Can People Tell AI-Generated Content Apart? Measure Discernment, Not Self-Report

Ask a survey respondent "can you tell when content is AI-generated?" and the answer you get is mostly about confidence and self-image, not ability. If the research goal is to compare groups, to compare stimulus types, or to test whether a warning or a training session made people better at spotting AI output, the question has to change: have people make binary judgments on balanced material, then use signal detection theory to separate how well they discriminate from how willing they are to call something AI. This guide explains why self-report cannot stand in for an ability measure, then gives a recomputable worked example and a set of design rules you can apply directly.

"I think I can spot it" measures confidence, not accuracy

Confidence and accuracy can come apart, and the direction of the gap is sometimes the wrong way. In two nationally representative U.S. samples, Lyons and colleagues found that about three in four respondents overestimated their relative ability to distinguish legitimate news from false news, placing themselves 22 percentiles higher than warranted on average; this overconfidence was in turn associated with visiting untrustworthy websites more often, failing to tell true from false claims in a test, and reporting greater willingness to like or share false content (Lyons et al., PNAS 2021). Moving "I can detect it" straight into a questionnaire therefore measures self-efficacy. Self-efficacy and task performance are two variables that must be measured separately and then related to each other, not one standing in for the other.

The same pattern holds for AI-generated content. Across six experiments with 4,600 participants, Jakesch, Hancock and Naaman asked people to discern AI-written self-presentations produced by a state-of-the-art language model in professional, hospitality, and dating contexts. Participants could not reliably identify the AI text; a computational analysis of language features showed their judgments were steered by intuitive but flawed cues such as first-person pronouns, contractions, and family topics, and those same cues could be exploited by a model to produce text rated as "more human than human" (Jakesch et al., PNAS 2023). The evidence on faces runs the same way: once synthesis engines cleared the uncanny valley, AI-synthesized faces were judged human and even rated more trustworthy than real faces (Nightingale & Farid, PNAS 2022).

So "can people detect it?" is a performance question that needs a task, not an attitude question that can be asked directly with an agreement scale. Treating self-report as the answer leaves an unfalsifiable number in the report: a higher value only shows that respondents are more confident, not that they are more accurate.

Split a single judgment into two computable quantities: sensitivity and bias

Signal detection theory provides the tool. Classify both the material and the judgment into two categories, and make a binary decision: the material is truthfully "AI-generated" or "human-written," and the judgment is "AI" or "human." Treat "saying AI" as the signal. The four cells are: AI material judged AI (a hit), human material judged AI (a false alarm), AI material judged human (a miss), and human material judged human (a correct rejection). Once hit and false-alarm counts are recorded, two non-confounded indices follow:

• Sensitivity d′ = z(hit rate) − z(false-alarm rate), where z is the standard-normal quantile. A larger d′ means the two classes are genuinely better separated.
• Response bias (criterion) c = −½ × [ z(hit rate) + z(false-alarm rate) ]. A c near 0 means no general lean toward saying AI or human; negative c means a default lean toward "AI," positive c a default lean toward "human."

Why insist on separating them? Because many "working" metrics pull in both at once. In a methodological review, Batailler and colleagues show that the ability to distinguish true from false material and the bias to call material true or false regardless of its veracity are two distinct aspects, and that a single accuracy or mean-rating score is moved by both; reanalyzing existing datasets, they show how this confound can change conclusions (Batailler et al., 2021). Modirrousta-Galian and Higham reanalyzed five gamified "fake news inoculation" studies with ROC curves and found the training did not improve discrimination; it made people call all material false more often, that is, it mainly shifted the criterion (Modirrousta-Galian & Higham, 2023). Reporting only the share of correct judgments before and after would have misread a shift toward caution as a gain in skill. This is the same logic as keeping false inclusion and false exclusion separate in screener design.

