ALGOPATENT - Automatonophobia

Disclaimer: this article does not constitute legal advice. My views and understanding may have changed since the date of publication. Readers should consult qualified professionals for advice tailored to their specific circumstances.

Automatonophobia [pdf]

by François Veltz - Algopatent

Published on LinkedIn October 2024

Executive summary

The article, authored by Qatent's co-founder, confronts automatonophobia—the fear of robots—in AI-driven patent drafting. Reactions vary from eager adopters to anxious attorneys fearing obsolescence, yet machines will not replace them. Claims drafting demands human compromises across science, law, business strategy, and regulations, where LLMs grasp documented elements like terminology but falter on proprietary contexts. Automation excels at descriptions, offloading tedium such as reference numerals and formal checks. LLMs produce polished but generic text lacking disclosure sufficiency, technical specifics from drawings, intuitions, and corpus denoising. Humans provide lexical instincts; machines expand via synonyms. New horizons include vast portfolio handling, simulations, and analogy reasoning. Ultimately, smart interfaces augment professionals for superior outcomes—embrace the synergy.

Automatonophobia

Automatonophobia is a fear of human-like figures, such as robots.

As a co-founder of qatent, I am facing diverse reactions regarding emerging Artificial Intelligence applied to patents.

Some users are enthusiasts, early adopters and a few are incredible power users. Many are neutral or indifferent. But a few are very anxious and fear that machines will replace them someday. A typical reaction that I receive is "I hate what you are doing but if I do not dive into it I will be obsolete anytime soon". I would like to write to the latter people in this necessarily brief post.

Replacing patent attorneys with machines will not happen in my opinion, for at least several non-exhaustive reasons:

- claims drafting will foreseeably remain the prerogative of humans for a long time. In my view, patent claims are compromises between different aspects: a) science b) legal c) business, and d) regulatory. For a) science: LLMs have "digested" scientific embeddings indeed. They have "some" knowledge of science. For b) legal, certain words or expressions in claims are "socially" accepted ("cryptographic ledger") while some are not ("Cloud computing" in claims is generally not considered clear enough). To some extent, machines can capture such requirements (stable and shared terminology). In other words, handling big data may lead LLMs to catch at least part of these two first documented realities. But for c) business insights and d) regulatory requirements, LLMs have no clues whatsoever and cannot access this contextual data alone (unless the context is packed and provided to machines on purpose, which is a challenge in itself). For example, business considerations may guide the drafter to choose certain words and/or claim structures compared to other alternatives (e.g. single versus multi-provider of services, contributory infringement, claims' categories). The same applies to regulatory affairs (which for example can force having certain mechanisms in place, such as redundancy of systems for safety-critical systems).

- once claims are being defined by humans, the generation of the detailed description can be largely automated, yes. Luckily. In the age of AI, who wants to continue dealing with reference numerals management and other boring tasks such as antecedence checks, dependency checks, formal errors, potential irregularities, and the like? Such low-level tasks are wasting our brain time, which preferably can be refocused on strategic or at least more interesting matters. This is an obvious and natural shift, not a complete reengineering of the profession.

- Large Language Models are the masters of illusion: they are increasingly good at producing good-looking content and hopefully semantically relevant ones (when properly guided by humans). But LLMs have limitations and even strong ones.

- one of these current limitations is that out-of-the-box LLMs hardly can bring sufficiency of disclosure or enablement. They use generic wording. By design, they try to avoid semantic risks. With our classification work at qatent, one key finding is that annotating drawings generally causes humans, not machines, to describe the very details that are needed (e.g. preferred modes, process conditions details such as temperatures, type of materials, and the like). And experimental data, such as clinical data, are out of reach for machines, probably for a long time.

- when dealing with fallbacks, humans have intuitions and machines have almost none. Humans have intuitions regarding lexical directions and machines cannot - or hardly can - have those. When a lexical direction is indicated by humans, then machines can provide exhaustive lists of synonyms, hyperonyms (more generic terms), and hyponyms (more specific terms). Contents can thus be "amplified" by machines (I'll try to elaborate in a next post about this major shift in patents).

- LLMs mimic published contents, being trained on the entirety of the corpus. But this database is full of bad, at least debatable if not obsolete drafting practices (e.g. "means for"). Only humans can help denoising or deobfuscating.

- there are many new opportunities related to emerging man-machine cooperation. For example: handle entire portfolios, of thousands of patent families, that are now out of reach for traditional law firms, even largely staffed. Reinclusion in combination of parts of past patent applications. Coherent generations. Clarity scoring. Content densification. Coupling of text generation with physical simulation. Visualization options. Reasoning by analogy (the term "transposition" cannot be found in patent law articles and rules, examination guidelines, or case Law!). And much, much more.

Of course, some hypotheses may further redefine these continuously evolving man-machine boundaries: e.g. surge of AGI, or at least of AIs coupled with reasoning capabilities and/or physical simulation models.

As of today, temporary and partial conclusions:

- patent attorneys will be augmented by machines, not replaced by them;

- Man and machine are complementary: in particular, humans have intuitions while machines are still blind eidetic systems;

- the real challenge is to have good man-machine interfaces to get the best of both worlds. This requires identifying, understanding, and then leveraging respective limitations and complementarities;

- the assistance of machines opens up many new opportunities;

- Man and machine works can and will be better than those made by man alone, for applicants' benefits.

Do not fear AI in patents, be part of it: let's meet and discuss new opportunities!

What do you think about this new frontier?

Page updated

Report abuse