Series · What the science says
AIs silently distort documents — why "a generalist assistant + a PDF" is not enough
Manent — July 15, 2026
"Take a generalist AI assistant, attach your documentation, ask your questions."
That's the reflex of many companies. Research has just shown
why it's risky — and it's exactly the risk Manent
is designed against.
What research says
In 2026, a team at Microsoft Research published
DELEGATE-52, an evaluation of 19 language models — including
the best on the market — on tasks where they are delegated work on
a document over the course of a long conversation. The finding is striking:
the models silently distort the content. Not
visible mistakes — a number that drifts, a date that migrates, two names that
get swapped, a value that shifts by a few units — while
the text stays impeccable and convincing.
The scale: with the best models, on the order of 25 % of the
content altered over about twenty exchanges, with sharp
drops concentrated on a few turns. And a counterintuitive result:
giving the model tools makes things worse — the more it
manipulates, the more it drifts.
Honesty about scope: DELEGATE-52 measures
the editing of documents, not simple question-and-answer. So we do not
conclude "a generalist assistant gets 25 % of your answers wrong". What
matters is the mechanism: an AI that works from
memory, without ever verifying each claim against the source, drifts in a
plausible and invisible way — and this mechanism threatens any use of a model
on your documents.
Why "a generalist assistant + an attached PDF" inherits the problem
The naive setup chains together exactly the conditions the study penalizes:
- It paraphrases from memory instead of citing — right where
the values drift.
- Nothing verifies the answer against the source: it comes out
as is, plausible, without a second look.
- It doesn't know how to stay silent: it answers even when the
document doesn't cover the question.
- The citation, if there is one, is not verifiable —
an approximate page number, sometimes invented.
- Critical content is regenerated like the rest: a
reworded safety instruction is a potentially false instruction.
What we do at Manent
Each guardrail answers a recommendation from research — verify,
restrict, keep the human in the loop:
- Grounding verification. Every claim in an answer
is checked against its sources. Unsupported → it is not shown.
- Exact citation. File, section, page (or minute of a
video). The drift becomes visible: the technician opens the page
and verifies.
- The right to stay silent. Without a reliable source, the assistant says
so and refers to an expert, instead of inventing to fill the gap.
- Safety cited word for word, never reworded — and
verified with a reinforced model when a safety instruction is at
stake.
- Read-only. The assistant never modifies your
documents. New knowledge becomes a draft, published
only after validation by a human on your team.
- Your data stays with you, in an isolated workspace.
Side by side
| | Generalist assistant + PDF | Manent |
|---|
| Verifies each claim? | no | yes, against the source |
| Cites a verifiable source? | rarely | file + page, to open |
| Knows how to stay silent? | no | yes — it abstains |
| Safety instruction | reworded | cited word for word |
| New knowledge | no trace | draft → human validation |
| Your data | shared servers | isolated workspace |
The difference, in one sentence: a generalist assistant
handed a PDF answers "from memory, according to your document"; Manent answers
"according to page 42 of this file — and if I have no source, I don't
answer". In the field, a wrong torque value recited with
confidence costs more than an "I'm not sure, here's the page".
What we don't claim: no tool is
magically immune. The right proof is not a marketing number, it's to
measure citation accuracy on YOUR corpus — which the
audit report and the measurement bench do, on your documents.
References
- Laban, P., Schnabel, T., Neville, J. (2026). LLMs Corrupt Your
Documents When You Delegate (DELEGATE-52). Microsoft Research.
arXiv preprint 2604.15597 ; code: microsoft/DELEGATE52.
- Independent coverage: The Register, Windows News (May 2026).
Measure reliability on my corpus