AI patent drafting put to the test

Since the release of ChatGPT in late 2022, many inventors have taken to using AI to draft their own patent applications without professional assistance. Fortunately for me (as a patent drafter), the AI-drafted patent applications I’ve seen to date have been universally awful. I have no doubt there will be scandal within the next few years as these AI-drafted patent applications are taken up for examination and the inventors learn they have lost out due to unrectifiable defects in their wording.

BRM Patent Attorneys are long-term sponsors of the Inventors Association of Australia (Victoria), and recently I thought I saw an opportunity to present a cautionary tale to the group. Author, inventor and early adopter of AI, T.S. Blackwell-Hart, presented to the group on how to get the most out of the AI tools. As T.S. outlined the basics of AI, the risks of hallucinations, and prompts that can be used to minimise those risks, the idea of putting AI patent drafting to the test was born.

What followed was a three-way challenge starting with a drafting exercise of the type used to train patent attorneys. The exercise describes a simple invention and the relevant background technology. The challenge was to ‘claim’ the invention – that is, to prepare wording to define the coverage of a patent application.

I am a patent attorney with 20 years’ experience in IP. For this exercise, I drafted my version of the main claims and went up against T.S. plus AI, and against an AI-novice plus AI plus some suggestions from T.S.

The results came as surprise: The AI-drafted claims were not terrible! So, let’s have a look at the exercise, the claims, what went right, what went wrong and what can be learnt.

Legal background

Generally speaking:

• a patent claim can be thought of as list of features that something must have to be covered by the claim;
• the validity of a claim is judged as at its ‘priority date’, which is usually the filing date of the initial patent application;
• a claim is invalid if it covers anything that was already publicly known and/or anything that obviously follows on from what was publicly known.

In technical terms, the last point corresponds to the ‘novelty’ and ‘inventive step’ requirements. These are the two key requirements to focus on for this discussion, but there are several other requirements – e.g. the claims must be clear, and the patent application must ‘support’, and provide ‘enabling’ details, commensurate with the scope of the claims.

The novelty and inventive step requirements lead to a trade-off: the more a claim covers, the less likely it is to be valid. A quality patent will include ‘dependent claims’ that cover less than the main claims, to provide fallback positions in case the main claims are later held to be too broad (e.g. because new evidence of relevant older technology comes to light).

Technical background

The exercise explained: ‘Bird feeders are known which have rotatable guards in order to prevent animals other than birds from gaining access to the food. In trying to cross the guard, an animal (e.g. a squirrel) causes the guard to rotate. The sudden rotation of the guard surprises the squirrel, so that it slips or jumps off the guard.’

The invention

The exercise also set out the hypothetical inventors’ explanation: ‘Our solution to this problem is to equip the guard with motorised rotation that is triggered automatically when weight applied to the guard exceeds a predetermined threshold. The motor provides a rapid start and sufficient rotational speed to surprise a squirrel, while the feeding unit remains stationary, allowing birds to continue feeding.’

In the illustrated version of the invention, the guard 9 was supported by compression spring 14. When a squirrel pressed down on the guard 9, the spring 14 compressed, triggering switch 12 to energise motor 13 to rotationally drive the guard 9 via gears 15, 11.

The claims

The highlighting is mine. Based on past experience, I had expected to cover the AI claims in red, signifying major problems that would render the claims unenforceable against a close copy of the illustrated version of the invention. But the AI claims are much better than expected. There is no red – instead there is:

• yellow, highlighting wording that I regard as less than ideal; and
• green, highlighting drafting choices worthy of discussion.

Wording that is less than ideal

Two of the features highlighted yellow enable competitors to produce similar products without infringing.

