Register inheritance
The output carries the emotional register of the immediate session context rather than the register the external audience requires. A proposal drafted after a frank internal conversation about a competitor reproduces that conversation’s framing in language that will reach the funder. The model did not reason about the audience shift; the most proximate context saturated generation and carried through.
Over-literal instruction-latching
The model treats the nearest-stated instruction as the operative goal rather than as an input to reasoning about what would serve the underlying purpose. Told to be thorough and write in prose, it buries the decisions and recommendations several pages in, so a senior reader with minutes rather than hours cannot act on the document without reading it in full. The instruction was followed; the goal was abandoned.
Evidence substitution by evaluative language
A finding is asserted and the work of convincing the reader of its significance is done by adjectives such as striking, compelling, or transformative, rather than by the numbers, the scope, or the concrete consequence. The adjectives arrive in place of the evidence that would earn them, performing confidence rather than conveying it.
Plausible citation
A claim is followed by a citation to a real author on an adjacent topic whose actual work does not support the specific claim at the scope asserted. The most dangerous version is the real reference that does not say what the text claims: it survives any check short of fetching the source and reading the passage, and is far harder to detect than a fabricated reference that a quick search resolves.
Audience-unaware design
A tool is designed without a specific user in view and defaults to an abstract user who already knows the things the tool is supposed to teach. A pedagogical visualisation for children renders bars of equal width regardless of magnitude, breaking the point it exists to make. Each instance is a failure of the audience model.
Content loss during format conversion
Asked to convert a document to another format, the model treats the task as regeneration rather than transformation. What returns has the requested format and much of the original content silently rewritten, restructured, or abbreviated, with no flag raised. It is hard to catch because a plausibly formatted file does not invite a check of whether the original survived.
Key facts
- Six named failure modes recur in AI-assisted knowledge work: register inheritance, over-literal instruction-latching, evidence substitution by evaluative language, plausible citation, audience-unaware design, and content loss during format conversion.
- None is a capability failure; each would have been caught by a reasoning step that was not run.
- The most dangerous citation failure is a real reference that does not support the claim, not a fabricated one.
Last checked . If a figure here is out of date, tell us.