A patent application succeeds or fails on what it can prove, not on how confidently an AI model predicted it. That holds whether a person or an AI drafting tool wrote it. The application needs synthesized compounds, real data, and claims scoped to the evidence, with the fallback ladder filed early since the US and EPO differ on fixing gaps.
AI now routinely sits somewhere in the discovery pipeline for new drugs, antibodies, and enzymes. A model screens billions of virtual compounds or predicts which antibody sequence will bind a target epitope. Humans confirm, refine, and test the result.
Patent offices don't examine how a candidate was found. They examine whether the application, as filed, demonstrates that the applicant actually possessed the invention and enables one skilled in the art to make and use it. AI-generated candidates don't get a pass on that standard; if anything they invite more scrutiny, because a computational prediction and experimental proof are not the same thing, and examiners and courts increasingly know it.
This article is about that disclosure problem: what to put in the specification, how to frame claims, and where US and European practice diverge. Inventorship, who counts as an "inventor" when a model contributed to the discovery, is a real question, but a separate one; here, we assume that's settled (for now) and ask whether the application supports what's claimed.
There's a mirror-image AI story running alongside the discovery one. On the front end, a model screens candidates and hands back a ranked, confidence-scored list; on the back end, AI-assisted drafting tools increasingly turn that output into a first-draft specification, then carry the case through prosecution: drafting office action responses, generating claim charts, and tracing citations back to the file wrapper.
Solve Intelligence is a platform built for that second job, handling the chemical structures and sequence data a discovery model produces and carrying them from invention disclosure through drafting, prosecution, and beyond.
The two jobs rhyme more than they first appear to: a discovery model's ranked list is confidence-scored inference, not a reduction to practice, and a drafting tool's first-pass claim language is a starting point, not a demonstration that the disclosure actually enables it.
That failure mode isn’t hypothetical: a drafting tool that writes “binds target X with high affinity” into a background section because that’s the language the discovery model returned has made the same mistake that this article is about. That is, quietly converting a confidence score (how sure the model is in its own prediction) into a stated fact about binding strength that only a measured Kd could actually establish.
Neither symmetry lowers the bar. A drafting tool is only as useful as its ability to preserve the exact distinction this article is about, what was predicted versus what the specification can actually prove, and the rest of this article assumes that discipline whether a person or an AI-assisted tool is holding the pen.
A typical AI discovery output is a ranked list: candidate structures or sequences, each with a predicted property (e.g., binding affinity, expression level, catalytic activity) and a confidence score. None of that is experimental data; it's an inference bounded by how well the training data represents the chemistry or biology being claimed.
That distinction matters for drafting, because patent law cares about what was reduced to practice or reliably predicted from the specification, not what a model estimated. Treat the AI output as a triage step, and build the specification around whichever compounds or sequences were actually synthesized, expressed, and tested, with enough real data to justify how far the claims reach beyond them.
None of this means every prediction deserves equally little weight. A model with an established, validated track record in a narrow domain sits closer to the kind of prediction that can count as plausible from the filing date than a novel, unvalidated model’s output does, which is worth documenting alongside the prediction itself. That is a reason to weigh the prediction more heavily, not a reason to skip confirmatory data.
A claim to a molecule or biologic is either structural (defined by what it is: a formula, a sequence, a set of CDRs) or functional (defined by what it does: binds a target, inhibits an enzyme). Structural claims are narrower but easy to support once the compound is disclosed. Functional claims are valuable precisely because they're broad, covering molecules the applicant never made, and that breadth is what draws challenge.
The Supreme Court's unanimous 2023 decision in Amgen v. Sanofi confirmed that patent claims to an antibody genus, like claims of any scope, must be supported by disclosure sufficient to enable a skilled artisan to make and use the full breadth of what is claimed. The reasoning reaches any functionally-defined genus, including AI-nominated compound classes defined by predicted activity rather than shared structure.
AI output tends to push drafters toward functional claiming, since the model's own language is often functional ("compounds predicted to inhibit target X"). Resist that where the data doesn't support it: if the examples share a real structural motif, claim the motif; if they don't, keep the claim honestly functional and match the evidence to that scope.
