This one is going to make some noise. Just the way we like it. Because it brings the AI debate back to reality.
Together with Gary Marcus and Valerio Capraro, we've just released a new paper: "Human… | Walter Quattrociocchi | 12 comments
Walter Quattrociocchi announces a paper co-authored with Gary Marcus and Valerio Capraro, titled “Human likeness is a misplaced benchmark for artificial intelligence.” The authors reassess claims that large language models have passed the Turing test and thereby demonstrated human-like intelligence.
They reviewed 35 studies that used the Turing test or related methods to evaluate language models. Although 16 reported positive results, only three used the classical three-party version of the test, and just one reported a pass. In that study, GPT-4.5, specifically instructed to impersonate a human, was judged human in 73% of five-minute trials.
The authors argue that producing human-like responses does not prove human-like intelligence. Surface similarity reveals little about whether a system can generalize reliably, recognize uncertainty, revise its approach, or adapt when circumstances change. They also question why human imitation should be the central measure of AI progress, given that humans are biased, inconsistent, manipulable, and often wrong.
Instead, they propose evaluating AI according to measurable benefits, reliable generalization, independent judgment, effective error correction, and support for human autonomy and welfare. They warn that systems optimized for convincing impersonation may increase manipulation risks and encourage replacement rather than augmentation of human abilities.
The accompanying comments broadly support this position, emphasizing that intelligence need not resemble biological cognition and that practical usefulness matters more than sounding human. Several commenters also distinguish successful imitation from genuine transformation, understanding, or thought.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Gary Marcus, Valerio Capraro, Alan Turing, and Sapienza University of Rome
🤖 AI prefill
📰What
Walter Quattrociocchi announces a paper co-authored with Gary Marcus and Valerio Capraro, titled “Human likeness is a misplaced benchmark for artificial intelligence.” The authors reassess claims that large language models have passed the…
🤖 AI prefill
📅WhenWhen did this happen, or what is the timeframe?
📍Where
Rome
🤖 AI prefill
💡Why
Centers on human likeness as a benchmark for artificial intelligence, Turing test, and large language models
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Turing Test EvidenceTuring test, imitation game, GPT-4.5, human identification, 35 papersBeyond Human Likenesshuman likeness, intelligence benchmark, reliable generalization, adaptive reframing, error correctionHuman Fallibility And AI Risksbias, inconsistency, manipulation, human judgment, impersonationAI Human Benefithuman welfare, human autonomy, measurable benefits, helpfulness, verification
Key Phrases:human likeness as a benchmark for artificial intelligence99%Turing test98%large language models95%classical three-party Turing test92%GPT-4.5 human impersonation90%human-like intelligence91%reliable generalization90%independent judgment86%adaptive reframing82%human autonomy82%
Manipulation Score: LowHow much this article uses emotional framing vs. straight reporting
Loaded Language2425%
Absolutist1815%
Threat/Panic2025%
Us vs Them1015%
Engagement Bait4220%
Manipulation score look right?
Factual Claims (10)
mediumscientific
The authors released a paper titled "Human likeness is a misplaced benchmark for artificial intelligence."
mediumscientific
The paper examined 35 papers claiming to use the Turing test or a variation of it to evaluate large language models.
mediumscientific
Sixteen of the examined papers reported positive results.
mediumscientific
Only three of the 35 papers included the classical three-party Turing test.
mediumscientific
Only one examined study reported passing the classical three-party Turing test.
hightechnology
In that study, GPT-4.5, explicitly prompted to impersonate a human, was identified as human in 73% of five-minute trials.
mediumscientific
The authors' previous research, published in PNAS, found that human and artificial judgments can converge on similar answers while using different observable criteria.
mediumscientific
Human-like outputs do not establish that an AI system can generalize reliably, recognize the limits of its knowledge, or reconsider its approach when circumstances change.
mediumtechnology
Systems designed to convincingly impersonate people create new opportunities for manipulation.
mediumscientific
The authors advocate measuring reliable generalization, independent judgment, human autonomy, and measurable benefits to people rather than human imitation.
Comments
Login to view and post comments.