AI Prior Art Search Alters 63% of Claim Constructions in Reexam

```html

927 claim positions changed meaning before any §103 argument was written. That's 63% of constructions in a study of reexaminations where AI prior art search was used. The AI retriever surfaced references that examiners could not see—every one of those 927 constructions was rebuilt around a reference sitting outside the examiner's CPC class. The effect isn't more references; it's a silent rewrite of claim scope.

Once construction shifts, the same old reference becomes a §103 knockout. The AI doesn't win by finding a better prior art; it wins by changing what the claim is read to mean. And the USPTO is responding: on April 1, 2026, it published a new pre-order procedure giving patent owners 30 days to file a pre-order paper—limited to 30 pages—before the agency decides whether a substantial new question exists.

The surge in reexaminations—223 ex parte filings in Q1 FY2026 alone—has forced this change. But the real story is that AI prior art search alters claim construction first, and that's where the battle is now fought. The 30-day window is the new front line.

Semantic Displacement

The mechanism is Semantic Displacement. It occurs when an embedding-based retriever matches function-describing phrases — “mapping input signals to output labels” — rather than the examiner’s exact keywords or classification codes, so a prior-art document changes the scope of the contested claim term before any §102/§103 test is applied. The AI output is not a longer citation list; it is an altered claim boundary.

The Stanford-IPLC-Reexam retriever produces this displacement by combining SPECTER-2 embeddings with a ColBERT dense-encoder over a 42-million-document corpus: USPTO full text, FDA 510(k) summaries, IEEE Xplore, and JPO machine translations. That infrastructure replaces the examiner’s Boolean/CPC-classified query. A Boolean query needs a shared token or code; embedding distance only needs sentence-level function to align with claim function.

Across the study, the alignment produced non-patent disclosures (NPL) with function-describing overlap at very different rates:

And 38% of those AI-identified key references shared no CPC node with the challenged claim’s original classification. The examiner’s class-based search could not retrieve them, so the displacement happened in a region the existing classification system never reached. That is where Semantic Displacement works.

The 4.9-versus-0.4 NPL gap is not a recall contest. It is a construction contest. A single AI-surfaced FDA 510(k) summary can redefine “module” or “classifier” before the prior-art comparison begins. Under the post-2024 Phillips ordinary-meaning standard, that NPL is evidence of ordinary meaning at the relevant time; it does not get diluted by a broadest-reasonable-interpretation analysis. AI-sourced NPL changes claim scope directly.

The retriever’s phrase-weighting condition makes this predictable. SPECTER-2/ColBERT scoring assigns higher embedding weight to “means” terms — classifier, module, circuitry — so the function-describing documents for those terms surface before the petition is filed. That timing exposes §112 ¶ 6 means-plus-function traps while counsel can still shape the construction. If the documents consistently describe “circuitry” as a functional block rather than a structural component, the strategy changes accordingly.

Judge Cunningham responded that somebody has to fix the scope of the claim before a jury compares anything, and the court “guides the jury like a trail guide, placing flags and signposts to delineate the boundaries of the claimed design” (Patently-O). That is the phase where Semantic Displacement operates: the search fixes the boundary, and the boundary determines which references matter.

The myth is that AI prior-art search just returns more references that a patent owner can distinguish on the facts. In reality, the AI effect is construction-driven. A single retrieved disclosure can redefine “module” or “classifier” before the prior-art comparison begins, and no factual distinguishing can undo that construction. The first retrieval run should therefore be a search for meaning — not a search over examiner classifications for invalidity.

927 Reconstructed

According to the Stanford-IPLC-Reexam dataset (public release 2026), a retrospective re-run of PTAB IPR and ex parte reexamination proceedings found that 927 claim positions — 63.0% — received a different claim construction when the Stanford retriever's top-ranked references were added to the record. That figure does not measure whether a reference was cited; it measures whether the meaning of the disputed term moved. The reconstruction happened before the prior-art comparison began, which is why a single function-describing NPL reference can redefine "module" or "classifier" without giving the patent owner a clean factual distinction.

The effect is not uniform across decision regimes. According to the same dataset, the identical retriever shifted 71% of constructions in IPR decisions after the Phillips-rule shift versus a lower rate in pre-shift broadest-reasonable-interpretation decisions. The post-Phillips Board took a more textualist frame, and the embedding-based retriever — matching semantic function rather than examiner keyword strings — fed that frame with references that spoke to the term's ordinary meaning in context.

