| The thesis: a machine does not decide what is true about your company by checking whether it is correct. It decides by consensus, by what the largest, most consistent body of sources agrees on. That means a correct claim standing alone loses to a wrong one repeated everywhere. In the Interpreter Era, accuracy is no longer a verdict. It is a vote, and you can lose it while being right. |
We inherited a comforting idea about truth: that it is a property of statements, independent of how many people hold it. A thing is true or it is not, regardless of the vote. That idea still governs how leaders think about their own companies. We know what we are, the reasoning goes, so if we simply state it clearly, the truth will out. It is a reasonable belief, and against a machine it is dangerously incomplete.
Because a model has no access to what is actually true about you. It has access only to what is said about you, at scale, and a mechanism for resolving the disagreements. That mechanism is not a truth detector. It is closer to a poll.
How a machine adjudicates what you are
Consider what the model is actually doing when it answers a question about your company. It is not consulting a verified record of fact. It is drawing on a vast field of statements, some from you, most not, some current, many stale, and producing the answer that is most consistent with the weight of that field. When the statements agree, it reports the agreement with confidence. When they conflict, it does not investigate which is correct. It leans toward the version that is most represented, most repeated, most corroborated across the sources it trusts.
That is a consensus procedure, not a verification procedure. It answers the question what do the sources agree on, not the question what is actually the case. Most of the time those two questions have the same answer, which is exactly why the difference is easy to miss. The trouble arrives in the cases where they diverge, and for a company whose reality has changed faster than its record, they diverge constantly.
Correct but lonely loses to wrong but repeated
Here is the hard consequence, stated without softening. A claim about you that is accurate but thinly represented, said by you, on your site, once, will lose to a claim that is inaccurate but widely repeated across many sources. The machine is not being perverse. It is doing exactly what it was built to do: weight by consensus. But the effect is that truth and volume have been quietly swapped. The winning claim is the popular one, and popularity is not correlated with accuracy in the cases you care about most.
This is why the instinct to simply state your truth clearly, once, fails. A single correct source is a minority vote. It does not matter how right it is if it is outnumbered. We touched the transmission side of this in Your Brand Is a Rumor Again, but the deeper point is epistemic: the machine has no category for a lonely truth. To its procedure, unrepeated and untrue look nearly identical.
Why being right stopped being enough
For most of business history, being right eventually won, because reality had ways of asserting itself. The product worked or it did not. The customer was served or was not. Truth had a home-field advantage, because it was tethered to outcomes people experienced directly.
The consensus machine severs that tether at the moment of first judgment. Before anyone experiences your product or talks to your team, the machine has already returned its verdict based on the vote of the sources, and reality has not yet had a chance to weigh in. Being right still matters enormously downstream. But it no longer protects you upstream, where the machine forms the impression that decides whether downstream ever happens. Right, alone and unrepeated, is not a strategy anymore. It is a starting position that can lose.
What actually builds consensus
If truth is adjudicated by consensus, then the work is to build the consensus honestly, to make the accurate version of you the most represented and best corroborated one in the field the machine reads. Consensus is not one thing; the machine assembles it from several signals.
- Volume of agreement. How many sources say the accurate thing, not just whether you say it. A truth repeated across many independent places reads as settled.
- Consistency. Whether those sources say it the same way. Ten phrasings of your truth look like disagreement; one phrasing repeated looks like fact.
- Corroboration. Whether independent, credible voices affirm it, not only you. Outside agreement counts for far more than self-assertion in a consensus system.
- Authority of the agreeing sources. A consensus among high-trust sources outweighs a louder one among low-trust ones.
Those are, not by accident, the things the ACES lens measures and the discipline of Model Response Optimization works to move. Winning the vote is not a dark art. It is making the true account of you the one the sources actually agree on.
The line that matters
It would be easy to read all this as license to manufacture agreement, to flood the field with your own repeated claims and call it consensus. Resist that, and not only for ethics, though the ethics are real. A manufactured consensus is fragile. It rests on a narrow, self-referential base that the next model, or the next investigation, can see through, and when it collapses it takes your credibility with it. Real corroboration, independent sources that agree because the thing is true, is what holds under scrutiny. The goal is not to win the popularity contest by stuffing the ballot. It is to earn a genuine consensus around what is actually so, so that the popular answer and the correct one are the same answer. That is the only version of this that survives contact with a system whose entire job is to detect what many sources independently confirm.
Questions this raises
How does AI decide what's true about my company?
Not by verifying facts, but by consensus. It draws on the field of what is said about you across many sources and returns the answer most consistent with the weight of that field. When sources conflict, it leans toward the most represented and corroborated version rather than investigating which is correct. It answers what the sources agree on, not what is actually the case.
Why does AI repeat inaccurate things about us even though we've stated the truth?
Because a single accurate source is a minority vote. If an inaccurate claim is repeated across many sources and your correction appears in only one place, the consensus procedure weights the popular claim over the lonely true one. To the machine, unrepeated and untrue look almost the same.
Isn't being right and stating it clearly enough?
It is necessary but no longer sufficient. Being right still wins downstream, once people experience your product or talk to your team. But the machine forms its verdict upstream, before any of that, based on the vote of the sources. A correct claim that is thinly represented can lose that early vote, which determines whether downstream ever happens.
How do we build consensus without faking it?
By earning genuine agreement: making the accurate account of you widely and consistently represented, and corroborated by independent, credible sources rather than only asserted by you. Manufactured consensus rests on a narrow self-referential base that collapses under scrutiny; real corroboration holds because the thing is actually true.
Is this the same as your rumor or trust arguments?
It is the layer beneath them. The rumor piece was about lost provenance; the trust piece about who now dispenses trust. This is about the procedure itself: how machine truth gets adjudicated, by consensus rather than verification, which is why volume and agreement can override accuracy.
The bottom line
We would like to believe that truth defends itself, that being right is a fortress. Against a consensus machine it is not a fortress, it is a single vote, and a single vote can be outnumbered by a confident, repeated error. This is not a reason to abandon truth for volume. It is a reason to stop assuming truth travels on its own and to start doing the work of building an honest consensus around it, so that the version of you the sources agree on and the version that is actually true are finally the same thing. In the Interpreter Era, being right is where you start. Being agreed with is how you win.
See what the sources currently agree on about you. Request your Brand Interpretation Assessment, or read the Interpreter Architecture whitepaper for the full thesis.