The Number That Survives
Track a single figure as it moves from its origin toward public conversation, and a mechanism becomes visible.
At origin, the finding is specific: a model achieved a given accuracy, on a given benchmark, under given conditions and limitations. Restated by the party closest to the claim, the conditions soften into a promise: the accuracy transforms into a statement about what the technology does, full stop. Carried further into general coverage, the benchmark drops away entirely, leaving only the comparison that makes for an easier sentence, typically against human performance. Repeated again by voices optimising for attention, the claim becomes a statement about consequence and inevitability, about jobs and disruption rather than about the technology itself. By the time it settles into ordinary conversation, it is no longer a finding at all. It is an imperative: adopt now, or be left behind.
The number itself is unchanged at every stage. What disappears is everything that gave the number meaning: the benchmark, the conditions, the limitations, the methodology. This is the Signal Degradation Curve: information fidelity falls with each retelling, while the confidence with which the claim is stated rises. The two move in opposite directions, and the gap between them is where decisions go wrong.
| Stage | Information fidelity | Representative claim |
|---|---|---|
| Origin | 100% | Model achieved 94% accuracy on benchmark X under conditions Y with limitations Z. |
| Vendor | 80% | Our AI achieves 94% accuracy, transforming how businesses operate. |
| Media | 60% | New AI system outperforms humans with 94% accuracy rate. |
| Amplifier | 40% | AI is now 94% as good as humans. Your job is at risk. |
| Cultural | 20% | AI can do almost anything now. We need to adopt immediately or be left behind. |
Why the Direction Is Never Corrected
The pattern is not random noise. It has a consistent shape, because each stage of retelling optimises for something other than accuracy. A restatement closer to the claim's commercial origin tends to simplify because caveats do not sell. A restatement further from the origin tends to simplify because nuance does not compress into a headline, a slide, or a scroll. At no point in this chain does anyone whose incentive is to be believed have a reason to reintroduce the conditions that were dropped a stage earlier.
That is why degradation runs in one direction only. A claim does not sharpen as it travels. It only ever loses resolution, and the loss compounds, because each stage inherits the already-simplified version rather than the original.
A claim does not sharpen as it travels.
What This Explains
Builder.ai and nate were not isolated frauds so much as visible endpoints of this same mechanism, operating at the extreme. nate's claim of full automation could not have withstood scrutiny of the underlying process: the actual automation rate was zero. Builder.ai's figures were similarly detached from the operational reality they described, with reported revenue running to roughly four times the actual amount, while internal communications worked to keep the extent of human labour behind the product obscured. In both cases, the claim that reached investors and customers had already been stripped of the operational detail that would have made it checkable. The claim that arrived was confident. The claim that would have been useful was the one several stages back, still carrying its conditions.
The claim that arrived was confident. The claim that would have been useful was the one several stages back, still carrying its conditions.
The Signal Degradation Curve is not a prediction about any particular technology. It is a description of what happens to information once it is repeated for reasons other than accuracy, observed consistently enough across the AI market to be treated as structural rather than incidental.
The question the curve leaves open, deliberately, is how an organisation might trace a claim back to its origin before acting on it. That is a separate conversation.