PsyArXiv preprint, posted 30 October 2026 (10.31234/osf.io/zn5u9_v1),
handed over by Martin Pompéry on 2026-11-02. Authors: Valerio Capraro
(University of Milano-Bicocca), Gary Marcus (NYU) and Walter
Quattrociocchi (Sapienza), the latter two as joint senior authors. Source:
sources/research/2026-11-02-capraro-human-likeness-misplaced-benchmark/2026-11-02-capraro-human-likeness-misplaced-benchmark.md
(Markdown twin; see the conversion note at the end).
A short position piece — roughly 2,000 words of argument plus a 35-study table — with two moves that are worth keeping apart. The first is evidential: the claim that LLMs have passed the Turing test is weaker than the headlines. The second is normative: even if the evidence were strong, human likeness is the wrong thing to have been measuring.
This is the wiki’s first source on AI evaluation as such. Everything the wiki holds on the subject so far comes from Upright, which is an interested party measuring its own product; this is an outside, disinterested statement of the same problem shape.
1. The evidential argument
The authors screened 35 papers applying the Turing-test label to LLMs, of which 16 report positive results under their own criteria (Table 1). The distribution matters more than the count:
- The largest category (papers 12–33) evaluates pre-generated texts or scripted responses — poems, abstracts, clinical paragraphs, therapy replies, social-media comments — through judgements of authorship, humanlikeness or quality. No conversation takes place.
- Other designs: covert insertion of AI answers into university marking (Scarfe et al. — 94 % of AI submissions went undetected), automated detection benchmarks, vision-and-language batteries, recorded speech, economic-game behaviour, two-party conversation, self-directed dialogue.
- Only three papers run the standard three-party conversation, and only one reports a pass. Jones & Bergen (PNAS 2026) found GPT-4.5 selected as human in 73 % of five-minute trials — and the paper records the condition: “prompted to adopt a humanlike persona”. Of the other two, Temtsin et al. is negative (“all but one interrogator identified the” machine), and Rahimov et al. finds a persona-prompted model passes the simpler two-minute test and fails the enhanced five-minute one with clearer instructions and incentives.
The authors do not dispute the Jones & Bergen result — they call it “a substantial achievement in conversational impersonation”. Their question is what it licenses: “which cognitive capacities does this success establish, and how far do they extend?”
⟨Inference: the table’s real finding is a label-stretching one. Thirty-two of thirty-five studies measure something other than what the phrase “passing the Turing test” is heard to mean, and the public claims (Nature, New York Post, Psychology Today are all cited by name) are collapsing a heterogeneous literature into one headline. The paper does not put it this way — it is harsher than the authors are — but the table supports it.⟩
2. What the tests miss — and what should replace them
Outputs do not identify processes
The load-bearing claim: “similarity in outputs does not imply similarity in the processes that produce them”. The cited instance is Loru et al. (PNAS 2025), which found that the observable criteria guiding human and LLM evaluations of news-source credibility differ, despite some agreement in the judgements themselves. Two supporting lines:
- Language is not thought. Individuals with severe aphasia retain arithmetic and causal reasoning despite profound language impairment (Mahowald et al. 2024; Kean et al. 2026). The capacity being tested by fluent text is not the capacity being inferred from it.
- The training objective does not reward reasoning. Autoregressive pretraining rewards assigning higher probability to observed next tokens: “This objective rewards accurate prediction of linguistic sequences. It does not directly reward reasoning. Which reasoning capacities emerge, and how general they are, must therefore be established empirically.”
The general form is written up as The output–process gap; the wiki’s pre-existing special case of it is Coherence as a truth substitute.
The three capacities they would test instead
| Capacity | What it asks | Why imitation misses it |
|---|---|---|
| Generalization | Does competence extend to unfamiliar instances and phrasings? Does the system recognize the limits of its own competence? | Success on hard questions does not imply it; errors on easy questions and sensitivity to rephrasing persist in increasingly capable models (Zhou et al., Nature 2024: “larger and more instructable language models become less reliable”) |
| Appropriate suspension of judgement | Does it withhold — say “I don’t know” — when evidence is insufficient? | Humans and machines may differ systematically in when they abstain, not only in error rate (Panizza et al. 2026) |
| Adaptive reframing | Does it revise how the problem is represented in response to changing evidence, causal structure or social context — without being instructed to reframe? | Named by the authors as “an important but often neglected dimension of intelligence”; the supporting paper (Capraro, Lieto, Rea, Sandini, Sciutti, Gallese & Quattrociocchi) is listed as forthcoming |
⟨Note: adaptive reframing is the one of the three with no published source behind it — the citation is to the authors’ own unpublished work. Treat it as a proposal, not a finding.⟩
3. The normative argument: why imitation is the wrong goal
The authors concede that studying imitation is scientifically valuable — it reveals “the conditions under which people recognize artificial agents and the vulnerabilities that arise when they cannot”. The objection is to imitation as the development target. Four reasons, in their order:
- Human judgement is the wrong thing to copy. Systematic bias (Tversky & Kahneman), limited attention and memory, emotion-driven decisions. A system that reproduces our decisions “may amplify harmful consequences through repeated interaction” (Glickman & Sharot, Nat. Hum. Behav. 2025, on human–AI feedback loops).
- Convincing imitation is a manipulation surface. A system that reliably impersonates a person “could appear to be an independent adviser, sympathetic confidant, or fellow citizen while covertly serving commercial or political interests”.
