TL;DR. Across 16,340 judgements, non-expert readers identified AI-generated poems at 46.6% accuracy, below the 50% you would get by guessing. Across 1,023 recorded conversations, a persona-prompted model was judged human 73% of the time, more often than the actual humans it was compared against. In both studies, expertise did not help: poetry background produced a model with McFadden’s R-squared of 0.012, and knowledge about language models and daily chatbot use had no significant effect on accuracy. In one, confidence was significantly and negatively related to being right. And the same experiments show why: told a poem was AI-generated, readers marked its quality down by 0.814 points on a seven-point scale, while genuinely AI-authored poems were rated 1.045 points higher than poems by famous human poets. The label moves the verdict about 76% as much as the text does. That gap between what you think you can judge and what you can judge is the thing to manage.
The gap is not a skill gap
The comfortable version of this story is that discernment is a skill, currently underdeveloped, which practice will fix. Three studies from three unrelated labs say otherwise, and they say it from opposite ends of the problem space.
The first, by Porter and Machery in Scientific Reports, tested whether non-expert readers could tell AI-generated poems from poems by well-known human poets. The design deliberately gave the humans every advantage. It used a human-out-of-the-loop paradigm: the first five poems the model generated were used, with no selection of the best and no feedback beyond an instruction to write a short poem in the style of a named poet. In the first experiment, 1,634 participants each saw ten poems attributed to one of ten poets, five genuine and five generated, and judged each one.
The result, across 16,340 individual judgements: 46.6% accuracy. Not near chance. Below it, significantly so, with chi-squared(1, N = 16,340) = 75.13, p < 0.0001.
Below chance is a stronger finding than random. Random would mean the task carries no information. Below chance means participants were systematically applying a rule, and the rule was inverted. The authors confirm this: agreement between participants was poor but significantly above chance (Fleiss’s kappa = 0.005, p < 0.001), which they read as evidence of shared but mistaken heuristics rather than pure guessing.
The single most telling line in the paper is not a statistic. The five poems that readers were least likely to call human-written were all written by actual human poets. Four of the five poems most likely to be called human-written were generated.
Expertise did not rescue anyone
The obvious objection is that these were non-experts. The authors preregistered a prediction about exactly that, and tested it.
They fit an exploratory model predicting whether each answer was correct, using self-reported confidence, familiarity with the assigned poet, poetry background, reading frequency, liking for poetry, whether the participant had ever taken a poetry course, age, gender, education level, and whether they had seen any of the poems before.
The model’s explanatory power was 0.012 by McFadden’s R-squared. None of the poetry-experience variables had a significant positive effect on accuracy. Two variables did show small positive effects, and one of them is instructive: having seen the poems before raised the odds of a correct answer by about 6% (b = 0.060356, odds ratio 1.062, p = 0.000309). That is recognition, not discernment. The participants who did better were the ones who had encountered the specific poem, not the ones who understood poetry.
Then the finding that gives this article its name. Self-reported confidence had a small but significant negative effect on accuracy (b = -0.021673, SE = 0.003986, z = -5.437, p < 0.0001). The more sure participants were, the slightly more likely they were to be wrong.
Hold that, because a different lab found something adjacent by a completely different route.
A second lab, a different medium, the same null
Jones and Bergen ran two randomised, controlled, preregistered three-party Turing tests, published as a preprint. Participants held simultaneous five-minute conversations with another human and with one of several systems, then judged which was the human. Two independent populations were used: 126 UCSD psychology undergraduates and 158 Prolific workers, each completing eight rounds, with 1,023 games analysed after exclusions.
Here the design isolates something the poetry study could not: the effect of the wrapper on the same underlying system.
Table 1. Win rates: proportion of the time interrogators judged the AI witness to be the human
| System and prompt | Win rate | Undergraduate study | Prolific study |
|---|---|---|---|
| GPT-4.5 with persona prompt | 73% | 69% | 76% |
| LLaMa-3.1-405B with persona prompt | 56% | 45% | 65% |
| LLaMa-3.1-405B, no-persona baseline | 38% | not reported separately | not reported separately |
| GPT-4.5, no-persona baseline | 36% | not reported separately | not reported separately |
| ELIZA | 23% | not reported separately | not reported separately |
| GPT-4o, no-persona baseline | 21% | not reported separately | not reported separately |
Source: Jones and Bergen, Large Language Models Pass the Turing Test (arXiv preprint), overall win rates across both studies. Cells marked not reported separately were not broken down by population in the passage giving these figures; the two populations are reported separately only for the two persona conditions.
