AI ethical advice became the preferred option when people could judge the words without knowing who wrote them. In a preregistered study of 642 participants, AI-generated guidance was chosen for 53.7 per cent of the dilemmas when its source was hidden. When the same advice was labelled as AI-generated, its choice rate fell to 46.8 per cent. The result does not show that a machine is objectively more ethical than a person. It shows that the label changed how people valued otherwise identical advice.
The experiment, published as a Registered Report in Scientific Reports in March 2026, tested a form of algorithm aversion. People began with a strong preference for human help. After they saw the quality of both answers, that aversion narrowed substantially. Source disclosure then moved the choice back towards the human expert, even though the content being compared had not changed.
How the AI Ethical Advice Study Was Designed
Researchers Lennart Meincke, Gideon Nave and Christian Terwiesch used twenty dilemmas originally published in The New York Times column The Ethicist. They compared answers written by the human columnist with responses produced by GPT-4. Each participant in the main study saw five dilemmas selected at random.
The work began with a pilot involving 187 laypeople, MBA students and a smaller panel of scholars and clergy. Participants rated the usefulness of advice without knowing its source and showed no significant difference between AI and expert guidance. In direct choices, 57 per cent selected the AI answer. The strongest statistically significant preference appeared among the lay participants, who selected it 59.6 per cent of the time.
The larger main study randomised participants across three conditions. One group saw only the dilemmas and said whether they would prefer advice from a human expert or an AI. A second group saw both answers with their sources disclosed. A third saw both answers without source labels. The published study reports 658 recruits and a final analysed sample of 642 after exclusions.
People Wanted a Human Before They Read the Advice
Before any answer was shown, 72.6 per cent preferred a human adviser. Put the other way, AI was selected in only 27.4 per cent of the dilemmas. That is the cleanest measure of the reputation gap. Participants were not rejecting a weak answer because they had not seen one. They were rejecting the idea of algorithmic ethical guidance in advance.
Once labelled answers were available, the AI choice rate rose to 46.8 per cent. People still leaned towards the human source, but much less strongly. The quality of the actual advice displaced a large part of the initial aversion. In the unlabelled condition, the AI rate rose again to 53.7 per cent, a statistically significant increase over full disclosure.
This pattern is different from simple deception. The researchers did not falsely label the AI answer as human. They either supplied the real labels or supplied no labels. The hidden-source condition therefore tested how the content performed when provenance was removed from the judgement. The result suggests people discounted the AI answer partly because it was AI, not because of a difference they consistently detected in the words.
Why Source Disclosure Changed the Choice
Ethical advice is not only information. It is also a relationship of trust, responsibility and perceived understanding. A human adviser can be held accountable, can ask follow-up questions and has lived experience of the norms being discussed. A participant may therefore rationally value human authorship even if one isolated answer reads no better than the machine’s.
At the same time, the experiment shows how easily reputation can overwhelm content. People may carry a general belief that AI cannot understand morality, then revise that belief after encountering a competent answer. The label can trigger assumptions about coldness, manipulation or lack of accountability before the advice itself is weighed.
The reverse risk also matters. An AI system trained to sound supportive can produce polished guidance that feels wise without recognising missing facts or the consequences of being wrong. LiveAIWire’s analysis of AI sycophancy shows that models can affirm a user’s conduct more readily than humans. Ethical advice that pleases the person asking is not necessarily advice that deserves to be followed.
The Study Tested Preference, Not Moral Authority
The paper’s central outcome was which answer people chose and how useful they rated it. It did not establish that GPT-4 reached morally correct conclusions, because many personal dilemmas do not have a single measurable correct answer. It also did not follow participants to see whether they acted on the guidance or whether the outcomes were beneficial.
The answer sets had structural differences too. The authors note that the human expert tended to give clearer directives, while the AI was more abstract. Participants might prefer one style for reasons unrelated to ethical quality. The dilemmas came from a newspaper advice column, so the findings cannot automatically be transferred to law, medicine, safeguarding or crisis decisions where professional duties and missing context are more important.
The study’s registered protocol strengthens confidence that the main hypotheses and analysis were specified in advance. Registration does not remove every limitation, but it reduces the freedom to design the explanation after seeing the result. The journal page also describes the work as a Registered Report and gives the acceptance date for its first-stage protocol.
AI Ethical Advice Creates a Disclosure Dilemma
Hiding the source made the AI answer more popular, but that is not an argument for hiding it in real products. Provenance can be relevant to a user’s decision, especially when an answer has no human professional standing or when a platform has commercial incentives to deepen engagement. The ethical way to respond to algorithm aversion is to improve evidence, accountability and explanation, not to make the algorithm invisible.
Disclosure should also be specific enough to be useful. A generic statement that AI assisted the answer says little about who chose the training data, whether a human reviewed the output or what happens when the system encounters danger. The design questions become more serious in companion products, where LiveAIWire’s reporting on AI companions and loneliness found that immediate comfort can coexist with unresolved long-term risks.
Users deserve both the content and its chain of responsibility. A service can identify that an answer was generated by AI, explain that it is not a professional judgement, show relevant sources where possible and provide a route to a human for consequential decisions. That approach may preserve some aversion, but it lets people decide with the information that genuinely matters.
What the Result Says About Trust in AI
Trust is more malleable than the opening preference suggested. Most participants initially wanted a human, yet experience of the answer nearly doubled the AI choice rate even when the label remained visible. This is evidence that people can update their view of algorithmic guidance rather than rejecting it categorically.
The magnitude may also depend on culture, familiarity with chatbots and the kind of dilemma presented. The analysed participants were recruited through an online platform and responded to a fixed set of English-language advice-column scenarios. A workplace conflict, a family obligation and a question involving religious duty may not elicit the same source preferences. Replication across communities would show whether the disclosure effect is broad or concentrated in particular audiences.
Scale changes the stakes even when an individual answer seems benign. A human columnist advises one reader at a time and carries a recognisable editorial identity. A chatbot can produce millions of personalised moral responses, remember the conversation and optimise for continued engagement. Small tendencies towards abstraction, affirmation or risk avoidance can therefore become systematic influences on behaviour.
It is also evidence that trust cannot be inferred from output quality alone. A person may prefer a human because responsibility matters, or prefer the AI because its abstraction feels impartial. Neither preference proves the underlying advice is sound. Good systems should therefore support scrutiny instead of optimising only for whether the user chooses or likes the response.
The striking part of the study is the split between what people expected and what they selected. Before reading anything, nearly three quarters wanted the human. When authorship disappeared, the AI answer won a small majority of choices. The words were capable of changing minds, but so was the label attached to them. Any serious use of AI ethical advice has to account for both.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.