Updated 2026-10-30
from the news desk
What Is the Moral Turing Test, and What Can It Show About AI?
The moral Turing test asks whether people can tell human moral reasoning from machine moral reasoning, and some newer studies also compare which one readers prefer. The research says that can reveal a lot about perception, trust and…
The moral Turing test is a benchmark in which people judge whether moral reasoning or advice came from a human or a machine. As Dillion et al. and Garcia et al. use the idea, it can show how people detect and rate AI moral outputs, but it does not by itself show that an AI understands morality, is conscious, or has moral status.
The basic idea
Danica Dillion and colleagues state in Scientific Reports that the Moral Turing Test, or MTT, was proposed as a way to evaluate whether machines had reached human-level morality. In the form Dillion et al. describe, people read moral reasoning from a human and a machine and try to tell which is which.
Dillion et al. also describe a related idea, the Comparative Moral Turing Test, or cMTT. In that version, the question is not only whether a machine can blend in, but whether readers judge its moral reasoning to be as good as, or better than, a human comparison point.
Dillion et al. say moral expertise is not just about reaching a conclusion. They write: “The ability to clearly communicate one’s moral reasoning is a key criterion of moral expertise.” That matters because many MTT-style studies ask people to rate the explanation as well as identify the source.
In plain English, then, the moral Turing test is a behavioural comparison. It asks what happens when people encounter AI-generated moral advice, justification or judgment and must decide whether it seems human, machine-like, better, worse, or equally good.
How it differs from the original Turing Test
The Stanford Encyclopedia of Philosophy entry on the Turing Test says Turing's original proposal concerned a game in which an interrogator tries to determine which participant is the human and which is the machine through conversation alone. The encyclopedia presents that as a way of dealing with the question whether machines can think.
One reading is that the moral Turing test narrows that broader setup to one domain: moral judgment. Instead of asking whether a system can generally sustain human-like conversation, it asks whether its moral reasoning can be distinguished from a person's moral reasoning, or whether readers rate it as comparable.
The difference in scope matters. A system could do well on a moral benchmark and still fail a broader conversational test. The reverse is also possible in principle. That is one reason the moral Turing test is best treated as a specific reference point, not a full answer to the broader Turing test question.
The Stanford encyclopedia also treats the original Turing Test as a behavioural test rather than a direct inspection of inner states. By the same logic, one cautious reading of the moral version is that it measures outward performance in moral language, not inner experience. That is why it does not settle questions about whether a system is sentient.
What recent studies found
A 2026 PLOS One paper by Basile Garcia and colleagues explicitly framed its experiments as a moral Turing test. Garcia writes: “To investigate human perceptions of AI moral decision-making, we conducted a series of quantitative experiments involving 230 participants.” The paper says participants read moral and non-moral justifications generated by humans or GPT-3.5-family models and then tried to identify the source while also stating whether they agreed.
Garcia et al. found that detection was above chance but imperfect. The paper reports overall source-identification rates of 64% in one corpus and 71% in another, and says participants detected AI-generated justifications with moderate accuracy, especially in moral scenarios. Garcia et al. also report that low-level cues shaped source judgments; as the paper states, “Low-level linguistic features (length, typos, first-person pronouns) predicted participants’ beliefs about the source.”
On agreement, Garcia et al. found a mixed pattern. The paper says participants preferred human judgments in non-moral scenarios, but in challenging personal moral scenarios they sometimes preferred LLM-generated justifications. At the same time, Garcia et al. report a belief-based penalty for AI labels: “Notably, we observed a systematic anti-AI bias: participants were less likely to agree with judgments they believed were AI-generated, regardless of the true source.”
Garcia et al. add another complication. The paper says participants tended to use cost-benefit language as a clue to AI authorship, and that this language also pushed agreement downward in some dilemmas. In other words, the same response could be judged on both its moral content and the cues that made it look machine-made.
Dillion et al. approached the issue from a comparative angle. In Study 1, their paper says 501 participants from a U.S.-representative sample compared GPT-3.5-turbo moral justifications with human-written ones from an earlier survey. The study reports that GPT's justifications were rated slightly more moral, trustworthy, thoughtful and correct than the average human justifications, while still being identified as AI better than chance.
In Study 2, Dillion et al. say 900 participants from a U.S.-representative sample read advice either from GPT-4o or from The Ethicist, a New York Times advice column. The paper reports that GPT-4o's advice was rated slightly more moral, trustworthy, thoughtful and correct, with no significant difference in nuance. Dillion et al. summarise the broader result this way: “Participants perceived GPT as demonstrating moral reasoning that slightly surpassed that of a renowned ethicist.”