A recomputable example: the same 75% accuracy, different ability and bias

Suppose a detection task uses 40 AI-generated texts and 40 human-written texts, so the truth is split evenly. Two groups each judge all 80 items; the results are the illustrative numbers below (chosen to demonstrate the arithmetic, not as benchmarks):

Raw accuracy is identical, 75% for both. Unpacked with the formulas, the two groups are not the same. Group A: hit rate 0.75, false-alarm rate 0.25; z(0.75) = 0.674 and z(0.25) = −0.674, so d′ = 0.674 − (−0.674) = 1.35 and c = −½ × (0.674 − 0.674) = 0, with no general lean toward either response. Group B: hit rate 0.90, false-alarm rate 0.40; z(0.90) = 1.282 and z(0.40) = −0.253, so d′ = 1.282 − (−0.253) = 1.53 and c = −½ × (1.282 − 0.253) = −0.51, a clear lean toward calling everything AI.

Group B has higher sensitivity, but its score is also inflated by a near-blanket "AI" tendency; Group A discriminates slightly less well but judges in a more balanced way. Writing "both groups scored the same, so their detection ability is equal" collapses ability and bias into one number that can no longer be taken apart, and the two correspond to different remedies: raising sensitivity depends on materials, cues, and feedback, while correcting bias often just means adjusting the baseline prompt or the stimulus mix. The d′ and criterion formulas and their use on experimental data are laid out by Stanislaw and Todorov (Stanislaw & Todorov, 1999).

The stimulus mix manufactures false highs: report the AI-to-human ratio

Raw accuracy has an even more hidden problem: it changes with the mix of material. Suppose a set contains 10 AI-generated and 90 human-written items. A respondent who answers "human" every single time gets 90 items "right," a raw accuracy of 90%, while their hit rate on AI items is 0 and their real sensitivity is zero. Put the same person in a set of 50 AI and 50 human items and the blanket "human" strategy drops to 50%. The same person using the same strategy can look weak or strong depending only on the ratio.

This is why any report must give three things together: the proportion of AI to human stimuli, the hit rate and false-alarm rate rather than a single overall accuracy, and the d′ and criterion derived from them. "Detection accuracy was 78%" with no ratio and no hit/false-alarm split leaves the reader unable to tell whether the number reflects real discrimination or a stable bias toward one category. Keeping the mix near 50% and stating it in the method section is the precondition for comparing results across groups and batches.

Four rules for designing the task and computing the indices

Turning the principles into practice, these four are the easiest to skip and the most damaging to conclusions:

• Use a binary judgment, not a subjective "how AI does this seem" rating. Only an AI/human two-way outcome yields hit and false-alarm rates directly; if a 1–5 plausibility rating is unavoidable, analyze it as a whole with an ROC/AUC approach rather than comparing pre-post means (that is exactly how Modirrousta-Galian and Higham found the training had not improved discrimination).
• Show each item once and randomize the presentation order. Repeating an item invites memory and learning effects that turn later high accuracy into a false improvement; draw items from several models and several contexts so the conclusion is not tied to a single model.
• Measure self-report and discrimination separately, and state the order. Confidence can be reported as an independent variable alongside d′ and criterion, to test whether confidence tracks ability, but it cannot be substituted for ability; the order of the two blocks belongs in the method because they can influence each other.
• Correct 0 and 1, and give an interval. When a hit or false-alarm rate is exactly 0 or 1, z is undefined, so apply a 1/(2N)-type correction before computing (Stanislaw & Todorov, 1999; the origin of the rule is in Snodgrass & Corwin, 1988). Use at least 20–40 trials per cell and report an uncertainty range for d′ and criterion, as with any confidence interval, rather than a single point estimate.

On "one judgment per item, mixed presentation," a large online study gives an order-of-magnitude reference. Across two studies with 15,016 participants, Groh and colleagues compared ordinary humans, a leading machine model, and machine-informed humans on deepfake videos: overall human accuracy was close to the model's, but they made different kinds of errors; people who saw an accurate model prediction were more accurate than either alone, while inaccurate predictions often reduced human accuracy (Groh et al., PNAS 2022). Stimulus type, machine assistance, and human cognitive processing all change the result, so any conclusion must be bounded by the material types and presentation conditions actually tested.

What each of three question formats can and cannot support

Even when all three ask about "detecting AI-generated content," the strength of evidence differs sharply, and the choice depends on the decision the research must support:

For most product, content-safety, or platform questions, the third row is what is needed: comparing the ability of different groups and determining whether a warning, a label, or a training changed ability or merely moved the threshold. The second row suits item-difficulty screening and cue diagnosis, and the first suits measuring how confident people are. Mixing them up is the most common source of unwarranted conclusions in this kind of study.