The illustrated version of the invention has a food container 3 and a food tray 5, whereas it would seem that there are other versions of the invention that should be covered, even if those versions are not ideal, e.g. it seems that versions of the invention that have a food tray 5 but no food container should be covered. The patent attorney and AI-novice claims both cover this option. Indeed, both of these claims would cover version in which the container and tray 5 are swapped for a ‘feeding portion’ or ‘feeding unit’ as simple as a bird seed block:

Likewise, it would seem that the weight-sensing switch could be relocated so that it is not ‘positioned between the guard [9] and feeding unit [5]’.

‘Rapidly’ is a relative term in relation to which there is scope for argument. How rapid is ‘rapid’? More than likely, when the patent application is taken up for examination, the patent examiner would object that ‘rapidly’ is unclear. Hopefully that objection could be overcome with amendment (e.g. to replace ‘rapidly’ with ‘at at least XXX rpm’ if ‘at at least XXX rpm’ is adequately disclosed in the original patent application).

If ‘rapidly’ remained in place when the patent was granted, it would then complicate enforcement because there would be scope for the competitor to argue:

• ‘my guard rotates but does not rotate rapidly, and therefore I don’t infringe’; and
• the claim is unclear and therefore invalid.

If the patent owner were well funded and chose to contest these issues in court, I would expect them to succeed. Nonetheless, the scope for argument is less than ideal:

• Firstly, there is the risk that the patent owner won’t succeed on these points.
• Secondly, the scope for such argument diminishes the deterrent value of the patent and increases the risk of having to initiate court proceedings to enforce the patent – if you threaten enforcement of your patent, a competitor (especially a well-funded competitor) is far more likely to acquiesce if there no scope for reasonable argument.

Drafting choices

The patent attorney (my) claim 1 covers a lot more than the other claim 1’s. It is not limited in terms of weight sensor (etc), and as such covers versions of the invention in which (for example):

1. the motor operates continuously;
2. the motor operates on a timer;
3. the motor is triggered in response to something other than a weight sensor, say optical sensors that react to a squirrel on top of the guard.

With this claim, a few extra words (e.g. to mention options 1, 2 and 3) outside the claim would be required to ‘support’ the claim.

My proposed dependent (fallback) claim 3 approximately corresponds to the other two claim 1’s.

The additional claim scope is not necessarily an advantage. It increases the risk of a patent examiner locating evidence of an old technology covered by the claim and in turn objecting that the invention (as defined by the claim) is not novel. Such an objection could be overcome by filing an amendment to fall back to claim 3, but facing the objection and filing the amendment would entail some cost.

The claim scope should be carefully considered and discussed with the inventors at the outset. If all commercially relevant versions of the invention will inevitably have a weight sensor, limiting the patent coverage to a ‘weight sensor’ at the outset is likely the pragmatic option.

In the same vein, all three claim 1’s are currently limited in terms of a motor. Could a commercially relevant version of the invention be implemented without a motor, e.g. could a mechanism akin to a Yankee screwdriver mechanism be used so the weight of the squirrel provides the motive force to turn the guard? Should the claims be expanded to cover such a possibility? Or would such an expansion simply increase the risk of objections without corresponding benefit?

Words such as ‘trigger’ and ‘weight sensor’ should be accompanied by discussion outside the claims to ‘support’ these broad terms and inform their meaning, e.g. it should be made clear that the illustrated versions of these features are not the only options, and context should be provided to ensure ‘weight sensor’ cannot be reasonably interpreted in a narrow sense that requires something that quantifies the weight of the squirrel.

‘When’ is not the same as ‘in response to’ in my proposed claim 2. There is scope to argue what ‘when’ means. This branch of the AI claim language might be interpreted so as to be satisfied by a motor that does not rotate in response to the weight, but rather is merely capable of rotating when the weight is applied. This wording has the potential to draw objection and/or lead to legal argument, in line with the above discussion of ‘rapidly’, albeit to a lesser degree. I have flagged this green rather than yellow because I regard it as a small point and within the bounds of what a competent patent attorney might write.

Similarly, it’s not really the ‘motor [that] is configured …’; rather the motor is (apparently) entirely conventional and it is the trigger (e.g. the combination of the spring and the switch) that causes the motor to rotate in response to the weight.