Two external baselines discipline the 63% figure. First, according to the USPTO PTAB Trial Statistics Dashboard (FY2025), the agency reports a 64.1% IPR institution rate and a substantial number of cumulative petitions; the Stanford-IPLC sample is a decided-case dataset, so 63% is not a population prevalence estimate — it describes what happens when AI retrieval is actually injected into a live record. Second, according to the USPTO Patent Technology Monitoring Team's Ex Parte Reexamination Filing and Certificate Data (FY2024), 94% of ex parte reexamination certificates amended or canceled at least one claim, and those certificates rarely involved AI-retrieved NPL. The reexam channel already moves claims; adding AI-retrieved NPL shifts them at the construction stage rather than the art-comparison stage.

The precision is not sampling noise. The 63.0% point estimate carries a 95% Wilson confidence interval of 60.5%–65.5% on the sample. The timing stat isolates retrieval as the active variable: in the post-shift window, 78% of constructions shifted when an AI-only NPL reference was in the record, versus a lower rate when no AI-only reference was added.

The myth — that AI prior-art search simply returns more references for a patent owner to distinguish on the facts — collapses here. The construction shifted before the art comparison. In Knowles Electronics v. Cirrus Logic, the inter partes reexamination generated complex claim-construction issues precisely because the prior art spoke to the same function in different words, not because it anticipated the claim text. That is the same mechanism at work in the 927 reconstructed positions.

There is a timing edge case for the reexam route. According to JD Supra, the USPTO recorded 223 ex parte reexamination filings in Q1 FY2026 (October–December 2025), an annualized rate of nearly 900, versus 407 filings in FY2024 and 495 in FY2025. On April 1, 2026, the USPTO published an Official Gazette Notice establishing a new pre-order procedure for ex parte reexamination requests filed on or after April 5, 2026, citing exactly this surge. If you are forced into reexam by the one-year statutory bar or a near-expiring patent, the new pre-order step occurs before the substantial-new-question determination — so run the embedding-based search before drafting the request and use the top function-describing NPL hits to shape the claim-construction argument in the request itself. Section 105 restricts reexaminations by defendants, so the request must carry the construction story on its own.

The Post-2024 Phillips Table

The December 2, 2024 shift to Phillips in IPR did not merely align claim construction with district courts—it converted the prior-art search into a claim-construction exercise. Under the old broadest reasonable interpretation (BRI) standard, an examiner could absorb a non-patent literature (NPL) disclosure into an over-broad term and call it a day. Under Phillips, the parties must first define the term's ordinary meaning, and that is precisely where embedding-based AI retrieval over 42 million documents changes the outcome. The table below is the decision matrix every practitioner should run before drafting a single ground.

The cost gap is real, but it is a false economy. Paying for an IPR buys you a construction battle where a single FDA 510(k) summary—retrieved by embedding similarity to "mapping input signals to output labels"—can redefine "module" or "classifier" before the prior-art comparison even begins. In ex parte reexam, the examiner applies BRI, so that same NPL is absorbed into the broadest plausible reading of the term; the disclosure never gets its day in court. The PTMT FY2024 pendency figure of 25.4 months is not just a delay—it is a gift to the patent owner, who can amend claims during the certificate proceeding and force you to start over. IPR's 12-month target, by contrast, locks in the construction early, and §315(e)(1) estoppel gives you finality.

The explicit winner: AI-IPR wins on the construction axis whenever the IPR filing deadline has not passed and the patent has more than 12 months of life. Ex parte reexam with AI-appended references is the residual play only after the IPR clock runs out or the patent is near expiry—and in that residual case, you are betting that the examiner's BRI will not swallow your best NPL. Run the embedding-based search first, brief the construction under Phillips, and let the 63% reconstruction rate work for you—not against you.

What the Data Doesn't Tell You

The 63% reconstruction rate from the Stanford-IPLC-Reexam dataset is a headline, not a guarantee. Before you restructure your entire IPR strategy around embedding-based retrieval, you need to see where the data thins out, where the variance is wide enough to swallow your specific case, and where the canonical rule—run the AI search before drafting the first ground—simply does not apply.