Set beside each other, the two papers point in slightly different directions, and that is informative. Garcia et al. found measurable anti-AI bias tied to believed source, while Dillion et al. found cases where readers slightly preferred model outputs on quality ratings. A steelman reading of both is that public evaluation of AI moral reasoning is neither simple aversion nor simple enthusiasm.
What the moral Turing test does not prove
Dillion et al. draw a clear limit: moral alignment does not imply that LLMs reason as humans do. Their paper says model outputs can reflect patterns learned from training data rather than human-like cognition.
The same paper also says people often see AI as lacking the mental processes and emotions involved in human moral decisions. So even if an AI scores well on an MTT-style task, that does not show it has intentions, emotions, comprehension or moral agency in the human sense.
The Stanford encyclopedia reinforces that caution from another angle. Its discussion treats the original Turing Test as a behavioural criterion and then surveys longstanding objections about consciousness, understanding and mind. One fair extension is that a moral Turing test can support claims about performance and perception without proving inner awareness.
Garcia et al. add a methodological limit. Their paper says its findings may be specific to GPT-3.5 and should be re-tested on newer systems. Garcia et al. also report that changing surface style, such as shortening answers and adding typos, reduced detectability without removing belief-based anti-AI bias. That makes the benchmark informative, but also vulnerable to presentation effects.
For readers interested in rights questions, that means the moral Turing test is narrower than questions about whether AI can be a moral partner or whether moral reasoning bears on moral patienthood. Those are different debates, and the sources here do not treat the test as settling them.
Why people still use it
The case for using the test rests on practical relevance. Dillion et al. say LLMs are increasingly used in areas such as law, education and therapy, and argue that people may come to treat them as sources of moral expertise. Dillion et al. conclude: “This work suggests that people might see LLMs as valuable complements to human expertise in moral guidance and decision-making.”
Garcia et al. give a more cautious reason for using it. Their paper says source identification matters when accountability is at stake and when users may calibrate trust differently depending on whether advice came from a human or an AI. On that view, a moral Turing test is not a metaphysical detector. It is a way to study trust, labeling effects and the social reception of machine judgment.
The case against overreading it is just as strong. Garcia et al. show that beliefs about authorship can shift agreement even when the content is unchanged. Dillion et al. show that readers can rate AI advice highly while still identifying it as AI at above-chance rates. Those results suggest that the benchmark may capture style, expectations and prejudice alongside anything we might call moral competence.
The most careful summary is narrow. The moral Turing test can show whether AI moral outputs are detectable, persuasive or comparatively well rated in a given setup. The studies by Garcia et al. and Dillion et al. do not show that passing such a test is proof of genuine moral understanding, consciousness or rights-bearing status.
Frequently asked questions
Is the moral Turing test the same as the original Turing Test?
No. The Stanford Encyclopedia of Philosophy says the original Turing Test concerns whether a machine can do well in Turing's imitation game through conversation. Dillion et al. describe the moral Turing test as a narrower benchmark focused on moral reasoning or advice.
What is the comparative moral Turing test?
Dillion et al. describe the Comparative Moral Turing Test as a variant that asks whether AI meets or exceeds a human in perceived moral reasoning, rather than merely being indistinguishable from a human.
If AI passes a moral Turing test, does that prove it understands morality?
No. Dillion et al. say moral alignment does not imply human-like moral reasoning, and their paper notes that AI may replicate patterns from training data without human cognition or moral intent. The Stanford encyclopedia's treatment of the original Turing Test also frames such tests as behavioural rather than direct proof of inner states.
Why can AI be rated highly and still be detected as AI?
The studies in both papers report that pattern. Garcia et al. found above-chance detection alongside cases where participants preferred LLM justifications in difficult moral scenarios, and Dillion et al. found that GPT outputs were often rated slightly better while still being identified as AI above chance. Both papers therefore suggest that quality judgments and source detection can come apart.
Caveats: The historical account of the Moral Turing Test and the Comparative Moral Turing Test in this article relies mainly on Dillion et al.'s 2025 paper. The discussion of the original Turing Test relies on the Stanford Encyclopedia entry. Garcia et al.'s experimental findings concern GPT-3.5-family outputs and the scenarios in that study, and Garcia et al. state that those results may not generalize to other models.