What is AI drift?
AI drift is what happens when an AI output reflects the source but not faithfully. The text holds together, the structure looks reviewed, the meaning is just slightly off. By the time anyone notices, the brief has gone out, the recommendation has been made, the decision has been signed off on what the AI produced rather than what the source said.
This is not model drift, data drift, or concept drift. Those describe a model changing over time, or the world shifting around it. AI drift here means something narrower and more immediate: a single output moving away from the specific source material that went into it.
Drift and unsupported content are different questions
Hallucination is one way AI output can be wrong: the output contains claims with no basis in any source. Another is when the output remains supported by the source but changes an important part of what the source meant. A condition can disappear, certainty can increase, scope can widen or relevant source information can be omitted.
Unsupported content asks whether the claim has source support. Drift asks whether the source's meaning survived the transformation.
Drift is harder because it does not look wrong. The output stays close to the source. It uses the right vocabulary. It covers the right topics. It reads like a faithful reflection of the source. But somewhere between the source and the output, meaning shifted. 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.
Language models can transform source meaning while summarizing, synthesizing and generating from source material. Better prompting, retrieval and model capability can reduce these failures, but they do not guarantee preservation of every condition, qualifier, exception or relationship.
What drift looks like
Here is drift 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.
Changed. "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 are all present and 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 forms drift takes
Drift is not a single failure mode. It shows up in distinct, separable forms, and Plumb traces movement across the dimensions that carry meaning: conditions, exceptions, certainty, scope, obligations, permissions, timing, numbers, entities, relationships, causality, and emphasis. Some of the ways meaning moves:
The meaning shifted
The claim reflects source material but with changed meaning. Softened, strengthened, redirected. The vocabulary stayed close. The signal moved.
The signal went missing
Source material with material weight is absent from the output. A risk, a condition, a constraint, a flag. Present in the source. Gone from the deliverable.
The claim has no source
The output contains content that cannot be traced to the source at all. This is the classical hallucination case, and the one that existing grounding and hallucination checks most directly address.
Why drift matters now
AI is moving from experiment to production in firms that produce decision-bearing work. Consulting briefs. Legal memos. Risk assessments. Vendor evaluations. Client-facing recommendations. Investment notes. The outputs of AI workflows are being acted on by partners, clients, and stakeholders, often without the source being checked.
When those outputs drift, the decision rests on what the AI produced, not on what the source said. The reputational, contractual, and fiduciary weight sits on the deliverable. The integrity sits in the source. Drift is the gap between the two.
AI is being deployed into formal decision systems without the infrastructure that keeps its outputs inspectable. That is the layer Plumb builds.
Drift and plausible distortion
The dangerous form of drift, where the output passes casual review while carrying changed meaning, is what we call plausible distortion. Drift is the broader phenomenon. Plausible distortion is the specific class of drift that reaches a decision without being noticed.
Not every form of drift is plausible distortion. An output that drifts and reads obviously off is just a bad output. Someone catches it. Plausible distortion is drift that looks right. That is the case Plumb is built for.
How Plumb Analysis addresses drift
Plumb Analysis reads AI-generated output against the bounded information behind it and surfaces where meaning changed, disappeared or was introduced. Two things go in: the material the AI worked from, and the AI output exactly as produced. What comes back is a structural account of what the output actually represents.
Findings are owned by specific source material rather than reported as a score. The claim support question, whether the source backs what the output said, is one part of it. The other is whether the source's meaning survived: why a supported claim can still get the source wrong.
Drift gets caught at the layer between AI and decision. Not as a quality score. Not as a badge. As infrastructure.
Is AI drift the same as model drift?
No. Model drift refers to a model's performance degrading over time as the world or the data changes. AI drift, as Plumb uses the term, refers to drift between an AI output and the source material it was produced from in a single interaction. The two are different problems. Plumb addresses the second.
Can better prompts prevent AI drift?
Better prompting can reduce these failures, but a strongly prompted output can still omit a condition, change certainty or lose a relevant qualification. Checking the output directly against the bounded source makes those changes inspectable.
Where does AI drift show up most?
Anywhere AI is turning source material into output someone acts on. Summaries, vendor evaluations, contract reviews, briefs from project notes, due diligence memos, client-facing recommendations. The common pattern is a source set, an AI output produced from it, and a human acting on the output without re-reading the source.
Why is drift difficult to catch?
Drift is difficult to notice because the output can remain closely connected to the source. The vocabulary is right. The structure is right. The problem is not absence of support; it is that important meaning moved while the output remained credible.
How does Plumb Analysis fit into an existing AI pipeline?
Plumb Analysis runs between the AI and the human. The source material and the generated output are passed to it before the output reaches a person. It returns a structural account of the output: what was supported, what changed, what was omitted, what was introduced. One integration. One contract. Every workflow.
See drift in your own work.
15 minutes. We run Plumb against a real output, yours or one of ours, and show you exactly what it finds.
Talk to Us →The products
Plumb reads what AI produced against the information behind it, and what AI did against what actually became true.
Plumb Analysis for what AI produces · Plumb Boundary for what AI does