Why it is not a capability failure
The capability is present in each case; what varies is whether the conditions of the task elicit it. A model that introduces citation errors when left to generate freely verifies citations accurately when auditing is required as an explicit step. A model that produces audience-agnostic prose recalibrates its register, structure, and depth when asked to construct a specific audience model before generating. The deficit is a property of default operating conditions, not of the model’s ceiling.
Where it comes from
Two structural properties of generation produce it. Pattern-completion on the training distribution pulls every decision toward the genre average, and competently executed, templatable work is massively overrepresented in what has been digitised, so the average is itself pattern-matched rather than reasoned. Attention attenuation over long sessions means an instruction given at the start is not forgotten but outweighed: fifty thousand tokens later, attention flows to the most proximate content rather than to the audience reasoning that set up the session.
Where it shows up
Most visibly in long, substantive sessions rather than short exchanges. In a two-turn exchange the instruction and the generation are close enough that the instruction carries full weight. In a four-hour session producing a serious deliverable, the instructions established at the start compete with everything accumulated since, and lose. Standard capability benchmarks run under short, controlled conditions and do not surface it, which is why a model at the frontier on benchmarks can produce the full catalogue of failures in deployment.
The human parallel
The class of failure is the characteristic failure mode of junior knowledge-work practitioners in every domain: producing a brief by asking what briefs look like rather than reasoning about what this brief needs to do. Experienced practitioners run an implicit test on every significant decision, whether this element earns its place for this purpose and audience or is here because it is a default. The model has no such accumulated test, and in the absence of architectural provision for it, reverts to the statistical average on every decision.
Key facts
- The intent deficit is the gap between what a language model can produce when directed and what it delivers by default, named by Public Policy Lab in 2026.
- It is caused by pattern-completion toward the training distribution’s genre average combined with attention attenuation over long sessions.
- It appears most in long, substantive work sessions and is not surfaced by short-context capability benchmarks.
- It is structurally the same failure mode as that of an inexperienced human practitioner generating from what outputs usually look like.
Last checked . If a figure here is out of date, tell us.