Semantic search widened the net for prior art. Attorney review still decides what it means.

Start with the commercial problem: why does prior-art search deserve this much attention?

Prior-art search determines whether an invention is worth pursuing and how broadly it can be claimed. A patent granted over undiscovered art is a weak asset, vulnerable in licensing, in litigation, and at the examiner's desk. For a client, the cost of a missed reference is rarely the search fee. It is the prosecution spend, the business decisions made on a false read, and the risk that a competitor invalidates the claim years later.

Traditionally this work rewarded patience and a good searcher's intuition. An experienced professional would build Boolean queries, refine classification codes, read broadly, and still worry about the reference nobody thought to look for. The volume made truly exhaustive review impossible, so we managed the recall-versus-precision problem by hand: cast wide and drown in noise, or cast narrow and miss things. That tension is the real reason the search step is slow and expensive, and it is precisely where a different retrieval method can help.

Semantic search is not a new idea. Why is it viable now, and not five years ago?

Three things converged. First, the models. A modern embedding model represents a claim or a disclosure as a vector that captures meaning rather than exact wording, so a query can find conceptually similar art even when the vocabulary differs. That matters in patents, where drafters routinely coin idiosyncratic terms for ordinary components.

Second, the training signal. Some of the strongest patent-search models are trained on examiner citations, the references examiners themselves pulled as X, Y, and A art. That gives the system a notion of relevance grounded in real prosecution rather than surface word overlap. Third, the engineering. Embeddings for the entire corpus can be pre-computed, so a search that once ran overnight now returns in seconds, and non-patent literature can be indexed alongside patents. Five years ago the models were weaker, the non-patent coverage was thin, and the tooling was not built for a prosecution workflow. The method existed; the accuracy and the plumbing did not.

Walk through how a search actually runs, step by step.

We start from the invention disclosure, not a keyword string. An attorney or a technical specialist writes a structured description of the inventive concept and, where it helps, the individual claim limitations. The system normalizes that input, resolves synonyms, and converts it into embeddings. It then searches the vector index across both patents and non-patent literature, retrieving conceptually near documents and ranking them by relevance.

Good tools cluster the results so related disclosures group together, and they let us search at the level of a single limitation, which is how patentability is genuinely assessed: element by element, not document by document. What comes back is a ranked, de-duplicated set of candidates, each with the passages that drove the match. From there it is human work. A patent attorney reads the top references, judges real relevance, maps disclosures against the claims, and decides what it means for novelty, obviousness, and claim scope. The machine finds and orders. The attorney interprets, and is accountable for the conclusion.

Patent work carries strict confidentiality and disclosure duties. How do you handle governance?

This is the part a peer should ask about first. Putting an unfiled invention into a third-party tool can, depending on the vendor's terms, amount to disclosing it to that provider, and that can jeopardize confidentiality and even patent rights. So before any client matter touches a system, we do diligence on data handling: where inputs go, whether they are retained, whether they train the model, and what the contract actually says. We favor arrangements where client data is segregated and not used for training, and for the most sensitive matters we keep the inventive concept out of external prompts entirely.

The USPTO's 2024 guidance is explicit that the duty of candor and good faith extends to the use of AI tools, and that information material to patentability must be disclosed regardless of how it was found. We also watch export-control and foreign-filing-license exposure when technical detail leaves our environment. None of this is optional. The efficiency is only worth having if the confidentiality and candor obligations are met first.

How do you validate accuracy, and who is accountable for what the tool returns?

A patent attorney is accountable, always. The tool produces candidates and a ranking; it does not produce a legal opinion. We treat its output as a well-informed reading list, not a conclusion.

On accuracy, I am careful with numbers. Published benchmarks report roughly ninety percent or better on certain retrieval tasks, but those figures describe specific datasets and settings, not a guarantee on your matter, and achieving high precision and high recall at the same time remains genuinely hard. So we validate in ways we can defend. We run known matters through the system and check whether it surfaces the art we already know is relevant. We keep the searcher's judgment in the loop rather than trusting a top-ten list. And because generative components can state something plausible and wrong, we never let the system summarize a reference in place of reading it. The honest framing for a client is that semantic search widens and speeds the net; attorney review is what turns a retrieved document into a patentability position.