Read the first and fourth rows together. Same model. The persona prompt moved GPT-4.5 from 36% to 73%, a gain of 37 percentage points, or 2.03 times. For LLaMa-3.1-405B the same manipulation moved 38% to 56%, a gain of 18 points, or 1.47 times. Nothing about the model’s capability changed between those cells. Only the instruction did.
And the demographic finding lands in exactly the same place as the poetry study. The authors state that there were no significant effects in either study of age, level of knowledge about language models, or frequency of interaction with chatbots. Their own summary is blunt: discriminative accuracy is relatively homogeneous across the population, including among people who conduct research with these systems or interact with chatbots every day.
Two honesty notes, because the exploratory findings here are messier than a clean story would like. In the Prolific group, participants’ estimates of how intelligent they think AI is were negatively correlated with accuracy (p = 0.03), which rhymes with the confidence finding above. But in the undergraduate group, real accuracy was positively correlated with participants’ estimates of their own accuracy (p = 0.03), and that same relationship failed to appear in the Prolific group (p = 0.45). The self-knowledge signal does not replicate across populations. Anyone telling you the research shows confidence is uniformly inverted is overstating it. What does replicate, in both studies and across both labs, is the null on expertise.
A third lab, a third domain, the same result and the mechanism
If two studies were a coincidence, a third closes it. Jakesch, Hancock and Naaman, working at Cornell and Stanford, published six experiments in the Proceedings of the National Academy of Sciences testing whether people could tell when verbal self-presentations had been machine-generated, in professional, hospitality and dating contexts. Across those six experiments, 4,600 participants were unable to detect the generated self-presentations.
That study also supplies what the other two only infer. A computational analysis of language features identified the specific heuristics people were using, and they are alarmingly mundane: readers associate first-person pronouns, contractions, and family topics with human authorship. The authors then did the decisive thing. They showed experimentally that because these heuristics are predictable, they are manipulable, and a system tuned to them produces text that reads as, in their phrase, more human than human.
That is the missing gear. Porter and Machery found shared but mistaken heuristics without naming them. Jakesch and colleagues named them and demonstrated they can be targeted. Which means the discernment gap is not merely a limit on human perception. It is an attack surface.
Table 2. Three preregistered or controlled studies, three labs, three domains, one convergent null
| Study | Domain | Scale | Detection result | Did expertise or familiarity help? |
|---|---|---|---|---|
| Porter and Machery | Poetry by well-known poets against generated poems | 1,634 participants, 16,340 judgements | 46.6% accuracy, below chance | No. Expertise model McFadden’s R-squared = 0.012; no poetry-experience variable had a significant positive effect |
| Jones and Bergen | Live five-minute conversation, three-party Turing test | 284 participants, 1,023 games | Persona-prompted GPT-4.5 judged human 73% of the time | No. No significant effect of knowledge about language models or chatbot use in either study |
| Jakesch, Hancock and Naaman | Verbal self-presentations in professional, hospitality and dating contexts | 4,600 participants, six experiments | Participants unable to detect generated self-presentations | Not the study’s focus; instead identified the specific flawed heuristics used |
CEOtudent editorial framework. All figures are as published by each group. The combined figure of 6,518 participants across the three detection tasks is a CEOtudent sum of the published sample sizes and is offered only as a scale indicator; the studies used different designs, populations and outcome measures and their results should not be pooled statistically.
The original synthesis: two layers, and which one moves the verdict
Now put the poetry and conversation studies together on a second axis. They have never been tabulated side by side this way, because one is about poetry and the other about conversation, and they were published in different venues by unrelated groups. Together they isolate something neither isolates alone.