In my view, the word ‘unit’ is fine in this context, but care should be given to even such innocuous words. In another context, there might be a design-around that entails separate components that are not a unit.

‘[C]haracterised in that’ is routine wording often called for by the European Patent Office, where it is understood to divide the claim into a ‘preamble’ setting out the combination of features that was already publicly known, and a ‘characterizing clause’ identifying the features that set the invention apart from what was publicly known.

In most countries, this two-part form is not required and an acknowledgement that certain features are publicly known might help anyone (e.g. a patent examiner or an infringer) arguing that the claim is not valid. Even in Europe, how best to divide the claim into this two-part form could change over time as patent examiners locate evidence of more relevant background technology. Hence, in my view, the AI-novice claim could be improved by replacing ‘characterised in that’ with ‘wherein’ (and reinstating ‘characterised in that’ much later on in the process, and only in Europe, and only when a European examiner calls for it).

Why are these AI claims better than expected?

Firstly, the input information (the drafting exercise) clearly described the relevant technical background and the invention, and how those two things differed. In real world drafting, a key part of a patent attorney’s skill is to ask inventors the right questions to draw out quality input information.

Secondly, the input information included the key paragraphs set out above which give a very clear suggestion as to what a reasonable claim should look like. The AI has evidently followed the hints to arrive at two claims consistent with the suggestion. As above, the suggested claim may not be of optimal scope.

More realistic input information may well have led to very different results, e.g. if the input information instead suggested that ‘the invention includes a conical guard formed of 0.8 mm thick steel’, the AI claims could well be limited to these specific details and therefore easily avoided, e.g. by adopting a 2 mm thick plastic guard.

Thirdly, this was a very simple example. Not only was the technology simple, but the nature of the technology was easy to put into words. The key components had commonly used generic names (e.g. ‘motor’ and ‘guard’) and interactions between these components were likewise straightforward.

So, can generic AI draft a quality patent application?

No: It can’t, not yet. Certainly not without professional assistance.

As above, these AI-drafted main claims are much better than I expected based on real world experience with AI patent drafting, but it doesn’t follow that this exercise suggests that patent drafting using generic AI drafting is an advisable course.

Even in the artificial confines of this exercise, beyond the main claims, the AI output was problematic. Neither set of AI output produced a complete patent application. The output elicited by the AI-novice included:

The Solution

The present invention addresses this problem by providing a bird feeder with a motorized guard that is activated automatically by a weight sensor.

In accordance with the invention, the bird feeder comprises:

• A feeding unit for storing and dispensing food.
• A guard rotatably mounted relative to the feeding unit.
• A motor and a weight sensor.

The weight sensor is configured to detect when a weight applied to the guard exceeds a predetermined threshold (e.g., 250 g). Upon reaching this threshold, the motor is energized to rapidly rotate the guard at a speed sufficient to surprise and dislodge the animal (e.g., 30–35 rpm).

Key Advantages:

• Active Deterrence: The motor provides an immediate, high-torque start that a squirrel cannot anticipate or counteract.
• Stability: The feeding unit remains stationary during rotation, ensuring that birds currently feeding are not disturbed or endangered by the moving guard.
• Geometric Efficiency: By utilizing a sloped shield (preferably at least 45°), the combination of gravity and centrifugal force ensures the animal cannot maintain a grip during motorized rotation.

(my emphasis)

At least arguably, this text renders the AI-novice claim 1 invalid because:

• it suggests the invention is limited to high torque, the feeding unit remaining stationary and the sloped shield (etc); and
• the claim is not limited to these things and is therefore not supported.

Can I prepare a patent application without a patent attorney?

Yes: Just as you can pull your own teeth without a dentist 🙂

This has been an interesting exercise and a way to demonstrate some of what we do as patent attorneys. The AI results were much better than expected, but ultimately reassuring that AI is not coming for my job just yet … well, not this year … I hope 🙂