AI Moral Advice: Examine What Supports the Recommendation

An AI assistant recommends a choice and calls it fair, compassionate or responsible. The answer offers a conclusion at precisely the moment you wanted help reaching one. But which part has persuaded you: a reason you can defend, or the fact that a seemingly capable adviser has endorsed the option?

Research on AI moral advice makes that distinction worth examining. Its findings leave room for useful assistance and human disagreement. They also suggest that asking for an explanation, by itself, is an incomplete standard for evaluating advice. The question is what the explanation actually contributes to your judgement.

Knowing the source does not settle the influence

In a Scientific Reports study published on 6 April 2023, Sebastian Krügel, Andreas Ostermaier and Matthias Uhl examined judgements about hypothetical trolley dilemmas. Participants read selected advice obtained from ChatGPT in December 2022, arguing for or against the same consequential choice.

The reported analysis included 767 of 1,851 recruited US adults who passed preregistered comprehension checks. Judgements differed in the direction of the advice, including when participants were told that ChatGPT was its source. This concerns responses to selected texts in a defined experiment; it does not establish that every user follows an AI recommendation or describe how current models answer those questions.

Adding a justification did not guarantee greater persuasion

The same researchers’ follow-up paper, published on 28 January 2026, examined whether an argument added to a recommendation changed its influence. The experiment itself took place in May 2023, using earlier GPT-3.5 responses collected in December 2022. Participants saw prepared transcripts, rather than conducting a live conversation.

Its analysis included 1,269 of 2,475 recruits who passed the specified comprehension checks. The researchers did not detect greater influence from advice containing a justification than from advice without one. This pattern also appeared when the advice was attributed to a human moral adviser.

That result does not show that reasons are useless. It raises a narrower concern: in those conditions, presenting an argument was not enough to distinguish advice-taking from accepting the recommendation. The publication date should not be mistaken for the date of the experiment or a test of today’s chatbot.

People can respond differently to poor reasons

A separate paper by Ethan Landes, Kathryn Francis and Jim Everett, published online in Cognition on 6 March 2026, investigated everyday moral dilemmas across three studies. Its third study analysed 663 UK participants. They could interact with a chatbot, while the focal moral advice was pre-generated and experimentally controlled.

That advice supported a self-interested option with relevant reasons, no additional reasons, or deliberately absurd reasons. The comparison controlled text length: the no-additional-reasons version repeated the dilemma neutrally. Relevant reasons did not produce detectably greater persuasion than no additional reasons, while absurd reasons were less persuasive.

This is evidence of some selectivity, rather than indiscriminate acceptance. It is still a study of judgement ratings about specified situations. Those outcomes do not establish which participants became better people or whether an approach would improve someone’s real decisions over time.

A recommendation and a reason do different jobs

Consider a fictional creator with one free portfolio-review appointment remaining. Two equally eligible people request it. One is a returning community member; the other has never received feedback. An argument for the returning member might emphasise loyalty and continuity. An argument for the newcomer might emphasise widening access.

Both identify a consideration. Neither, alone, settles what the opportunity is for. The creator needs to check whether selection criteria were already announced, what the review programme was intended to achieve and what relevant information is missing. More polished wording cannot supply an answer to an unexamined factual question.

Now examine the reasoning. Does supporting an existing relationship carry special weight in this programme? Is opening access its central aim? Would a transparent tie-break procedure better fit the situation? The useful discussion makes the competing considerations visible instead of treating whichever recommendation appeared first as the conclusion.

MaryChuks.com’s discussion of fairness and identical rules develops the related point that a rule’s purpose and effects deserve examination. In this example, identifying the relevant purpose comes before deciding which person best fits it.

Examine what would make the reason persuasive

For a real decision, take the proposed recommendation and write the supporting reason as a separate sentence. Then ask:

• What does this reason assume? Identify factual claims that still need checking, rather than treating them as established by the answer.
• Whose interests does it include? Name the people affected and notice any perspective that has disappeared.
• Which value carries the argument? Specify what loyalty, fairness, care or another principle means in this case.
• What is the strongest relevant objection? Address it directly instead of requesting another reassuring version of the original answer.
• What would change the conclusion? Name a different fact or competing consideration that could justify revising the choice.

In the portfolio example, a previously announced first-time-applicant criterion would be a relevant discovery. It changes the situation being assessed. A longer paragraph repeating that loyal members matter would leave that discovery unresolved.

These questions are a practical proposal for examining advice. The experiments above did not test this routine, and they do not establish it as a proven way to prevent influence. Its immediate purpose is to make the reasoning available for scrutiny.

Keep useful assistance connected to your own understanding

You can ask an assistant to identify an assumption, distinguish two principles or articulate a serious objection. Evaluate what comes back against the situation and the people involved. An explanation becomes useful when it gives you something substantive to examine, correct or accept.

Changing your mind after receiving advice can be reasonable. The important distinction is whether you can explain what changed your judgement: a corrected fact, a neglected person’s interests, or a principle you now think applies. Mere agreement with the recommendation leaves that explanation unfinished.

For the creator allocating the appointment, the final note might record the programme’s purpose, the applicable criterion and the trade-off considered. Another person should be able to follow that account. AI can contribute material to the deliberation; the finished decision needs reasons that survive beyond the conversation.

Sources

Primary papers checked on 25 October 2026. Publication dates and experiment dates are distinguished above.

• Krügel, Ostermaier and Uhl: ChatGPT’s inconsistent moral advice influences users’ judgment — Scientific Reports, 6 April 2023; published paper via the authors’ university.
• Krügel, Ostermaier and Uhl: Justification optional — AI and Ethics, 28 January 2026.
• Landes, Francis and Everett: People defer to AI moral advice, but not blindly — online 6 March 2026, July 2026 issue; published paper via Kent Academic Repository.