Duty of Evidence for AI Answers

Why hallucinations matter especially in patent work

Patent practice hinges on exact wording: is a feature supported in the description? Is the claim worded as in the application as filed? Did the examiner really cite D1 against feature M3? A model that makes a summary sound plausible does not do this work for you. If anything, it creates beliefs that the attorney later sells as fact.

Typical traps: an invention disclosure that "implies" a feature (the attorney doesn't cite it because there is no reference, although the model claimed the feature was described there); an Art. 123(2) basis that is not in the application as filed; a prior-art quote that is factually right but from a different document than the one the model named.

The rule in one line

This sounds trivial but is the rule broken most often in practice. A capable model readily produces paraphrases, conclusions and interpretations. The prompt has to force the model to deliver a quote plus reference, otherwise you get summaries that are barely traceable back to the source text.

How to enforce it in the prompt

Every evidence-bound prompt needs three components:

1. Explicit instruction: "Back every statement with a verbatim quote in quotation marks and a reference (e.g. para. [0023], claim 3, p. 4 l. 12)."
2. Output format: "Return the result as a table with columns Statement / Quote / Reference." Or as a list: "Statement, followed directly by the quote in quotation marks and the reference in parentheses."
3. Bailout clause: "If you cannot find a verbatim quote for a statement, write 'No reference found' instead of the statement. Do not invent quotes."

The third point is decisive. Without a bailout clause the model prefers a hallucinated quote over a "don't know". With the bailout clause you get an honest gap that is your trigger to check the source manually.

Duty of evidence in typical patent tasks

• Claim construction: every feature construction backed by a quote from the description or a figure (paragraph, figure label). No free interpretation without a textual anchor.
• Feature comparison against prior art: every hit against a feature with a quote from the prior-art document and a paragraph or claim number. A bare "D1 discloses this" without a quote is unusable.
• Art. 123(2) basis check: every amendment with a verbatim quote from the application as filed. Paraphrases do not count as a basis; only the specific reference does.
• Office-action analysis: every objection backed by a quote from the office action itself. Examiners often choose specific wording, and the original phrasing determines the right defence line.
• FTO triage: every relevance decision with a quote from claim 1 of the hit. No abstract-only matching.
• Client report: the duty of evidence applies internally too: the lay-friendly summary must rest on statements that were previously evidenced, not on free rephrasing.

Anti-patterns to avoid

• Quote without reference. "As the specification states, this is a device …": sounds fine, but where exactly? Without a reference the quote is worthless.
• Paraphrase in quotation marks. The model puts quotation marks around its own rephrasing. Spot-check whether the "quote" actually appears verbatim in the source.
• Reference without quote. "See para. [0023]": but what does it say? The statement must come with its quote, not just a pointer.
• Evidence only at the top. Models get "lazy" as prompts grow. The instruction must sit at the end of the prompt or be repeated, otherwise it fades.
• Quote from the chat history. The model quotes from an earlier answer in the same chat and passes it off as a quote from the source. Context management prevents this at the root.

How to verify the duty of evidence

1. Spot check: look up three random quotes from the AI output in the source document. If one is off, treat the whole output with suspicion and recheck.
2. Reviewer pass: a second model or a second prompt checks that every statement has a quote and a reference. See Multi-Model Workflows.
3. Your own notes in parallel: mark the important paragraphs yourself when first reading the source. Later compare the AI quotes against your list. If the AI quotes something that is not in your list, verify.

What AI does to your willingness to check

The three controls above assume something: that you are willing to leave a statement open and go and look it up. That willingness appears to be up for grabs as soon as an AI answer is on the screen. A preprint by Marcoccia, Quattrociocchi and Capraro (arXiv:2607.13562, July 2026) measured this across five experiments with 3,132 participants. On hard questions they declined to answer 36 % of the time without AI advice and only 6 % of the time with it. So they answered more often and were right about a third as often (27.5 % against 9.2 %), while their own rated confidence of being right rose from 29.6 to 75.9 on a 0-to-100 scale. The advice did not even have to be requested: when it appeared unprompted, like an AI summary above the search hits, abstention fell from 35 % to 1 %. Paying for correct answers lowered AI use slightly but did not restore the willingness to abstain.

The limits belong with it: the participants were recruited via Prolific in the US and tested on a single task type, visual details in films, and the AI advice was constructed to be wrong almost every time. The paper expressly says nothing about how accurate AI assistants are in general, and whether the effect holds when AI advice is mostly correct is untested. It is a preprint.

A reading, not a finding: if the effect also shows up among professionals in their own field, then checking cannot rest on your own attentiveness. The stop clause in the prompt and the spot check at the end become fixed steps that stand in for a willingness that fades under AI. And "I am sure" is a worse quality signal after using AI, because in the study rising confidence coincided with falling accuracy.

What the duty of evidence costs

If every statement has to carry a quote and a reference, the honest question is: where is the time saving if it all gets checked in the end? Verifying is work, and that work is the price of being able to trust a machine output at all; some call it the "truth tax", the levy on every AI statement. Skipping it saves time only on paper: an unchecked statement that the attorney later passes off as fact surfaces in opposition or infringement proceedings, and there it costs many times more.

In practice the maths usually still comes out positive. Reading an existing quote against the source is faster than finding the passage from scratch, and the honest gap from the bailout clause shows exactly where you still have to look. You can lower the tax by handing part of the checking back to the AI: have the model point to the exact passage, or let a second model run the reviewer pass. The balance differs by task: for hit screening and basis checks the duty of evidence quickly turns into a net gain; for freer writing the checking burden dominates at first. Better to weigh that consciously than to wish it away.

Connection to the other pillars

The duty of evidence is one of the four pillars for prompt engineering in patent practice. It only works in combination with the other three:

• Context management: a fresh chat with exactly the relevant documents. Without clean context, the model quotes from the chat history instead of the source.
• Paragraph-by-paragraph review: the model goes through the entire source, not just the prominent places. That way no evidence is left behind.
• Self-reflection + devil's advocate: at the end the model (or a second model) checks that every statement really has its evidence.