suggestibility.ai
Suggestibility is a failure mode. We named the company after it.
The word already meant something before it meant us. It describes how readily someone accepts what they are told — and language models have the trait badly. That is the problem this product exists to work around, which is why it is on the door.
The word has a meaning already
In psychology, suggestibility is the degree to which a person accepts and acts on information supplied by someone else. It is not a diagnosis, and it is not a polite word for gullible. It is an ordinary trait that everyone has to some degree, and it is measured rather than assumed.
How readily a person yields to a leading question, and how much they shift an answer after being told they got it wrong. This is the thread that runs through research on false confessions: the pressure of the question shapes the answer, and the person giving it often cannot tell that it did.
Details introduced after an event get absorbed into the memory of the event. Ask a witness how fast the cars were going when they smashed rather than hit, and the remembered speed changes. The memory is not being consulted so much as reconstructed, with the question as an ingredient.
Suggestibility is not a psychiatric condition and not a defect of intelligence. Highly capable people are suggestible under the right framing, which is exactly what makes it interesting — and exactly why a system that is confident and fluent can still be suggestible without appearing to be.
Machines have it too
Language models display a recognizable version of the same trait. It usually goes by sycophancy, and it shows up in behavior anyone who has used a chat model will recognize:
- They move toward a position the user has already stated, and away from one the user has already dismissed.
- They read the framing of a question as evidence about what answer is wanted — a question that assumes a flaw finds a flaw.
- They soften criticism when the person asking appears to be the author of the thing being criticized.
- They converge on an answer that is already on the table, rather than arriving at one separately.
None of this is a bug that a better model release quietly removes. A model trained to be helpful to the person in front of it is being trained, among other things, toward agreement. The trait and the usefulness come from the same place.
Why that breaks single-reviewer review
In a review, you are the one supplying the frame. You paste your own architecture decision, your own policy, your own threat model, and you ask whether it is sound. In doing so you have told the reviewer who wrote it and what you are hoping to hear.
Ask one model whether your design is good, and a large part of what you learn is how you asked.
The deeper problem is that agreement produced this way carries no information. If a reviewer would have agreed with the opposite document too, its approval of this one tells you nothing. You cannot detect that from the answer, because a suggestible reviewer sounds exactly like a convinced one. Confidence is not the tell.
Asking the same model a second time does not help, and asking it to critique its own answer helps less than it appears to — self-critique inherits the same priors in the same context window.
What independence has to mean
The countermeasure is not a better prompt. It is a procedure that makes convergence impossible rather than discouraged.
Every reviewer on a board is invoked concurrently, and each one receives only the artifact. No reviewer is shown another reviewer's findings, because a reviewer that can see the others is no longer an independent reviewer. The findings meet each other exactly once, at the synthesis step, after every reviewer has already committed to a position it cannot revise.
That ordering is the whole design. It is why consensus here means something narrow and checkable — several reviewers arrived at the same finding separately — rather than the much weaker thing it usually means, which is that nobody objected once the first answer was on the table.
It is also why every finding keeps the name of the reviewer that raised it. An attributed finding can be weighed. A blended one has to be taken on faith.
The minority report
In Philip K. Dick's story, the minority report is the dissenting prediction — the one vision that disagreed — and the system's authority depends on it never being read. The majority is cleaner without it. It is also, in the story, the one that happens to be right.
Averaging model outputs does the same thing for the same reason: the blended answer looks more authoritative precisely because the disagreement has been removed from view. So dissent is kept here as a first-class part of the output, recorded and attributed, not reconciled into the majority and not quietly dropped for reading smoothly.
A single reviewer objecting to the retry semantics in your design is not noise to be filtered. It is the most interesting sentence in the review, because it is the one place the question turned out to be genuinely contested.
What we do not claim
Reviewers drawn from different model families are not truly independent. They share large portions of their training data and a great deal of convention about how a helpful assistant behaves. Their mistakes are correlated. Separate families fail differently enough to be worth convening, and they are plainly better than asking one model twice — but they are not several unrelated human experts, and consensus among them is weaker evidence than consensus among people who trained in different decades under different assumptions.
What survives that caveat is still useful, and it is worth being precise about what it is. When reviewers with overlapping priors still split on a question, the split is rarely about the models. It usually means the artifact itself is underdetermined — it does not say enough to settle the point. That is a real signal about your document, available from correlated reviewers, and it is the one a single reviewer can never give you.
So: the name
Naming the company after the flaw is not modesty. It is the shortest way to say what the product is for. Suggestibility is the thing that makes a single AI reviewer feel authoritative and be worth very little. Everything here — parallel invocation, no shared context between reviewers, attribution on every finding, dissent preserved instead of averaged — exists to take it away from the answer.
How this compares to a human review board · Who sits on the board · What the score means