Large Language Models (LLMs) are AI tools trained on large datasets to generate human-like text. These tools are becoming increasingly ubiquitous in daily life, providing helpful assistance ranging from drafting emails and automating tasks to aiding creative pursuits.
Time pressures and cost constraints can incentivise patent applicants to seek aid from LLM based tools such as chatbots for patent drafting. Although LLM tools may seem like a worthy patent drafting assistant, many of these tools present a minefield of risks to the validity and commerciality of patent rights.
The disclosure problem
For a patent right to be valid, it must be new and inventive at the time of its “priority date”, which is the earliest filing date of the patent application disclosing the invention. Any public disclosure of an invention before this date can result in a loss of patent rights. Therefore, it is important to keep an invention confidential until a patent application is lodged.
Prompting an LLM tool using technical information relating to an invention may constitute an invalidating public disclosure if the terms of service of the tool allow for model training with user input, as the “confidential” inputted data is at risk of output to another user.
Some premium LLM tools advertise zero-retention environments, where user data is excluded from training data. In these systems, although data is deleted after processing, it still exists in memory during the process, leaving it potentially vulnerable to interception via “man-in-the-middle” attacks, data leakage via hallucination, or prompt injection attacks where users manipulate input to an LLM tool to exploit the model.
Completely air-gapped (totally disconnected from remote servers) LLM tools provide the safest LLM tool options for confidential data, but these tools can be inaccessible outside of large organisations or institutions.
Furthermore, to draft a patent specification providing the best commercial protection for an invention and to inform patent strategy, it is imperative that the patent attorney understands all relevant business and technical information. It follows that, to generate anything remotely resembling aspects of a commercially valuable patent specification, such detailed information must be provided to an LLM tool when the tool is prompted to undertake tasks related to patent drafting.
In Australia and New Zealand, patent attorney privilege protects communications between a registered patent attorney and their client from disclosure in court where it relates to intellectual property advice.[1],[2] This important tenet allows for full and frank communication between the client (or potential client) and patent attorney that remains secret, including all advice, strategic discussions, and technical/commercial information. Unfortunately, LLM tools have no such legal protection. This means that confidential information given as prompts to an LLM tool is at risk of being compelled to be disclosed in a lawsuit, which may legally jeopardise patent rights or contentious dispute cases related to patent rights.
The obfuscation problem
Patent specifications are legally precise documents, which are held to high standards of clarity and completeness. Terms used in patent specifications to describe an invention are generally carefully considered and selected for their legal definition, having regard to interpretative risks under local and international law.
However, LLM tools are liable to prioritise fluency over facts, resulting in dilution or distortion of description of an invention in favour of smooth prose. When using LLM tools for aid with patent drafting, there is a risk of generating output that does not adequately describe and capture an invention, or that includes description of incorrect or impossible technical principles. A patent specification including such flawed LLM tool generated output may not only risk the commerciality or value, but also the validity, of the patent right.
Moreover, relying upon predictive text generated by LLM tools raises concerns about the quality of description of inventive concepts. As discussed above, an invention described in a granted patent must be new and inventive. Because LLM tools generate output based upon training data that is known, such output may be at risk of being biased towards what has come before, downplaying inventive characteristics that are important to distinguish the invention from any prior art.
Some LLM tools advertise patent drafting capabilities and are designed to be trained on patent specification data. These purpose-built tools may alleviate some obfuscation concerns; however, the LLM transformer architecture inherently struggles with long form inputs and outputs – such as patent specifications and the technical documents upon which they tend to be based. A practical and legally cautious approach involves careful assessment of all LLM tool generated output by a jurisdictionally qualified patent attorney.
Most importantly, patent claims, which define the scope of legal protection of a granted patent, must distil an inventive concept into a legal definition while taking into account numerous factors including commercial interests, jurisdictionally relevant legal principles, and patent strategy. LLM tools, when employed appropriately, may provide helpful patent drafting assistance. Nevertheless, a skilled patent attorney will ask the right questions of a prospective patent applicant to inform preparation of a comprehensive and commercially minded patent specification, while an LLM tool does not “know” what it does not “know”.
Efficiency is no substitute for commercially powerful and enforceable patent protection
Security of confidential information is paramount when dealing with technical developments and inventions. Providing any sensitive technical or commercial information to LLM tools should be avoided to ensure confidentiality and reduce risk of patent invalidity.
A patent application first disclosing an invention sets the priority date for the entire patent family, and it can be difficult (or in some cases impossible) to remedy issues caused by a poor quality first filing. All prospective patent applicants should consult a qualified patent attorney to draft comprehensive patent specifications tailored to commerciality and strategy.
[1] (Australia) Section 200 of the Patents Act 1990
[2] (New Zealand) Section 54 of the Evidence Act 2006