Reputation when the reader is a machine

Imagine being introduced by someone who has read widely about you, retained an uncertain mixture of old and new information, and feels no social discomfort about summarising it with confidence. That is an interesting addition to the reputation business.

As AI systems help people research organisations and individuals, an introduction may begin with a generated answer. The reader might ask who has relevant experience, which organisation understands a problem, or whether a particular claim is credible. What appears in that answer could shape whether a conversation happens at all.

I think this makes the quality of the public record more important. It also makes it tempting to chase the wrong thing: finding a way to persuade the machine to say something flattering.

That temptation is understandable. Search already taught people to care about visibility, and generated answers can feel like a more concentrated form of it. But an answer can vary with the question, the available sources and the system involved. A favourable response to one carefully chosen prompt says little about how reliably someone will be represented elsewhere.

Google’s guidance on AI features in Search offers a useful dose of sobriety. It says familiar search practices remain relevant, important information should be available as text, and no special AI markup is required for those features. It also makes clear that eligibility does not guarantee inclusion. This guidance concerns Google’s systems; it is no universal formula for every model that might describe you.

The practical implication seems fairly ordinary. Make it easy to establish who did what, when they did it and where the evidence lives. Use clear names. Date material. Explain the scope of a project. Correct information that has become misleading. Someone encountering your work should be able to distinguish a completed result from an aspiration without conducting a small archaeological expedition.

This matters because fluency can hide weak foundations. NIST’s generative AI risk profile describes confabulation, where a system confidently generates false or erroneous content. A well-written answer about a person or organisation can therefore still be wrong. The presence of a citation is useful only if the cited material actually supports the statement beside it.

I would treat this as a reason to improve evidence and verification, rather than as an invitation to flood the internet. Publishing the same unsupported claim in ten places gives a reader more copies of one claim. It does not create ten independent confirmations. A machine may struggle with that distinction; people should avoid exploiting the struggle.

Provenance offers a better way to think about the problem. The W3C’s PROV framework describes information about the people, activities and entities involved in producing something, helping others assess its reliability. Most public writing does not need a formal provenance system. It can still adopt the underlying discipline: preserve the route back to the original work.

For example, a claim about a technical contribution is more useful when it links to the contribution, identifies the author’s role and explains the context. A claim about a result becomes more credible when the method and limitations are available. A correction becomes more helpful when readers can see what changed. These details help humans immediately, whatever an AI system eventually does with them.

There is a boundary around privacy. Making professional work legible does not require turning an entire life into public training material. Some information should remain private, and some belongs to other people. A coherent public record can be selective without being deceptive. The question is whether the published account is accurate within the scope it claims.

I also think there is value in accepting imperfect control. We cannot dictate every summary another person writes about us, and we should be cautious about expecting that power over machines. What we can do is make truthful material easier to find, challenge specific errors with evidence and keep our own claims proportionate.

Reputation has always involved a gap between the work and the story told about it. AI gives that gap another set of intermediaries. The durable response is to leave a record that withstands checking, including when the first reader never opens a browser tab.