What is plausible distortion?
Plausible distortion is what happens when an AI output looks correct but has changed the meaning of the source material decisions depend on. The output reads fluently, passes casual review, and reaches the decision-maker with the appearance of accuracy — but the signal has been softened, compressed, omitted, or reframed somewhere between the source and the output.
The output that looked right
Most conversations about AI accuracy are about hallucination — outputs that contain claims with no basis in any source. Hallucinations are the easier problem. They are often obviously wrong, and when they are not obvious, they tend to be checkable. Someone notices. Someone pushes back.
Plausible distortion is harder because it does not look wrong. The output stays close to the source. It uses the right vocabulary, covers the right topics, and reads like a faithful summary. But somewhere in that gap between evidence and output, meaning changed. A condition was dropped. A risk was softened. A qualifier disappeared. A conclusion was drawn that the source never actually reached.
The output was not wrong enough to trigger review. That is exactly when the damage happens.
AI sits between evidence and the person reasoning from it. What survived the output is not always what the source actually contained. Plausible distortion is the gap between the two.
Signal means the decision-relevant meaning in the source: risk, urgency, uncertainty, conditions, weighting, and context. Plausible distortion is what happens when signal changes or disappears between the source and the output.
A concrete example
Here is what plausible distortion looks like in practice. Source material from a vendor contract, and the AI output produced from it:
The vendor is obligated to deliver a remediation plan within 30 days of any material breach, subject to written approval from both parties before implementation.
The vendor is expected to deliver a remediation plan within 30 days of any material breach.
Drifted. "Obligated" softened to "expected." The written approval requirement was dropped entirely. The output reads correctly. The meaning changed.
Nothing in this output is invented. The vendor, the 30-day window, the remediation plan — all present, all accurate. But "obligated" and "expected" do not mean the same thing in a contract. And the approval requirement was decision-relevant. It disappeared with no signal that it was ever there.
The four forms it takes
Plausible distortion is not a single failure mode. It is a category with four distinct forms:
The signal got quieter
Source material is present in the output but its meaning has shifted. A risk is softened. An obligation is reframed as a suggestion. Uncertainty is resolved into a cleaner claim than the source justified.
The source had more
Source units with material signal are absent from the output. Nothing in the output signals the gap. The output looks complete. It is not.
The inference outran the evidence
The output presents a conclusion or recommendation as if the source directly supported it. Synthesis is not the problem. Unmarked synthesis is.
No source trace at all
Claims that cannot be traced to the source. This is the hallucination case — the one traditional detection tools focus on. It is the part they are most likely to catch.
Traditional hallucination detection addresses only the last category. The first three — meaning drift, omission, and unsupported synthesis — are what make plausible distortion the harder problem. They enter decision systems with the appearance of legitimacy.
Why it is harder to catch than hallucination
Hallucination is detectable because the claim has no source. You can ask: where did this come from? And the answer is: nowhere. That is a clear failure.
Plausible distortion is harder because the claim does have a source. The source just said something slightly different. "Obligated" became "expected." A 30-day hard deadline became a general expectation. A conditional approval requirement disappeared. Each of these changes is subtle enough to pass a reader who is not comparing the output directly against the source — which is most readers, most of the time.
That is the problem. Not that the output was wrong. That it was wrong in a way that looked right.
Why better prompts do not solve it
A common response to AI output quality problems is to improve the prompt. Use clearer instructions. Add constraints. Tell the model to preserve exact language. These steps reduce error frequency but do not eliminate plausible distortion.
AI models compress, weight, and reframe source material as a function of how they generate output — not as a function of instruction quality alone. A more carefully prompted output can still omit a key condition, soften a contractual obligation, or drop a qualifier that changes what the claim means. The model is not violating the instruction. It is doing what it does: producing a fluent output that represents its synthesis of the source, not a faithful transcription of it.
The only way to know what survived the output is to verify it against the source after generation.
How Plumb detects it
Plumb receives two artifacts: the bounded source material the AI reasoned over, and the AI output exactly as produced. It normalizes both into comparable semantic units, then verifies the output against the source across four dimensions.
Every claim in the output receives one of four verdicts:
The verification result can be appended to the output before it reaches the decision-maker — so the person acting on the output sees what it was actually entitled to claim.
Is plausible distortion the same as hallucination?
No. Hallucination is when an AI output contains claims that have no basis in any source. Plausible distortion is when the output stays close to the source but changes its meaning — softening a condition, dropping a qualifier, omitting a risk, or reframing a conclusion. Hallucinations are easier to catch because they are obviously wrong. Plausible distortion passes review because it looks right.
Where does plausible distortion show up?
Anywhere an AI output is produced from a bounded source and then acted on. Contract review. Investment memos. Consulting deliverables. Project briefs. Regulatory summaries. Risk assessments. The common thread is that someone reads the output and makes a decision — without going back to the source to check what it actually said.
Can you catch it manually?
Sometimes. If you are comparing a short output against a short source, a careful reader can catch meaning drift. But manual review does not scale, does not apply consistently, and does not catch omissions — because you cannot notice what is missing from an output unless you have already read the source. Most of the time, people read the output, not the source. That is the point of an AI output.
Is this a problem with specific AI models?
No. Plausible distortion is a function of how language models generate output — through compression, synthesis, and weighting of source material. It is not a bug in a specific model. It is a property of the generation process. Different models have different error profiles, but none eliminate meaning drift or omission entirely.
What is Plumb?
Plumb is a faithfulness verification layer for AI-generated outputs. It sits between a bounded source body and an AI-generated output and determines whether the output faithfully reflects what the source actually said. It does not generate, summarize, or score. It verifies. Plumb is built by Astrelle Inc., based in Toronto.
See what your AI outputs actually said
The demo takes 15 minutes. We run Plumb against a real output — yours or one of ours — and show you exactly what it finds.
Book a Demo →