From Paper to Patent: What to Cut, and Using a Graduate AI Assistant for Preliminary Search
For the same work, a paper and a patent require different content; this is technology transfer beyond the paper. I was reminded by my advisor only after submission that "this could be patented," and then realized I had done things in the wrong order and worried for two weeks for nothing. Here is what I learned from that experience.
Five dimensions of difference:
1. Writing goal. A paper proves "what I discovered and why"; a patent states "what scope I claim, and whether others doing the same thing would fall within it." In one sentence: a paper needs others to believe it; a patent needs others to be unable to work around it.
2. Timing of disclosure. Conference abstracts, preprints, defenses, and broad public disclosure can all affect novelty. My approach: if I intend to file, I first report to my advisor and the university's intellectual property office, then decide when to disclose. This is a pitfall I fell into—I reversed the order of reporting and submission, and could only remedy it afterward.
3. Subject of protection. The implementation details that "seem redundant" in a paper are often exactly what the patent needs: temperature ranges, ratio ranges, step order, and alternatives. A paper describes the mechanism; a patent describes steps that can be followed.
4. Use of data. A "performance improvement" figure in a paper must become implementable experimental data in a patent (examples + comparative examples). Writing only "the effect is good" will not hold up during examination.
5. Language. A paper can use "may/tends to," while claims must define clear boundaries; any alternatives you can think of should either be written into dependent claims or left as a sentence in the specification.
Outline of a technical disclosure (what I give to the patent agent):
· Technical field, in one sentence;
· Problems with prior art: who has done something similar, and where it falls short (this section needs support from literature search. I open MuchSci's paper search assistant for a preliminary check: Crossref's 180 million records, exact or fuzzy search by title, author, and DOI, filtering by year and citation count, with DOI and citation information automatically completed. Note that it searches papers; patent databases must be searched separately);
· My solution: steps + parameter ranges + why key parameters take this interval;
· Examples and effect data;
· Differences from prior art;
· Alternatives (material substitution, order adjustment, parameter extrapolation). The claims are finalized by the patent agent, but how detailed the disclosure is directly determines how broad the protection scope can be.
Three orders that are easy to get reversed:
· File first, then disclose—not submit the paper first and then think about filing;
· First clarify whether you want to protect a "structure/process" or a "use," then decide how many sets of claims to write;
· First organize the experimental data into tables before writing—"insufficient data for implementation" is the most common problem in patents, and going back to supplement experiments costs much more than for a paper.
The role of tools. Tools can be used for preliminary search, terminology consistency, and organizing data tables; files should be uniformly archived in a personal knowledge base (supports PDF, DOCX, PPTX, XLSX, TXT, and MD; personal use by default, not shared). Do not leave versions of the disclosure scattered in chat records. Judging "whether it is worth filing" and "whether the protection scope is sufficient" is the job of the advisor, patent agent, and intellectual property office; this type of graduate AI assistant does not participate.
Three boundaries:
The patent text is subject to the patent agent's final version; what I write is the disclosure, not the application document;
Novelty must be checked in both patent databases and paper databases; checking only one side will lead to errors;
Confidentiality—before filing, the disclosure is undisclosed content; you must judge for yourself whether it can be uploaded to tools (I confirmed that uploaded data is not used for model training).
Also, there is one point of academic integrity: the experimental conditions and sample sizes for the same set of data must be consistent in the paper and the patent.
A paper and a patent are not two formats of the same thing; they are two exits for the same piece of work.