DeepMind-linked study finds perceived AI authorship can reduce agreement with moral judgments
A PLOS One study highlighted by Google DeepMind finds people can distinguish human and LLM moral justifications only imperfectly, while perceived AI authorship itself can reduce agreement.
What the study tested
Researchers Basile Garcia, Crystal Qian and Stefano Palminteri compared how people respond to moral and non-moral justifications written by humans and large language models. Participants were asked both to identify the likely source and to say whether they agreed with the judgment.
What they found
Source detection was better than chance but remained below 75%. The researchers did not find a general preference for human-written judgments; in some especially difficult moral scenarios, machine-generated justifications were preferred. A separate pattern was more striking: people were less likely to agree with a judgment when they believed it had been generated by AI, regardless of the true source.
The study also links detection and agreement to linguistic cues such as response length, typos, first-person language and cost-benefit wording. That matters for AI product design because user trust may depend not only on model quality, but also on assumptions about who—or what—produced the advice.
Why it matters
As LLMs move into higher-stakes decision support, evaluations may need to measure human perception and source-attribution bias alongside model accuracy and safety. The result is a useful reminder that transparent AI systems can face a trust problem even when their reasoning is judged comparable to human reasoning.
The paper was published in PLOS One on 5 August 2026 and is listed by Google DeepMind in its research publications.
This article is built from the source material below. Open the originals for full context and the latest updates.
