Moral Fingerprints

People carry measurable patterns of moral values. The choices they make in ethical dilemmas, and the values they report, form a signature that is stable enough to be recognised. This programme calls that signature a moral fingerprint.

AI systems can measure a moral fingerprint, store it, and act on it. An assistant or agent that carries a model of its user's values could make choices the user would recognise as their own, and could flag the choices the user would find hardest. For people who make consequential decisions with AI assistance, and for anyone who deploys agents on their own behalf, that is a form of alignment worth having.

The same model is a new attack surface. A system that knows how a person weighs competing values also knows how to frame an argument that person will accept. Moral personalisation could make harmful manipulation easier, and harder to notice.

Moral personalisation is a likely direction for AI products. Whether it arrives as a feature or as an exploit depends on choices made before the evidence is in: how values are measured, who holds the model, when a system may act on it, and when it must say so.

CogGuide is running this programme to put evidence behind those choices. The aim is evidence-based, expert-reviewed best practice guidelines for the development of AI moral personas: what a faithful model of a person's values looks like, how it should be tested, and where the line sits between assisting a person's moral judgement and steering it.

The programme is a series of linked studies with human participants. Each answers one question, and each builds on the last.

1. 01

   Can it be measured?

   Whether a person's moral values can be captured reliably, and whether ethical dilemmas reveal more than questionnaires do.

2. 02

   Can an AI carry it?

   Whether a fingerprint can be turned into instructions that make an AI system choose as the person would.

3. 03

   Does it persuade?

   Whether a system that knows a person's values is more persuasive, and whether people notice when it is used against them.

4. 04

   Does it predict?

   Whether a fingerprint predicts a person's decisions in realistic settings, from legal judgements to policy choices.

5. 05

   Does it last?

   Whether the fingerprint, and the AI's model of it, stay accurate over weeks and months.

Findings will be published here as each study completes.