision-quality question as well as the trust question.
- Hold the advice identical and change only the label. Use at least four arms: no label, an AI label, a human label, and an AI label that states what a human reviewed. The last arm tests the assumption Altay and Gilardi identified.
- Seed the advice with known errors. Include recommendations that are wrong on purpose, at a realistic rate, so you can score reliance separately when the AI is right and when it is wrong.
- Measure reliance as a change in decision, over repeated decisions. Use a judge-advisor setup: participants make an initial decision, see the advice, then make a final one. The weight of advice measure used by Logg and colleagues captures how far the final answer moves.
- Put real consequences on the decision. Pay for accuracy or track an actual choice, such as a click or a purchase, so reported trust can be compared with what people do.
- Ask what participants think the label means. A short check on whether they believe a human reviewed the advice tells you whether the label was understood.
- Ask separately about the advice and about whoever delivered it. The Böhm and Schilke studies show the two can move apart.
- Power the study for small effects and analyze it by audience segment. Label effects in the studies above were often a few percentage points and varied with attitudes toward AI, so preregister the analysis, including the segments.
What regulated teams should do now
- Say what a person checked and who is accountable for the advice. Altay and Gilardi's results suggest part of the drop in trust comes from people assuming nobody did.
- Where the system can support one, show a calibrated confidence estimate on each recommendation. A confidence figure helps only if it tracks how often the AI is right, and Bansal and colleagues found confidence displays did at least as well as explanations.
- For high-stakes decisions, ask people to form their own view before they see the AI's, and expect them to like the process less.
- Judge the label by whether reliance rises when the AI is right and falls when it is wrong, and report results by audience segment alongside the overall trust ratings.
The Decision Lab perspective
Our workforce AI adoption research shows why a single disclosure treatment may not work for everyone. In a Fortune 500 AI adoption program, skeptics stalled at first contact because they doubted AI's value, while enthusiasts were already experimenting and ran into friction. A label sends the same cue to both groups even though their barriers differ, so we test disclosure by audience segment and by behavioral outcome.
Our Money Guided experiment on financial wellbeing messaging shows what a label test can look like in practice. We compared four framings of the product's value against a control in a 400-participant experiment using validated wellbeing scales, and the winning frame now shapes how the product presents itself to employees and to the HR managers who buy it. That result was measured before launch, so in-product behavior is the next test, and we would run a disclosure label through the same sequence of a controlled experiment followed by measurement in the live product.
Frequently asked questions
Do AI disclosure labels make people trust advice less?
Commonly, when trust is measured with ratings. Medical advice labeled as AI-generated was rated less reliable in two experiments with 2,280 people, and AI-labeled headlines were judged less accurate. Behavior moves less consistently: in a study of 3,308 people, an AI label lowered the author's rated competence without lowering willingness to follow the advice.
Is it safer not to label AI-generated advice?
Usually not, because staying silent carries its own risk. In Schilke and Reimann's experiments, being exposed as an AI user by someone else cost more trust than disclosing it, and disclosure is now a legal obligation for some AI systems used in the EU. The more useful decision is how to word the label and what to pair it with.
What is the difference between trust and reliance on AI?
Trust is what people report about advice or its source, usually on a rating scale. Reliance is what they do with the advice: how far their final decision moves toward the AI's recommendation. Because the two can diverge, a label study that measures only ratings cannot tell you whether decisions changed.
How do you measure overreliance on AI in an experiment?
Include AI recommendations that are deliberately wrong and compare how often people follow correct and incorrect advice. Overreliance shows up as following incorrect advice people would have rejected on their own. Collect each participant's initial answer before they see the advice, so you can tell when the AI turned a correct answer into a wrong one.
Does saying a human reviewed the AI advice help?
The evidence is mixed. Medical advice labeled as coming from a physician working with AI was still rated less reliable than physician-only advice. In news, AI-assisted wording was judged more credible than AI-generated, and skepticism toward AI labels came partly from people assuming no human had been involved. Test the exact wording with your own users before rolling it out.