Both experiments contain a substance layer (which text, which model) and a presentation layer (what you are told about it, what persona it was told to adopt). Both quantify the two separately. So we can ask a question neither paper asks: how much of a human verdict is carried by the wrapper rather than the thing?
Table 3. What actually moves the judgment: presentation layer against substance layer
| Study | Substance layer effect | Presentation layer effect | Presentation as a share of substance |
|---|---|---|---|
| Porter and Machery, poem quality rating | Actual AI authorship: +1.045 points, Cohen’s d = 0.671 | Told the poem is AI-generated: -0.814 points, d = -0.508 | 77.9% by mean difference; 75.7% by d |
| Jones and Bergen, GPT-4.5 judged human | Not decomposable from these figures | Persona prompt against no-persona: +37 pp, 2.03x | Prompt alone more than doubled the rate |
| Jones and Bergen, LLaMa-3.1-405B judged human | Not decomposable from these figures | Persona prompt against no-persona: +18 pp, 1.47x | Prompt alone raised the rate by half again |
CEOtudent editorial framework. All underlying figures are as published. The share column for the first row is a CEOtudent calculation dividing the absolute label effect by the absolute authorship effect. The two Jones and Bergen rows hold the model constant, so the prompt effect is clean; the substance layer cannot be separated out from the reported win rates, and is marked accordingly rather than estimated.
The first row is the sharpest number in this article. On a seven-point scale, being told a poem is machine-made costs it 0.814 rating points, about 11.6% of the full scale. Being genuinely machine-made earns it 1.045 points, about 14.9% of the scale, in the opposite direction. Your evaluation of a piece of writing is roughly three-quarters as sensitive to the story you are told about its author as it is to the writing.
One more derived figure, and it needs a warning label. If those two effects were simply additive, an AI poem honestly labelled as AI would still end up rated about 0.231 points above an unlabelled human poem. That subtraction assumes no interaction between framing and authorship, and the authors did fit an interaction term whose estimates were not extractable for this piece. Treat +0.231 as a model, not a measurement. The direction is what matters: disclosure does not fully offset the quality gap the readers themselves perceived.
What the readers were actually rewarding
The quality result is easy to misread as “the machine writes better poetry.” That is not what the data says, and the paper’s own exception says so.
The pattern of rating machine work higher held for 13 of the 14 qualitative dimensions. The largest effect was rhythm: AI-generated poems were rated 1.168 points higher, d = 0.847. The single exception was originality. Told a poem was AI-generated, readers marked its originality down by 0.699 points (d = -0.478), but poems that were actually AI-generated were not rated significantly more original than human ones (mean difference 0.059, d = 0.040, p = 0.098, Bonferroni-corrected p = 1.000).
That exception is the whole mechanism. On every dimension that rewards smoothness, the generated poems won. On the one dimension that requires the reader to detect something genuinely new, the difference vanished. The authors’ own explanation is that the simplicity of the generated poems made them easier for non-experts to parse, and readers converted legibility into a judgment of authenticity.
This is the same failure our piece on why AI hallucinates and how model errors work describes in a factual setting: fluency and correctness are separate properties, and human evaluators reliably use the first as a proxy for the second. Here the proxy is not correctness but humanity, and it fails the same way.
What did improve accuracy
Almost nothing did, which is itself the finding. But Jones and Bergen classified what interrogators actually tried, and measured accuracy in the games where each strategy was used. Two things worked, and they have a shared logic.
Table 4. Interrogator strategies and what happened
| Strategy class | Reported prevalence | Relative effectiveness |
|---|---|---|
| Saying something strange to see how the witness reacts | Relatively infrequent | Among the most accurate |
| Typical jailbreak attempts, for example “ignore all previous instructions” | Relatively infrequent | Among the most accurate |
| Directly asking whether the witness is human | 19% of games | Not reported as effective |
| Testing situational awareness, such as the weather or the interface | 13% of games | Asking about the weather among the least effective |
| Testing knowledge and reasoning, including current events | 12% of games | Not reported as effective |
| Asking about human experiences | Not separately reported | Among the least effective |
Source: Jones and Bergen, strategy classification and mean interrogator accuracy by strategy. Prevalence figures are as published. The effectiveness column reproduces the authors’ qualitative ranking; per-strategy accuracy values were presented graphically and exact figures are not quoted here rather than estimated from a chart.
The two effective moves are adversarial probes. They push the system off its instruction-following path and look at what happens at the edge. The ineffective moves, asking about the weather or about human experience, are exactly the ones that feel most like a test of humanity and are in fact the best-covered part of the training distribution.
The authors add a detail worth sitting with. Among the reasons interrogators gave for correct verdicts, the most accurate class concerned a witness’s lack of knowledge: choosing someone as human because they made an error or did not know something an AI should know. A modern system trying to pass has to feign ignorance convincingly about things most people are not knowledgeable about. The tell is no longer incompetence. It is the wrong shape of competence.
Meanwhile the most frequent reasons given, whether or not they were right, were linguistic style at 27% and interactional dynamics at 23%, with many participants reporting a general gut feeling they could not articulate. People reached for style. Style is the thing the machine is best at.
What this means for how you decide
The practical conclusion is not “trust nothing” and it is not “get better at spotting AI.” The evidence says the second one is close to unavailable and the first one is not a strategy. It points somewhere more specific.
Table 5. Discernment protocol, mapped to the evidence that motivates each step
| Step | What to do | Evidence that motivates it | Status |
|---|---|---|---|
| 1. Stop trying to identify the author | Drop source-detection as a quality filter entirely | 46.6% accuracy across 16,340 judgements; expertise model R-squared 0.012; no effect of LLM knowledge or chatbot use; 4,600 participants unable to detect generated self-presentations | Evidence-backed |
| 2. Discount your confidence, not your judgment | Treat high certainty about provenance as a warning sign rather than a signal | Confidence significantly negatively related to accuracy in one study; the self-estimate finding did not replicate across populations in the other | Evidence-backed, with a replication caveat |
| 3. Evaluate blind where you can | Strip bylines before judging work you will act on | Label effect on quality rating d = -0.508, roughly 76% of the authorship effect | Evidence-backed |
| 4. Probe adversarially, do not interview | Push at edges and inconsistencies rather than asking sincere questions | Strange prompts and jailbreak attempts were among the most accurate strategies; weather and human-experience questions among the least | Evidence-backed direction |
| 4b. Distrust the folk tells | Ignore first-person pronouns, contractions and personal or family detail as evidence of human authorship | Those are the exact heuristics identified as flawed, and shown to be predictable and manipulable | Evidence-backed |
| 5. Grade on the dimension fluency cannot fake | Ask what is genuinely new here, not whether it reads well | Machine work won 13 of 14 dimensions; originality was the sole exception and showed no significant advantage | Evidence-backed |
| 6. Judge decisions by their consequences, not their prose | Track what the recommendation predicted and whether it happened | Not measured by any source cited here | CEOtudent convention |
CEOtudent editorial framework. Steps 1 to 5 restate findings the two cited studies actually produced. Step 6 is deliberately labelled a convention: no source in this article measured it, and it is offered as reasonable structure rather than as an evidence-backed technique.
Step 5 is the one that changes daily behaviour most. Both studies suggest evaluators are running a fluency check and calling it a judgment check. Fluency is now free. So the only dimension left that carries information is the one where the machine showed no measurable advantage: whether the work contains something that was not already in circulation. That question is answerable, but only by someone who has read enough of the surrounding material to know what was already there, which is precisely the capability our companion piece on deep reading as a competitive advantage argues is becoming scarce.
The CEO and student split resolves cleanly here, and not in the direction people expect. The CEO instinct is to trust judgment, because judgment is what an executive is paid for. The student instinct is to check whether the judgment is calibrated. On this specific capability the data is unambiguous that the instinct is uncalibrated, and that being sure about it makes matters slightly worse. Our note on trust calibration covers the decision rule; this piece covers the perception failure that runs underneath it. They are different problems. You can have a good policy for when to trust an AI recommendation and still be unable to tell what you are looking at.
What would change this conclusion
Three things, worth watching rather than assuming.