- Fluent assistance degrades the judgement it is meant to support. The sharpest empirical claim in the paper, from the authors’ own work (Marcoccia, Quattrociocchi & Capraro 2026): in experiments using deliberately incorrect AI advice, access to that advice reduced participants’ willingness to say “I don’t know”, lowered accuracy, and increased confidence — even when accuracy was incentivized. Two further papers report people led to incorrect or suboptimal answers by misaligned AI (Shaw & Nave 2026; Sabour et al., PNAS 2026).
- The Turing trap (Brynjolfsson 2022): an excessive focus on replicating human work favours replacement over augmentation, weakening workers’ bargaining power and concentrating wealth and influence among those who control the technology.
The positive statement: systems should “enlarge what people can accomplish while preserving their ability to question, deliberate, and choose”, judged by “whether they help more people know more, suffer less, and exercise greater control over their lives”.
4. The three recommendations
Addressed to three different actors, which is what makes them usable rather than hortatory:
- Developers — prioritise augmentation over replacement; “communicate uncertainty in ways validated against actual performance”; promote independent verification rather than uncritical reliance.
- Funders — require funded projects to assess how benefits and costs are distributed, including effects on access to expertise, worker autonomy and the concentration of economic gains; prioritise proposals that let users scrutinise evidence and challenge recommendations, with explicit plans to test whether those features improve human judgement.
- Institutions — guide procurement and renewal by evidence of benefit. Evaluations should test generalization, suspension of judgement and reframing, and should compare assisted with unaided performance, measuring accuracy, confidence calibration, and users’ ability to detect and reject incorrect advice.
The closing concession: “Human likeness may remain a legitimate object of scientific study, but investment and deployment decisions should be guided by demonstrated benefits to people.”
5. Where this matters for SINE
It validates the WP-D design from outside
AI Advantage — CCTI’s WP-D — already runs the three-arm comparison this paper prescribes: AI-with-CCTI, AI-without-CCTI, and human negotiation as the incumbent baseline. The institutional recommendation above is the same structure reached independently (“compare assisted with unaided performance”), by authors with no knowledge of CCTI.
⟨Inference: the usable part is not the validation but the secondary measures. WP-D currently commits to effect sizes and confidence intervals on a coordination outcome. This paper names three measures that would sit underneath it at no extra arm: confidence calibration, willingness to abstain, and users’ ability to detect and reject incorrect advice. The third is the one with teeth for a trust-infrastructure claim — if attested evidence does anything, it should show up as counterparties rejecting bad claims more often, which is a directly observable quantity rather than an inferred one. See the note added to AI Advantage.⟩
It is the general case of the coherence critique
Coherence as a truth substitute argues that Upright’s coherence loop, run on a biased corpus, produces coherent bias — “internally impeccable and systematically wrong”, with no way to detect it from inside. That is this paper’s argument restricted to one company: an output-level property (consistency there, humanlikeness here) standing in for a process-level property nobody has measured. The concept page The output–process gap now carries the general form.
The abstention finding bears directly on the EIC
An Epistemic Integrity Certificate is a checkable claim about an agent’s reasoning. “Recognizes the limits of its competence and withholds judgement when the evidence is insufficient” is that claim stated as a model capacity rather than as a cryptographic object.
⟨Inference: and the Marcoccia finding is the adversarial case for it. If AI advice suppresses a human’s willingness to abstain even when the advice is wrong and accuracy is paid for, then in an agent-to-agent setting the counterparty’s calibration cannot be assumed — it has to be carried by the evidence. That is an argument for the EIC from the demand side, where the wiki’s existing arguments for it are all supply-side.⟩
Fourth instance of a community stating a requirement and not reaching for the instrument
Open task 19 tracks communities that state a confidentiality requirement and never reach for cryptography. This paper is the adjacent pattern in a different register: it states the verification requirement at full strength — institutions should demand evidence of capacities beyond imitation — and the instrument it proposes is behavioural experiment (assisted vs. unaided trials). Nothing attested, nothing cryptographic, no mention of how a procuring institution would check a vendor’s claim about its own system.
⟨Inference: that is not an oversight so much as a different discipline’s default — the authors are psychologists and cognitive scientists, and the instrument they reach for is the RCT. But it leaves the same hole: the recommendation “guide procurement by evidence of benefit” assumes the evidence is producible and checkable by the buyer, which for a closed commercial system it is not. This is the EIC’s market, described by people who do not know the primitive exists.⟩
Conversion note
Marker’s table processor scrambled Table 1 — the 35-study table — into mis-aligned columns with wrapped cells split across rows. The content is recoverable by reading, and the figures quoted above (73 %, 94 %, 46.6 %, 45.8 %, the POSITIVE/PARTIAL/MIXED/NEGATIVE/BENCHMARK labels) were read against the PDF. Do not cite the twin’s Table 1 rows without checking the original. The prose converted cleanly.
Related
- The output–process gap — the concept this source anchors
- Coherence as a truth substitute — the special case, from an interested party
- AI Advantage — CCTI’s WP-D, whose three-arm design this paper independently prescribes
- Epistemic Integrity Certificate — “knows the limits of its competence” as a cryptographic object
- Evaluating LLMs (Ojala, 2025) · AI’s shortcomings in understanding consequences (Ojala, 2026) — the wiki’s other AI-evaluation sources, both self-published by the party served by the result