There are two very different ways an AI answer can be wrong, and they need opposite fixes. The first is the echo chamber: a claim that sounds well-supported because it has been copied a thousand times. The second is genuine disagreement: a question where competent, honest experts land in different places. Treat them the same and you will either trust noise or paper over a real fault line.
The echo-chamber problem
The open web rewards repetition, not accuracy. A single confident blog post gets summarised, re-summarised, and cited until the original — often thin or wrong — is buried under its own reflections. Models trained on that corpus absorb the consensus as if it were evidence. Three failure modes compound it:
- Citation laundering — a guess in one article becomes a “fact” the moment a second article links to it, with the uncertainty stripped off in transit.
- SEO sludge — pages optimised to rank, not to be right, dominate what a model and a search tool both see first.
- Model collapse — as AI-written text floods the web and trains the next model, the system increasingly learns from its own echoes rather than from the world.
The tell is always the same: volume masquerading as evidence. Ten sources that all trace back to one unverified origin are one source wearing a crowd costume. A system that counts them as ten is not being rigorous — it is being fooled.
When experts genuinely disagree
The opposite trap is treating every question as if it has one settled answer. Plenty do not. What a drug’s real-world efficacy is, whether a market is consolidating or fragmenting, how a regulation will be enforced — on these, qualified people disagree, and the disagreement is the information. Averaging two expert views into a confident middle is not synthesis; it is the destruction of the most useful signal in the room.
Where the web’s problem is false consensus, the expert’s gift is honest dissent. A good system amplifies the second to immunise you against the first.
How we decide what to trust
Validation, for us, is not a confidence score bolted on at the end. It is four habits built into how an answer is assembled:
- Trace to provenance. Collapse every claim back to its primary source — the filing, the transcript, the dataset — not the article that cited the article that cited it.
- Weight source diversity. Independent corroboration counts; repetition of a single origin does not. Ten echoes of one claim score as one.
- Surface dissent, don’t average it. When credible sources or experts disagree, show the split and the reasoning on each side rather than manufacturing a false middle.
- Route the contested core to a human. When the question turns on lived experience the documents cannot settle, the model stops and asks someone who actually knows.
This is the same trust model we apply everywhere: state what can be verified against a real source, label what is merely inferred, and flag what is genuinely unknown or contested so a person can weigh it. The goal is not an answer that sounds certain. It is an answer you can audit — and, where the world honestly disagrees, one that tells you so.