[

Ask a patent attorney what the job involves and you will hear about technology and about law. The honest answer is that we write things down. It is also why we are such poor company at parties!

What we write down is a business asset. A client arrives with a disclosure and leaves with a right that has to survive an investment case or a fight, years later and rarely the one anybody predicted. The words on the page are where that is decided.

The core skill of the profession is drafting, and drafting is precision of language. The training takes years, and consists largely of adopting an ever more literal approach to language while expanding one’s technical and descriptive vocabulary. You learn to work out what a piece of text really means, as distinct from what its author hoped it would mean. You then learn to write down precisely what you mean. Most people can do neither. Afterwards, for the sake of society, you have to relearn how to be figurative under controlled conditions.

Can a machine do that?

Put that way, the question is too narrow. Whether AI can produce something that reads like a patent specification is settled. It can. The question worth asking is whether it can make the decisions that the text records.

I have drafted several tens of applications using specialist AI tools, so this is not merely an observer’s point of view. Such AI tools produce text that is statistically likely to resemble what a patent attorney might have drafted from the same disclosure. They do it reasonably well for the background, and to some extent for the detailed description, or what we used to call the specific description, before international practice combined together the specific description and the consistory clauses that used to sit ahead of the specific embodiments. There is a great deal of nuance in that last sentence which a short article cannot carry.

Where the tools fail is the claims, and the consistory clauses that now travel under the heading Summary.

That is not a complaint about prose. The claims are where the decisions sit: what to claim and what to give away, how much of the disclosure to spend now and how much to hold in reserve. The text is only the record of those choices.

Why the failure matters is a question of history.

The earliest English patents had no separate claims. The grant identified the invention by a short title, and a written specification became an established feature of the patent system during the eighteenth century. Claims emerged over the course of the nineteenth, and became a statutory requirement in 1883, when section 5(5) of the Patents, Designs, and Trade Marks Act required a complete specification to commence with the title and end with a distinct statement of the invention claimed. The specification contained a technical description and, where necessary, drawings, while the statutory claim appeared at the end of the complete specification. The abstract is a modern arrival, introduced to serve publication and searching.

Even once the claims had become the definition of the monopoly, the courts kept a measure of latitude. Lord Cairns coined the pith and marrow in Clark v Adie (1877) 2 App Cas 315, and for a long time afterwards the English courts were willing, in the right case, to look past the words to the substance of what had been invented.

Then came Electric & Musical Industries Ltd v Lissen Ltd (1939) 56 RPC 23, where Lord Russell of Killowen laid down that “what is not claimed is disclaimed”. The result was a notoriously strict, near-literal approach to construction that persisted well into the early years of the EPC. When I entered the profession in 1989, my superiors explained the difference between British peripheral drafting and German central claiming as though describing two mutually unintelligible faiths. In Munich you claimed the idea. In London you fenced the field, and whatever lay outside the fence belonged to the world.

Moderation arrived with Catnic Components Ltd v Hill & Smith Ltd [1982] RPC 183, where Lord Diplock gave us purposive construction, and with Improver Corp v Remington Consumer Products Ltd [1990] FSR 181, where Hoffmann J turned it into three workable questions. The third question is where the strain shows. To pull a variant back inside the claim, the court had to hold that a patentee who had written one word meant it figuratively, the figure being a synecdoche or a metonymy, denoting a class of which the written word was merely the best-known member. Lord Hoffmann was still quoting that passage of his own in Kirin-Amgen Inc v Hoechst Marion Roussel Ltd [2004] UKHL 46. A profession that chooses its words for a living was being told that its words stood for something other than what they said. Both cases were attempts to reach a sensible answer while keeping up the pretence that the exercise was one of pure construction. That pretence held until Actavis UK Ltd v Eli Lilly & Co [2017] UKSC 48, in which Lord Neuberger, giving the judgment of the court, separated interpretation from the scope of protection and put the metonymical torture of our language to bed.

Attorneys of a certain vintage greeted this as a licence to relax. It is nothing of the kind.

Under Actavis, normal interpretation comes first, and a patentee has no guarantee of the benefit of equivalents to reach beyond the normal meaning of the claims. The Court of Appeal made the point sharply in Salts Healthcare Ltd v Pelican Healthcare Ltd [2026] EWCA Civ 93, upholding a finding of non-infringement in the ostomy bag case. The variant achieved substantially the same result in substantially the same way. It still did not infringe, because the skilled reader would have concluded that strict compliance with the claim wording was intended. The court’s reasons were drafting reasons. The claims were cast in structural rather than functional terms. The technical advantages set out in the description were tied to the claimed configuration and not to other configurations. There were 10 independent claims, which suggested a deliberate carving up of the embodiments. And the limitation in issue appeared in some of those independent claims but not in others.

Every one of those points was settled at the drafting stage, years before anyone gave litigation a thought. Getting the normal meaning right is where the attorney’s skill is still in demand.

Which brings us back to the machine, and to what the independent attorney is for.

The tools earn their keep. They are quick and they take a good deal of routine work off the desk, and I expect to go on using them. What the machine does not touch is the part the client is paying for.

Here is what AI cannot do:

  • It cannot understand the scope of the technical disclosure in front of it, or how that disclosure nestles into the state of the art.
  • It cannot know what protection the client wants, still less what the client needs for the business.
  • It cannot judge which fallback positions will be worth having if unexpected prior art surfaces at a point when the disclosure can no longer be extended beyond what is directly and unambiguously derivable from the application as filed.
  • It cannot couch the specific language of a preferred feature in broader generalised text, hedging the inventor’s bets without a crystal ball.
  • It cannot weigh the relative merits of the various subordinate limitations available to an independent claim.

Each of those is a judgment about somebody’s business, taken under uncertainty, by an adviser who will answer for it afterwards. Each of them is settled in the wording. That is why the drafting is the job rather than the paperwork that follows it.

None of this is a complaint about the size of the models. AI works on probabilities and on rules. Probabilities are ill suited to nuance. Rules were made by wise men for the guidance of fools. So long as courts assess scope by careful analysis of the precise wording of the claims, our jobs are safe. Someone has to decide what the words should be, and stand behind the choice. For now.

If courts, or AI judges, ever revert to assessing the pith and marrow of a disclosure without regard to what the inventor objectively claimed, we might have more cause for concern. That does not seem likely. For now.

FICPI's view

FICPI is working to make sure that the independent IP attorney profession remains fit for the purpose by adapting to the present and the future. AI is a critical part of that, as reflected in the development of the FICPI AI Patent Drafting Masterclass this year. As a global community, built on trusted relationships, FICPI helps strengthen the practice of the independent IP community. FICPI Fora, Congresses, committees and meetings are opportunities to gather insights from the international IP attorney community on any issue, whether it be practice-related or topics of patent and trade mark law, including AI where speakers from around the world shared insights on the topic at our 23rd Open Forum in Budapest in September 2026. 

Next steps

BECOME A MEMBER OF FICPI

FICPI is the only international NGO whose membership consists exclusively of IP attorneys in private practice. As a result, the FICPI community is driven by a shared interest to promote solutions and advocacy for private IP practitioners.
BECOME A MEMBER OF FICPI
MEMBER LOGIN