Episode 10 Liz and Sumair sit down with Saskia van Ryneveld in the wake of the USPTO's first AI-predicated decision, a case that turned on the fact that no one could reconstruct what the drafting AI had actually seen or been told. If a generic chatbot can land you in front of an examiner with no provenance, what should patent professionals actually be using? Liz walks through her real drafting workflow with inventors and small companies, where specialist AI fits, where the attorney still owns the work, and why the "closed universe" of a prosecution file is the opposite of what a general-purpose LLM is built for. The through line: patent drafting isn't autocomplete, and the next twelve months of USPTO scrutiny will reward tools that log every input, every prompt, and every edit.
The USPTO has issued its first AI-predicated discipline order, and the facts are almost mundane. A patent attorney registered since 2002 drafted a claim construction with one generative AI tool and reviewed it with a different one. The citations pointed to parts of the specification that did not exist. He caught the mistake roughly a day later and fixed it. The court did not sanction him. The Office of Enrollment and Discipline publicly reprimanded him anyway.
Elizabeth Shipsides, a patent attorney at HSML Intellectual Property Law who has worked in-house, at firms, and as a USPTO examiner, sees the order as a warning about process more than about one person's carelessness. Speaking with Saskia van Reineveldt and Sumer, who heads European business at IP Author, she laid out why the case matters, and what it suggests about how patent professionals should and should not use AI.
The reprimand came from the OED rather than the court because registered practitioners carry a duty of reasonable inquiry under 37 CFR 11.18. Signing a paper certifies that it is accurate and not misleading. That duty applies even when the underlying case sits in district court, and it does not transfer to software. Sumer put it bluntly: you cannot make a mistake because your AI made one. The same applies to the fiduciary duty owed to clients, not just to the Office.
Nothing in the order suggests intent. Shipsides reads it as careless rather than dishonest, and she is still astonished that anyone would submit a draft written by one AI and checked by another without reading it personally. The order does not name the tools, only that generative AI was used.
The most plausible explanation for the failure is inconsistency between the tools. If the reviewing tool never received the specification, it had nothing to check paragraph numbers against, and AI is built to produce fluent, confident answers from whatever it has been given. Shipsides suspects the second tool simply lacked the same inputs as the first.
Sumer drew a parallel to European practice. Under Article 123(2) EPC, every amendment has to be supported by the application as originally filed. A general model that has not been given the priority document or the application as filed cannot know it is working inside that closed universe. A tool built for prosecution can have that constraint engineered in. A general chat window cannot.
Shipsides's practical conclusions:
The time spent putting documents into two different tools is time that isn't meaningful legal work.
Sumer's argument for vertical tools centers on memory. A general model accumulates context from everything you have asked it, and earlier instructions get diluted as new material piles on. Even prompt order matters: put the task first, then a very long context, and the answer differs from the one you get when you restate the task after the context. A tool where each project starts clean keeps the context for one matter consistent, with no leakage from previous work.
Shipsides's experience fits. Her drafts felt more personalized to her in April and May, picking up her tone and voice, and less so now. She wonders whether turning off training on her data is the cause. Her colleagues ran the same tool and produced drafts she liked less than her own, so some personalization survives. She would still expect a dedicated tool to hold her style better than a general one does.
The same inconsistency shows up on the client side. Roughly half a dozen of her inventors have run her drafted applications through their own AI tools and returned text with misnumbered figures and paragraphs, redundant material, or a structure that no longer matches the original. Sometimes the tool had not read the document at all. In Claude, she has seen a .docx uploaded to a project file get converted to a text file, and the model admitted it had not read it when she asked. Her fix is to ask directly whether the document was read. Her wish is for attorney and inventor to share one tool.
Sumer described a workflow for startups that already have specifications drafted with AI. He sends the inventors his own version of the specification before drafting claims, so both sides are talking about the same embodiments and the scope is locked early. An AI chatbot working from the invention disclosure can produce that specification quickly.
Shipsides draws a firm line: claims and drawings are hers. She will let AI propose descriptions of diagrams for the detailed description, and she still changes them most of the time. The background section is a matter of preference, and she often cuts what AI writes. Everything gets reviewed. On slow days, when the tool lags, doing it herself is faster.
Where she finds AI most useful is brainstorming and strategy for office action responses. She reads the spec, the office action, and the cited art while the tool does the same, then compares what each of them found. Matching the tool's findings is validating: a second pair of eyes agreeing. In one run against Claude (she recalls Opus 5), she finished faster than the model and landed on the same points.
AI is good at spotting weaknesses. It doesn't necessarily know how to execute on them.
The limits show up in strategy. Claude tends to be argument-prone, she said. In one case, where the cited reference described different embodiments, it advised simply arguing. Shipsides, a former examiner, would rather amend and advance prosecution, accepting a 103 rejection next round instead of risking a second non-final. Examiners are human, and prosecution is a negotiation in which you give the examiner something. Some weaknesses the tool flags are ones she knows no examiner would pursue. The model cannot read that context.
Sumer added that models differ in temperament: some agree with everything, some push back, some lead with the worst case. Shipsides's tool argues with her, which she suspects it learned from her. She finds that overriding it also builds confidence, since she has evaluated a proposal and said no.
Her verdict is split. AI raises the quality of the work. She is not sure it saves time, partly because she argues with it. For claim charts she likes IP Author's color coding of elements as fully disclosed, partially disclosed, or weak, which is fast, though the decision on how strong a weakness is remains hers. For verifying that amendments have support in the application, she has found AI helpful on the prosecution side.
Claude finds details well, but ChatGPT writes warmer client emails. One told her to say the files needed closing out.
The order sets a baseline without banning anything: AI can draft and review, and the practitioner still owns every citation in what gets signed. The sensible response is structural. Load the same documents into one environment, let the tool keep the record, and check each paragraph reference yourself before signing. The attorney in this case caught his error within a day. The reprimand suggests that catching it before filing is the standard.