When this approach does not apply

• When the goal is a conclusion about an individual. d′ and criterion are summary indices for a given condition; judging one person requires enough trials, and the answer holds only for the material type tested, not for platforms or model versions that person has not seen.
• When the material itself goes stale. Generation models and synthesis engines keep changing, so a d′ measured at one time reflects that time only; the source, generation tool, and collection date must be recorded or the result cannot be reproduced.
• When detection is equated with resistance to deception. Whether someone can tell AI content apart and whether they will be deceived or share it in a real setting are different questions; real settings add motivation, time pressure, and social context that a lab discrimination index cannot replace.
• When the sample is not representative. A d′ from a convenience online sample does not generalize to a population; age, education, and digital-experience differences belong in the quota and stratification, following sample quota design and low-quality response detection.

In addition, material involving identifiable faces or voices requires informed consent and privacy safeguards, and this article is about measurement method, not a content-moderation or compliance conclusion.

Where WeJot fits in this chain

Making the design above run usually comes down to three execution tasks: presenting the material, controlling the sample, and exporting analyzable fields. In WeJot you can build the AI and human material as two question sets and present them in a randomized order, use conditional logic to control the ratio each respondent sees, and prevent the same item from repeating for one person; use quotas to hold age, education, and similar strata constant so that d′ comparisons share a common sample base; and export "material type," "judgment," and "self-rated confidence" as separate fields that feed the 2×2 computation directly.

To be clear about the boundary: the signal-detection computation, the judgment of material quality, and the applicability limits of the conclusion remain with the research team. What WeJot can do is make the presentation controllable, the sample structure checkable, and the export recomputable, not decide on the team's behalf whether the material represents current AI output. When the research goal is itself to test synthetic respondents, compare with whether AI-simulated respondents can replace real ones: one asks whether respondents can detect AI content, the other asks whether AI can replace respondents.

Sources and method basis

• Lyons, Montgomery, Guess & Nyhan (2021), PNAS 118(23) — Overconfidence in news judgments is associated with false news susceptibility (two nationally representative U.S. samples; about three in four overestimated relative discernment, by 22 percentiles on average; overconfidence correlated with visiting untrustworthy sites, failing a true/false test, and greater willingness to share false content. Basis for treating self-report as distinct from ability.)
• Jakesch, Hancock & Naaman (2023), PNAS 120(11):e2208839120 — Human heuristics for AI-generated language are flawed (six experiments, N=4,600; participants could not reliably identify strong-model AI self-presentations; flawed cues such as first-person pronouns, contractions, and family topics, exploitable to produce "more human than human" text.)
• Batailler, Brannon, Teas & Gawronski (2021), Perspectives on Psychological Science — A Signal Detection Approach to Understanding the Identification of Fake News (uses SDT to distinguish the ability to discriminate true from false material from response bias, and shows via reanalysis how a single accuracy score is moved by both.)
• Modirrousta-Galian & Higham (2023), Journal of Experimental Psychology: General 152(9):2411–2437 — Gamified inoculation interventions do not improve discrimination between true and fake news (ROC reanalysis of five studies: training did not raise discrimination but shifted the criterion toward calling material false. Basis for separating ability from bias in reporting.)
• Stanislaw & Todorov (1999), Behavior Research Methods, Instruments, & Computers 31(1):137–149 — Calculation of signal detection theory measures (computation of d′ and criterion c and correction when hit or false-alarm rates are 0/1.)
• Snodgrass & Corwin (1988), Journal of Experimental Psychology: General 117(1):34–50 — Pragmatics of measuring recognition memory (origin of the 1/(2N) correction for extreme hit and false-alarm rates.)
• Groh, Epstein, Firestone & Picard (2022), PNAS 119(1):e2110013119 — Deepfake detection by human crowds, machines, and machine-informed crowds (two studies, 15,016 participants; humans and a leading model were similarly accurate but made different errors; machine information helped when the model was right and hurt when it was wrong.)
• Nightingale & Farid (2022), PNAS 119(32):e2120481119 — AI-synthesized faces are indistinguishable from real faces and more trustworthy (synthesized faces judged human and rated more trustworthy, showing detection difficulty changes with generation technology.)