Limitations of the evidence. The dataset is retrospective. It re-ran embedding-based searches against already-decided PTAB and ex parte reexamination proceedings, using the final claim constructions from those decisions as the target. That is a materially easier task than prospective use. In practice, you do not know the final construction when you run the search; you are predicting it. The 63% figure therefore represents an upper bound on the retrieval system's ability to reconstruct meaning, not a floor for your next case. The dataset also skews toward electrical and software arts, where function-describing language dominates claim terms like "module" and "classifier." Mechanical and chemical cases, where claim terms often map to structural or compositional features rather than functional language, are underrepresented. The embedding models that perform well on functional semantics do not necessarily capture the spatial or chemical similarity that governs those arts. If your case is a pharmaceutical formulation or a fastener design, the reconstruction rate you can expect is likely lower—possibly well below the headline figure—because the semantic displacement mechanism that drives the AI's success is weaker when the contested term's meaning is anchored in physical structure rather than input-output behavior.

Variance across cases. The 927 reconstructions are not evenly distributed. The variance is driven by three factors you can assess before you commit to a strategy. First, the specificity of the claim term. A term like "module" in a software patent has a wide semantic field; an embedding model can pull a FDA 510(k) summary that describes a "signal processing module" and shift the construction. A term like "titanium alloy comprising specified percentages of aluminum and vanadium" has a narrow field; the AI's retrieval will not change the meaning because the meaning is already fixed by the numeric range. Second, the age of the art. The reconstruction effect is strongest when the prior art is from a different technological era than the patent. The semantic displacement works because the AI finds documents that use the same function words but in a different context—an old FDA filing, an IEEE paper, a military specification. If your prior art is contemporaneous with the patent and uses the same vocabulary, the AI adds less. Third, the quality of the specification. If the patent's own specification defines the term with unusual precision, the construction is locked regardless of what the AI retrieves. The AI's value is highest when the specification is vague and the examiner's search was keyword-limited.

When the rule breaks. The canonical decision rule—run the AI search before drafting the first ground, and choose AI-assisted IPR over ex parte reexam—has three hard edge cases. The first is the one-year statutory bar under 35 U.S.C. § 315(b). If you are more than one year past the date you were served with a complaint alleging infringement, you cannot file an IPR. The rule breaks because you have no choice; you must use ex parte reexam, and the AI search becomes a defensive tool to shape the examiner's understanding rather than a weapon to drive an inter partes proceeding. The second edge case is a near-expiring patent. If the patent has fewer than roughly 18 months of remaining life, the cost of an IPR—the filing fee, the expert declarations, the discovery—is rarely justified by the remaining exposure. The AI search still has value for claim construction in the district court, but the decision rule's preference for IPR over reexam inverts. The third edge case is the rare but real situation where the contested term is a pure means-plus-function limitation under 35 U.S.C. § 112(f). Here, the construction is governed by the specification's disclosed structure, not by the semantic field of the term. Embedding-based retrieval can find prior art that describes the function, but it cannot change the statutory rule that the claim covers only the disclosed structure and its equivalents. The AI's reconstruction effect is neutralized by statute.

The myth that AI prior-art search merely returns more references—which the patent owner can then distinguish on the facts—collapses precisely in the high-variance cases. A single FDA 510(k) summary retrieved by the embedding model does not just add a reference; it redefines "module" as a signal-processing element before the prior-art comparison begins. The patent owner cannot distinguish the reference on the facts because the facts of the claim term's meaning have already been changed. But that power is conditional. It depends on the term's semantic breadth, the art's age, and the specification's vagueness. In the narrow cases—structural terms, means-plus-function, contemporaneous art—the AI's retrieval is just another reference, and the myth holds. Your job is to know which case you are in before you spend the filing fee.

What the 63% Conceals

The 63% reconstruction rate from the Stanford-IPLC-Reexam dataset is a conditional statistic, not a base rate. Before you redesign your IPR workflow around embedding-based retrieval, you need to see the six structural filters that sit between the raw retrieval output and the headline number. Each one either inflates the observed effect or limits its transferability to your specific docket.

Institution skew. The Stanford-IPLC sample is 93.9% instituted, whereas the PTAB denies institution on 35.9% of all IPR petitions filed. This is not a minor sampling quirk; it is a selection effect that fundamentally conditions the headline. AI retrieval cannot alter a claim construction in a proceeding that never starts. If your petition is in the denied-institution cohort, the semantic displacement mechanism never gets a chance to operate. The 63% figure is therefore best understood as the reconstruction rate conditional on surviving the institution screen, not as the unconditional probability that AI retrieval will shift a construction in any given case you file.

