How Source-Backed AI Reduces Risk in IP Work

• An AI statement about a disclosure, a reference or an office action is only usable if the practitioner can see the passage it rests on. The session names two failure modes: misreading what a prior art paragraph actually says, and claimed subject matter that the specification does not support.
• The same question about an invention disclosure, asked twice: the plain chat answer is well structured but points to nothing; with source search on, each statement links to the section of the disclosure it comes from, so it can be checked before anyone relies on it.
• When comparing a disclosure with a cited reference, each difference comes with the highlighted passage in the reference, which removes most of the back-and-forth between PDFs.
• Limits were shown openly: at the time of the session (February 2026), a German-language reference returned a translated excerpt but no jump to the passage in the original document.
• In drafting, two checks: whether the subject matter of an independent claim appears in the summary, and whether the background contains wording that could be used against the application, such as "the present invention offers a solution to the limitations of the prior art".
• In an office action response, two checks: locating the passage of the cited art the examiner relies on, and confirming that a proposed amendment (here, adding a vacuum chamber to claim 13) has a basis in the application as filed, to avoid a new matter or added-matter objection.

Thomas Chazot, Head of Growth, DeepIP (host)

Romain Vidal, DeepIP (at the time of the session)