Minesoft, 2025 Yılını Patent Açısından Değerlendiriyor.

· Patent Raporu

Bu makale, İngilizce yayınlanmıştır.

The Examiner Has a New Assistant. So Should You.

AI-driven prior art search went live at the USPTO and EPO in 2025. Here's what that means for applicants.

2025 was the year AI search moved out of the back office and into live examination workflows. Not in a limited internal test, and not just at one office. At both the USPTO and the EPO, AI-driven prior art search became part of the actual process, touching real applications.  Looking at how these systems work in practice reveals important lessons about precision, recall, and the continuing role of human judgment in patentability analysis.

The USPTO’s ASAP! pilot brings AI in before the examiner. An internal AI system uses CPC classification alongside an application's specification, claims, and abstract to search USPTO-accessible patent corpora and limited non-patent literature to produce an AI-Assisted Search Results Notice (ASRN), sent directly to applicants before substantive examination begins. The ASRN lists up to ten references ranked by similarity. The USPTO is clear that ASAP! is a proof of concept, designed to test whether early AI results affect prosecution strategy and quality. But the fact that AI search results are now reaching applicants' inboxes before an examiner has read a single claim is, by any measure, an inflection point.

The EPO's approach works differently. ANSERA (Another Search Era) and its cloud implementation ANSERA-based SEARCH (AbS) embed AI inside the examiner’s search console. Examiners can run concept-based queries, use find-similar functions, execute figure-based searches, and use advanced operators across worldwide patent and non-patent literature collections, with AI-driven ranking and automated relevance markers to highlight potentially relevant passages. And while ANSERA was launched prior to 2025, the EPO has achieved a roll-out to 2500 examiners at national patent offices around the world. The infrastructure is no longer experimental, it is being institutionalized.

So, how much can we trust these machines? On precision and broader relevance, the story is relatively positive. Both ASAP! and ANSERA/AbS are designed to quickly surface the closest prior art in enormous corpora where human examiners are already overstretched. Early commentary suggests that in many cases the AI’s top-ranked documents align with what examiners would later cite, potentially reducing time spent wading through irrelevant hits and sharpening the focus of early patentability discussions. In that sense, AI is a natural extension of longstanding efforts at both offices to improve search tools and make better use of classification and fulltext data.

Trust is harder to earn when you consider recall and coverage of claim features. ASAP! is limited to USPTO-accessible data and caps its output at ten references. The USPTO makes no claim those ten documents cover every limitation in the independent claims or fully describe the prior-art landscape. ANSERA and AbS have substantially broader coverage, but EPO carefully frames them as tools for improving completeness and consistency, not as guarantees of exhaustive recall. The concern practitioners keep raising is a fair one: AI finds what looks similar, but the most consequential prior art is often the unusual piece, the non-obvious combination, the document that won't surface without a creatively crafted search strategy.

There's also the question of legal relevance and if AI can truly understand claim scope. ASAP! operates on text and structure, but whether a cited document anticipates or renders obvious a specific functional limitation or a nuanced technical effect is still a question for human reasoning. The EPO's 2025 Guidelines for Examination are explicit that examiners retain full responsibility for novelty and inventive-step assessments and for mapping each claim feature to the prior art. The AI can classify, allocate, suggest and summarise but the judgment call stays human. Good examination quality, wherever you measure it, has to be assessed against substantive outcomes, not against how often humans happen to agree with the machine's ranking.

For now, the safest reading is that both offices are building human-in-the-loop AI rather than asking anyone to simply trust the machine. USPTO examiners aren't bound by ASRN output and are expected to conduct their own searches in parallel; they can supplement, contradict or ignore the AI's ranking entirely when forming rejections. At the EPO, ANSERA, AbS and related tools sit within an explicitly human-centric framework: AI handles pattern recognition at scale; humans handle interpretation and judgment. Trust, in this design, comes from giving examiners better lenses. Not from taking decisions away from them.

Sophisticated applicants won't choose between office AI and independent search. They'll use both, document the difference, and stay in control of the picture.

That brings us to the practical question for IP professionals. If you can't fully trust the machine, how do you work with it? The better approach for your search is to complement AI search methods with other, user-controlled search approaches. Platforms like Minesoft's PatBase and Origin are built precisely for this gap, offering concept-based and similarity search using AI across more than 180 million documents from 109 jurisdictions, with the ability to iterate strategies transparently and produce an audit-ready record of prior-art review before filing and throughout prosecution. That kind of documented, human-in-the-loop process isn't just good practice. In a landscape where examiners arrive at interviews having already seen AI-ranked references, it's a competitive necessity.

In 2025, both the USPTO and EPO operationalised and upscaled AI prior art search, as a sophisticated applicant you will want to match up:  use both AI and other search methods, document the difference, and stay in control of the picture. Rather than staying sceptic about AI, you use it well.

Sources: