For decades, the description was the quiet part of a patent application. The music played in the claims; the description supplied background, embodiments, and fallback positions. As long as the claims were clear on their own, the exact wording of the description had few consequences in examination or litigation. Two developments are now unsettling this division of labor — and, of all places, they converge at exactly the same point. Legally, the EPO’s Enlarged Board of Appeal, with G 1/24, has elevated the description to a permanent tool of interpretation. Technologically, artificial intelligence (AI) is increasingly taking over the drafting of that very text. For law firms, this creates a risk that is easily overlooked: the AI’s efficiency gain is concentrated exactly where the duty of care is rising the fastest.
What G 1/24 Actually Decided
With its decision of 18 June 2025, the Enlarged Board of Appeal put an end to a long-simmering divergence in EPO case law. The core holding: the description and drawings must always be consulted when interpreting the claims for the purpose of assessing patentability — not only once a claim, read on its own, appears unclear or ambiguous. The underlying case concerned the interpretation of a single term (“gathered sheet”) in a patent; the Board used it to settle a fundamental question about how Articles 69 and 84 EPC interact.
Two points matter for practice. First, the decision harmonizes EPO practice with that of the Unified Patent Court and the national courts. What is written in the description will henceforth be read alongside the claims according to the same basic principle in examination, opposition, and infringement proceedings alike. The description thus moves from the periphery to the center of claim interpretation. Second — and precision matters here — the Board holds that the description must always be consulted for interpretation; it does not hold that the description may override an otherwise clear claim at will. The practical effect is therefore not a blank check to reinterpret every claim from the description. But it does shift the weight considerably: every definition, every use of a term, every statement about “the invention” can now influence how the claims are interpreted.
This is reinforced by a further, still open issue. Referral G 1/25 raises the question whether the description must be adapted to the claims prior to grant. Both directions carry risk: leaving contradictions in place may dilute the interpretation; adapting the description risks impermissible extensions (Art. 123(2) EPC) and unintended shifts in the scope of protection. However G 1/25 is decided, the message is already clear: the description is no longer neutral background text but an effective tool of interpretation.
From Background Text to a Sharp Tool
This changes the character of drafting itself. In the past, the description could be written generously, redundantly, and with many variants — extra sentences rarely did harm. After G 1/24, the opposite is true: every sentence now carries a potential interpretive effect. A definition that is broader or narrower than intended now counts. An offhand statement about the purpose or advantage of “the invention” can support a narrowing interpretation. An embodiment that promises more than the claim actually covers can become a point of attack. And terms used one way in one place and differently elsewhere create exactly the kind of inconsistency a court will now have to resolve — not necessarily in the patentee’s favor.
The description has thus become an instrument that demands consistency, terminological discipline, and a precise awareness of the interpretive consequences of every word. These are exactly the qualities that AI-generated text systematically struggles with.
Where the AI-Drafted Description Tips Over
AI can produce descriptions in a fraction of the time previously required — fluent, extensive, plausible. That is tempting, because the description is the most text-heavy and seemingly most mechanical part of drafting. But it is precisely the typical weaknesses of generative models that strike at the risks G 1/24 has sharpened:
Terminological inconsistency. Language models aim for variation and fluent phrasing, not rigid term fidelity. A model will use the same term slightly differently in different places, or introduce synonyms that carry a subtly different meaning. In a world where the description is always consulted for interpretation, that is no longer a cosmetic flaw — it is an interpretive trap.
Unsupported assertions and pseudo-embodiments. AI tends to fill gaps “sensibly” — with advantages, mechanisms of action, or variants that are not technically supported. Such passages can steer claim interpretation in an unwanted direction, and — should G 1/25 confirm a duty to adapt — become a source of Article 123(2) problems when later deleted.
Scope bias. Generative models, when in doubt, produce more text, not less. But more text means more surface area for contradictions, overreaching definitions, and incidental commitments — in other words, more of exactly what a court now consults.
Imported phrasing patterns. AI draws on training data full of other parties’ patents. It adopts “characterizing” phrases, implicit disclaimers, or purpose-bound language that quietly shifts the scope of protection — phrasing an experienced drafter would have deliberately avoided.