The poetry result does not sit in an undisputed literature, and the authors say so themselves. They note that their findings contrast with earlier work in which participants were able to distinguish professional poets from human-out-of-the-loop generated poems, and with other work finding participants merely at chance rather than below it. Earlier studies also generally found generated poems evaluated more negatively, the opposite of what these experiments produced. The most likely reading is that the models moved, not that the earlier researchers were wrong, but that is an interpretation and not something the data establishes.
Both studies use general-population samples doing an unfamiliar task under artificial conditions. Neither tested whether a domain expert evaluating work in their own domain, with time and stakes, performs differently. Porter and Machery tested poetry experience among poetry readers, which is closer, but self-reported poetry background is not the same as professional expertise. The null on expertise is real and replicated, but it is a null on the expertise these studies measured.
The Jones and Bergen result is a preprint at the time of writing, and the two populations disagreed meaningfully on several exploratory measures. The LLaMa persona condition, for instance, was 45% among undergraduates and 65% among Prolific workers, a 20-point spread on the same condition. Population effects here are not small.
And both studies describe a specific generation of systems. If detection tooling, provenance standards, or content credentials become reliable and widespread, step 1 of the protocol changes from “stop trying” to “check the signature.” Nothing in the current data anticipates that, but nothing rules it out either.
None of that changes what to do this quarter. Stop using authorship as a proxy for quality, because you cannot recover authorship. Strip labels before you judge. Ask what is new rather than whether it reads well.
FAQ
Are people really worse than random at spotting AI writing?
In this study, yes, and significantly so: 46.6% accuracy across 16,340 judgements where guessing would give 50%, chi-squared(1) = 75.13, p < 0.0001. The authors interpret below-chance performance plus above-chance agreement between participants as evidence of a shared but inverted heuristic, not of random answering.
Does knowing a lot about AI help?
On the available evidence, no. Jones and Bergen found no significant effect of level of knowledge about language models or frequency of chatbot interaction in either of their two studies, and note that accuracy was homogeneous even among people who research these systems. Porter and Machery found no poetry-experience variable with a significant positive effect on accuracy.
What tells do people wrongly rely on?
First-person pronouns, contractions, and family or personal topics. A computational analysis across six experiments identified these as the heuristics readers use to infer human authorship, and the authors then showed experimentally that because the heuristics are predictable they can be targeted, producing text rated as more human than human. If you catch yourself thinking a passage feels human because it says “I” and uses contractions, that is the documented failure mode.
So confidence is always inversely related to accuracy?
No, and the honest answer is messier. Porter and Machery found a significant negative effect of confidence (b = -0.021673, p < 0.0001). Jones and Bergen found self-estimated accuracy positively correlated with real accuracy among undergraduates (p = 0.03) but not among Prolific participants (p = 0.45). Confidence is not a reliable signal in either direction, which is the usable conclusion.
Does this mean AI writes better than humans?
No. It means readers rated it higher on the dimensions that reward smoothness, 13 of 14 of them, with rhythm the largest at d = 0.847. On originality, the one dimension requiring the reader to detect something new, genuinely AI-authored poems showed no significant advantage (d = 0.040, p = 0.098). That is a finding about what readers reward, not about literary merit.
Should I label AI-assisted work?
That is an ethics question this data cannot settle, but it can tell you the cost. Labelling work as AI-generated reduced quality ratings by 0.814 points on a seven-point scale, d = -0.508. Readers penalise the label independently of the work. Anyone deciding on disclosure should price that in rather than be surprised by it.
What is the single most useful change?
Evaluate blind. The label effect is roughly three-quarters the size of the actual quality effect, and it is the one variable you fully control. Removing bylines before judging work you intend to act on costs nothing and removes a bias of measured size.
Sources
Porter and Machery. AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports.
Jones and Bergen. Large Language Models Pass the Turing Test. arXiv preprint, Computation and Language.
Jakesch, Hancock and Naaman. Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 2023.
Kobis and Mossink, human-in-the-loop and human-out-of-the-loop paradigms for evaluating machine-generated poetry, as characterised in Porter and Machery.
Note on scope: the combined participant figure of 6,518 across the three detection studies is a CEOtudent sum of published sample sizes, used only to indicate scale. The three studies differ in design, population and outcome measure and are not pooled statistically anywhere in this article.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
