AI in Patent Drafting: Key Principles
Traffic rules on the drafting superhighway.
Our patent drafting tools have been quite a hit, ever since we launched a first version late in December 2023.
Drafting patents is an extremely specialised field, with complex nuances, legal and technical. Expertise matters, and experience - a hell lot.
This is central to our considerations when building drafting tools for experts who write patents. We are here to empower highly competent professionals, and it is important to hold certain principles dear.
Lets talk about them, because they are relevant not only for patents but drafting problems in general.
Drafter in the driving seat
AI and computer systems aren’t new to complex and high stakes tasks.
An example of a high stakes task that can go awry without a human intervention, is flying an airplane. Make no mistake, computer systems onboard planes are extremely good at flying the plane, but the point remains.
There are two things that every airplane with an autopilot, and software flying the plane is constructed to respect:
One, the captain and his co-pilots can take control anytime.
Two, the plane is never flown without a captain, and co-pilot, Not only are they trained to fly the plane themselves, they are trained to keep a close eye on the plane when it is flying itself, ready to take over, intervene or alter the inputs.
Software and AI are essential tools but do not (and are not meant to) replace the captain and co-pilot.
Photo by Şahin Sezer Dinçer on Unsplash
Leverage Language Models, for Language not Knowledge
I speak to drafting professionals everyday, and see a disconnect in what they think AI is meant to accomplish, and what AI is able to accomplish.
This is not to say that their expectations are high, and what LLMs deliver are not.
It is to say that many are expecting something else, while losing sight of what can actually be achieved.
LLMs are Language Models. Not Knowledge Models.
They are good at a lot of things that have to do with language: extrapolating, ideating, expanding, summarising, drafting in a particular style, processing knowledge elements into comprehensible text, etc. (This is HUGE. Computers were terrible at this 3 years ago)
They are also terrible at things like: research, finding things, working with large amounts of text, tasks that involve multistep logic, comprehension and application. In fact, they perform worse on some of these tasks than existing technologies.
This is a feature, not a bug.
LLMs are statistical programs. Not following only a set of logical steps, but extrapolating based on a statistical model.
As statistical programs, the results will always require human supervision. Because while of a high quality, the result is one of many likely good results, and there will always be a matter of choice and iteration.
Photo by Şahin Sezer Dinçer on Unsplash
In order to leverage what’s good and reduce the involvement of AI in its areas of weakness, drafting tools like ours should separate the knowledge elements from the language elements.
The drafter brings in the knowledge elements, and the tool constructs the language, style and extrapolates, summaries, and provides a first draft adhering to the knowledge elements, with minimal unnecessary elements or hallucinations.
Hallucinations are a function of a language model straying from the knowledge elements or expanding into areas that go beyond the user’s requirements.
Safety and Security
As a drafting professional, it is essential that my inputs and outputs are not used by any third parties, or accessed by them. It is important that they don’t train or improve any language models, or allow a third party to enrich themselves with data. It is important that data is encrypted and stored securely.
The best approach to this today is to rely on enterprise grade access and security when running closed-source language models such as GPT 3.5 or GPT-4, with a careful eye on their terms of use so that it captures the essentials above (example: Microsoft Azure). With open source models improving at a quick pace, the day is not far when organisations can self-host the models and compute, obviating the need to rely on enterprise grade software API endpoints.
AI is your organisation’s best friend but also can be its greatest enemy.
Organisations must provide their drafting professionals a streamlined and safe way to access AI to get work done.
If you don’t, they are likely to use AI in haphazard and dangerous ways.
Photo by Alice Dietrich on Unsplash
For example: Young associates who draft claims that are all over the place using ChatGPT? Parts of the invention descriptions read like an AI model hallucinating?
Not only does having a streamlined tool help ensure that the necessary work is being done, allocated efficiently between humans and AI models, but also to ensure that one is not doing the work that the other is supposed to.
Drafting tools should focus on adding the right safeguards, requisite bells and whistles, training modules and increase (not decrease) levers for the draftsmen to control when drafting important sections, such as the claims section.
Everything else should also be based on additional touch points carefully calibrated to ensure human intervention at the right time, place and of the right quality.
AI is your pretext to empower your organisation in streamlining how drafting is undertaken, and to bring together style, formatting and the knowledge expectations that go into producing a perfect draft.
But to do that, it is important to have these principles in place.