Definitions with a life of their own. A model likes to write its own definitions into the description. If such a definition — narrower or broader — departs from the intended claim meaning, it now has a direct effect on interpretation under G 1/24.
The insidious part is that these errors don’t look like errors. AI text reads competently and fluently. The inconsistency is not hidden in an obvious blunder but in the interplay of passages that each look innocuous on their own — and only reveal their effect once an infringement dispute arises.
The Firm’s Liability Risk
For a law firm, this bundles into a specific risk. The patent attorney is responsible; they sign the application, and they are liable if an unnoticed inconsistency in the description narrows the client’s scope of protection or brings a patent down in litigation. That responsibility cannot be delegated to a model.
The real core of the issue is a paradox. AI saves the most time exactly where G 1/24 demands the greatest care. The description used to be the part one could most safely “let run through” — and it is now the part that demands the most precise, interpretation-aware scrutiny. The apparent efficiency gain is therefore deceptive: whoever merely skims the AI-drafted description hasn’t worked faster — they have simply pushed the risk, invisibly, into the future: into examination, opposition, and infringement proceedings.
On top of that, the necessary review is not the fast kind of review. Checking an AI-drafted description for interpretive consequences means reading the entire text against the claims: for consistent terminology, for definitions, for overreaching statements about “the invention,” for unsupported embodiments. That comparison is cognitively demanding and often no faster than drafting it oneself — sometimes slower, because someone else’s plausible-sounding phrasing first has to be deconstructed to recognize its interpretive effect. The AI’s time advantage shrinks in direct proportion to the care that G 1/24 now requires.
This is not an argument against AI. It is a precise mapping of its risk: the description is the place where AI’s characteristic weaknesses meet the stakes that G 1/24 has raised. This is exactly where it is decided whether a firm has mastered AI — or whether AI has written a liability problem into the file.
What Firms Should Do Now
The analysis points to concrete safeguards that preserve AI’s efficiency gain without increasing liability:
- Terminology and definition review. Before filing, a targeted check: is every claim-relevant term used consistently and in its intended meaning throughout the description? AI-generated synonyms and self-made definitions must be explicitly reviewed.
- Consistency check against the claims. The description must be read as an interpretive instrument, not as continuous prose. Every statement about purpose, advantage, or “the invention” must be examined for its potentially narrowing effect.
- Restraint with boilerplate. What counts under G 1/24 should be intentional. Overreaching embodiments and unsupported assertions are more of a risk than a fallback position.
- G 1/25 readiness. As long as the duty to adapt remains unresolved, the description should be drafted so that later conformity amendments create as little Article 123(2) exposure as possible.
- AI as a draft, not a result. The sensible use case is an accelerated first draft followed by a documented, human interpretive review — not unreviewed adoption.
- Process and documentation. A firm quality-control process — who performs the interpretive review and how it is recorded — protects both the client and the firm.
Conclusion for IP Practice
- G 1/24 makes the description a permanent tool of interpretation. It is now always consulted to interpret the claims in examination, opposition, and — harmonized — before the UPC and the national courts. Every sentence now has an interpretive consequence.
- AI text is weakest exactly where it now matters most. Terminological inconsistency, unsupported statements, scope bias, and imported phrasing patterns are exactly the errors G 1/24 makes dangerous — and they don’t look like errors.
- The efficiency gain concentrates the liability risk. AI saves time precisely at the point that now demands the most careful, interpretation-aware review. A superficial check doesn’t save time — it merely shifts risk.
- The review is not the fast kind of review. Checking the description against the claims is demanding and often no faster than drafting it oneself — the real time saved is smaller than it appears.
- The firm’s value shifts toward interpretive skill. It is not the fastest drafting that wins, but the drafting that combines AI speed with human, litigation-aware judgment about every word.
G 1/24 thus provides an unexpectedly concrete legal basis for a claim that often stays abstract in the debate over AI in the patent profession: the human remains indispensable in the process — not as a matter of principle, but because the description has become a sharp tool that can only be safely wielded with judgment. Whoever uses AI as an accelerated draft and elevates the interpretive review to the core deliverable gains both: speed and safety. Whoever lets the description “run through unchecked” writes the risk into the file.