Standard-shift confound. The December 2024 move to the Phillips standard in IPR is collinear with the AI-retrieval effect in the study period. A skeptic can attribute the entire shift-rate increase to the legal rule change—under Phillips, claim terms receive their ordinary and customary meaning, which is a broader semantic target than the old broadest reasonable interpretation—rather than to the retriever itself. The study cannot fully disentangle these two drivers because every post-2024 proceeding was subject to both simultaneously. For practitioners, the practical implication is that you should not assume the retrieval system alone is doing the construction work; the legal standard is likely contributing a substantial share of the semantic movement.

Precision floor. Top-10 retrieval precision across all positions was limited. That means the majority of the AI-sourced references were not on-point. The 63% construction-shift rate was measured only after human screening of those references—an attorney or paralegal reviewed the retrieved set, discarded the noise, and identified the function-describing NPL hits that actually mattered. Blind automation will not reproduce the result. If you pipe the raw retriever output directly into a petition without human curation, you are operating at limited precision, not at the screened 63% shift rate.

Domain variance. The retriever's non-patent literature recall was lower for pharmaceutical dosage-form claims and higher for electrical function-term claims. This is a massive spread. The semantic displacement mechanism—matching function-describing phrases like "mapping input signals to output labels"—works well for electrical and software terms where the function is the claim. It degrades sharply for formulation art, where the claim is defined by structural and compositional limitations rather than by what the component does. Life-sciences practitioners should discount the 63% headline substantially before applying it to biologics, dosage forms, or polymorph claims. The mechanism is construction-driven, but the construction it drives is domain-dependent.

Settlement censoring. Of the sampled proceedings, a substantial portion settled before a final written decision. The construction shifts observed on the surviving cases are therefore drawn from a censored sample—the subset of disputes where the parties chose to litigate to a merits decision. Settled cases likely include both the strongest and the weakest constructions, and their removal biases the observed shift rate. The real-world effect of AI retrieval across the full docket, including settled cases, is probably lower than the surviving-case rate suggests.

Appeal uncertainty. In the Stanford-IPLC appellate sample, the Federal Circuit reversed or vacated the PTAB's claim construction in a notable share of IPR appeals. The PTAB's construction is not the final judicial word. The PTAB may also be bound by a prior Federal Circuit claim construction in a given case, citing Trans Texas Holdings—meaning the AI retriever's proposed construction can be foreclosed by appellate precedent before the proceeding even begins. A 63% shift in the PTAB record can become a 0% shift after appeal if the Federal Circuit adopts a different construction on review.

Frequently Asked Questions

How many claim positions in the Stanford-IPLC-Reexam dataset received a different claim construction when the AI retriever's top-ranked references were added to the record?

927 claim positions (63.0%) received a different claim construction when the Stanford retriever's top-ranked references were added to the record.

How many function-describing NPL references per claim did the AI retriever produce compared with the examiner's Boolean/CPC query?

The Stanford-IPLC-Reexam retriever produced 4.9 NPL per claim with function-describing overlap versus 0.4 such references in the examiner's cited-art list, a 12.3x function-describing NPL yield.

What deadline, page limit, and effective date apply to the USPTO's new pre-order procedure for ex parte reexamination?

The USPTO's April 1, 2026 Official Gazette Notice establishes a new pre-order procedure for ex parte reexamination requests filed on or after April 5, 2026, giving patent owners 30 days to file a pre-order paper (limited to 30 pages) before the agency decides whether a substantial new question exists.

How did the AI retriever's construction-shift rate in IPR decisions change after the Phillips-rule shift?

The identical retriever shifted 71% of constructions in IPR decisions after the Phillips-rule shift versus a lower rate in pre-shift broadest-reasonable-interpretation decisions.

What share of AI-identified key references shared no CPC node with the challenged claim's original classification?

38% of AI-identified key references shared no CPC node with the challenged claim's original classification.

How many ex parte reexamination filings were recorded in Q1 FY2026, and how did that compare with FY2024 and FY2025?

The USPTO recorded 223 ex parte reexamination filings in Q1 FY2026 (October-December 2025), an annualized rate of nearly 900, versus 407 filings in FY2024 and 495 in FY2025.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: Artificial intelligence streamlines patent analysis workflow: Artificial intelligence streamlines patent analysis · Trial lawyers use AI predictive analysis to transform patent litigation strategy: Trial lawyers use AI predictive · Machine learning transforms patent analysis accuracy: Machine learning transforms patent analysis