"Prophetic examples,” write-ups of experiments designed but not yet run, written in present or future tense, remain permitted in both the US and at the EPO, and are routine when not every embodiment can be tested before filing.
The AI-specific risk is circularity: a prediction is fine support when it rests on established structure-activity relationships, but weak when its only basis is the same model's output being used to justify the claim built on that output. Examiners and courts are increasingly unwilling to accept that bootstrapping. Back the prophetic examples that matter most to claim scope with at least some confirmatory data, and be selective about which candidates get full write-ups.
Disclosing a structure isn't enough, the application must also teach how to make it. This is one of the classic In re Wands enablement factors, which weigh guidance given, working examples, predictability of the chemistry, and the experimentation a skilled chemist would need.
AI-generated candidates create a specific failure mode: a model can propose structures that are synthetically difficult or inaccessible, since nothing in training penalized synthesizability. Confirm a viable route before a candidate goes into the application, and disclose a general synthetic scheme plus representative compounds spread across the claimed genus. A genus built on one synthesized compound and forty prophetic ones is built on very little.
AI tools predicting binding paratopes or epitope contacts are now routine in antibody discovery, but post-Amgen the bar for turning a predicted binder into a defensible claim is high. Sequence claims (defined heavy/light chain pairs, defined CDRs) remain the most defensible starting point, and must go into a sequence listing under WIPO ST.26, now required by both the USPTO and EPO. Broader structure-function claims need real characterization across multiple family members: epitope mapping, binding kinetics, and enough examples that a court won't view it as a "hunting license" over the genus.
One useful distinction: in April 2026, the Federal Circuit upheld method-of-treatment claims using a well-known antibody genus (anti-CGRP antagonist antibodies for migraine) in Teva Pharmaceuticals International GmbH v. Eli Lilly and Company, because the inventive contribution was the therapeutic use, not the antibody structure, letting the specification lean on background knowledge of the genus rather than re-proving it. If the AI-discovered element is a new use for an existing, well-characterized antibody class, claiming the use can be far easier than claiming the composition.
Given how exposed broad functional and genus claims now are, file breadth with a real ladder underneath it: the broadest claim the data supports, then intermediate claims (narrower structural families, specific binding mechanisms), down to the specific compounds or sequences actually tested.
Jurisdiction shapes how that ladder gets built. In the US, enablement and written description are assessed as of the filing date under 35 U.S.C. § 112(a); continuation practice allows some scope adjustment later, but new claim language still needs support already present in the filing.
At the EPO, both sufficiency (Article 83 EPC) and inventive step (Article 56 EPC) can turn on plausibility: the Enlarged Board's G 2/21 (2023) made it clear that post-filed evidence to support a claimed technical effect for inventive step is allowed, but only if that effect was already plausible from the original filing.
In other words, post-filed evidence cannot cure a sufficiency gap. The EPC’s added-matter rule under Article 123(2) requires that any amendment be directly and unambiguously derivable from the application as filed. This creates a genuine tension with impermissible broadening. The strictness of that standard is considerably greater than what US practice requires. The upshot: at the EPO especially, the fallback ladder needs to exist in the application as filed, not get built during prosecution.
AI-assisted drafting tools mirror the discovery-side triage here too: just as a discovery model ranks candidates without proving any of them, a drafting tool can track which rung of the fallback ladder has support in the as-filed text, flag EPO added-matter risk before an amendment goes in, and keep sequence listings formatted to WIPO ST.26 as claims move through prosecution.
But the parallel cuts both ways. Just as the discovery model’s ranking doesn’t substitute for confirmatory data, the drafting tool’s tracking doesn’t substitute for the underlying specification actually carrying that support: it can enforce the ladder, it can’t manufacture the rungs.
None of this changes because a model helped find the candidate, it just raises the stakes on managing the gap between what the model predicted and what the application can actually support. If a second AI system is helping draft and prosecute that application, it needs to be managing the same gap, not papering over it.
