The Same Question, A Different Company: Why Public AI Systems Don't Agree About You

2026-06-06

The Same Question, A Different Company

Ask ChatGPT whether a company is financially sound. Then ask Gemini. Then Claude, Grok, and Perplexity. You will not get one answer with minor wording changes. You will often get materially different accounts — different facts, different emphasis, different conclusions about the same organisation, drawn at the same moment.

This is not a glitch. It is the defining feature of the current landscape. A 2026 review of the leading public AI systems found that no single one wins across every category; they have specialised, and they disagree. On one independent accuracy test, the worst-performing system fabricated information roughly 20 per cent of the time — while another, built to decline when uncertain, scored close to zero. That is not a rounding difference. One invents; the other stays silent. Both are answering questions about your organisation right now.

For most leadership teams, this is invisible. They check what "the AI" says by opening one system, reading one answer, and assuming it stands for the rest. It does not. There is a spread — and every stakeholder lands somewhere different inside it.

The audience is already fragmented

Consider who is asking. An investor doing first-pass research may favour one system. A regulator's staff may use another, procured and approved for government work. A journalist on deadline reaches for whatever is fastest. A prospective hire asks the system their friends use. Each receives a version of your organisation — and none of them is cross-checking against the others.

So there is no single account to address. The clean verdict an analyst reads on one system may directly contradict the cautious, heavily caveated answer a regulator reads on another. Both feel authoritative. Both arrive in the confident, fluent tone these systems share whether or not the underlying claim is true. Neither reader has any reason to suspect the version they received is one of several.

Divergence is a signal, not just noise

There is a useful pattern hidden in this. When independently-built systems converge on the same answer, that agreement means something — it suggests the claim is well-supported across the public record they draw from. When they diverge in ways that change the conclusion, the divergence marks contested ground: a place where the public information about your organisation is thin, stale, contradictory, or simply wrong on one side.

Those are exactly the points worth knowing before someone else finds them. A split between systems over whether your organisation settled a particular matter, or who currently leads it, or what a discontinued product actually did, is an early sign that the record is unsettled — and that a wrong version is in circulation somewhere.

Companies that learn this late tend to learn it from the outside. The recent defamation claims over fabricated AI statements — a solar firm wrongly told the public it faced an attorney-general action, an individual described with invented criminal records — did not start with the company noticing. They started with someone else seeing the answer first.

What this asks of leadership

The instinct is to want one clean answer you can point to and, if needed, correct. That instinct no longer matches reality. There is no single answer, because there is no single system shaping perception. There are several, they disagree, and the disagreement shifts each time one is updated.

So the question is not "what does AI say about us?" It is "where do the systems disagree about us, and which version is reaching which stakeholder?" Answering it means querying each of them, on the questions that matter, and reading the spread rather than a single reply.

Clarity here does not come from finding the answer. It comes from seeing all of them at once — and knowing, before an investor, a regulator, or a reporter does, where the accounts of your organisation quietly come apart.