What did adoption look like inside the firm, and what went wrong first?

The early skepticism was reasonable, and it improved the rollout. Our searchers had decades of pattern recognition, and their first concern was that a black box would push them toward false confidence. They were right to worry.

Our first mistake was treating the tool as a replacement for query craft rather than an addition to it. When people typed a thin description, they got thin results and concluded the system did not work. The fix was process, not software. We rebuilt intake so a search starts from a proper structured disclosure. We trained people to search limitation by limitation, and to read the driving passages rather than the relevance score. We also set a firm rule that the tool never closes a search: a person decides when coverage is sufficient. Adoption improved once the team saw it surfacing art their keyword strategies had missed, and once they trusted that it augmented their judgment instead of grading it.

What measurable outcomes have you seen, and how should a reader weight them?

The clearest change is time. Searches that used to take hours of iterative querying now return a strong candidate set in minutes, which lets the attorney spend the recovered time on analysis rather than retrieval. The second change is coverage. Because the search keys on meaning, it reaches art that keyword strategies miss, particularly disclosures that describe the same concept in different words and relevant non-patent literature. The third is timing: we can give a client a preliminary patentability read earlier, before large prosecution costs are committed.

I would weight these with some caution. Minutes-versus-hours describes the retrieval step, not the whole matter; the attorney review that follows still takes real time and is where the value sits. Broader coverage lowers the chance of a nasty surprise but does not promise you found everything, because no search does. And an earlier read is a better-informed starting point, not a verdict. In our experience the gains are real and repeatable, but they compound the professional's work rather than replace it.

What would you say to a skeptical peer, or to a client nervous about AI in their patent work?

To a peer, I would say the failure mode is over-trust, not the technology. Semantic search is very good at retrieval and ranking, and genuinely poor at judgment. If you let a relevance score stand in for reading the reference, or a generated summary stand in for the document, you will eventually miss something or assert something that is not there. Used as a retrieval aid under attorney review, it is defensible, and it is better than keyword search alone.

To a client, I am direct about scope. This strengthens a patentability search. It is not a freedom-to-operate or clearance opinion, which is a different analysis with different stakes. I also tell them exactly what we do to protect their invention, because a nervous client is usually worried about confidentiality, and that concern is legitimate. The reassurance is not that the tool is infallible. It is that a qualified attorney reviews everything material, the confidentiality controls are real, and the disclosure duties to the office are met.

Can you give a concrete example where it mattered, kept representative rather than a named client?

A representative one. We were assessing patentability for a mechanical-adjacent invention where the client and our own searchers were fairly confident the concept was clear. The keyword and classification search came back clean. When we ran the same disclosure through semantic search at the limitation level, the system ranked a non-patent reference highly: a conference paper that described the core mechanism using entirely different terminology, in a different field, which is exactly why it never surfaced in a keyword pass.

On review, the attorney judged it genuinely material to one independent claim. That changed the advice. Rather than file broadly and invite an office action or a later validity challenge, we narrowed the claim, drafted around the disclosure, and listed the reference. The point is not that the machine was clever. The point is that meaning-based retrieval reached a document that word-based retrieval structurally could not, and a person made the call about what it meant. That combination is the value.

What does this mean for the profession and the client relationship over the next few years?

I think the retrieval and first-read layers of patent work keep getting faster, and that resets where attorneys add value. Less of it will sit in finding documents and more in judgment: claim strategy, portfolio construction, and reading how a given examiner and art unit are likely to respond.

That is good for clients if firms are honest about it. The risk to the profession is a race to look automated, where speed is marketed and the accountable review quietly thins out. I do not want to compete there. For the client relationship, the change I care about is candor. We can show our search method, explain its coverage and its limits, and be specific about what a human checked. Clients are becoming sophisticated about AI, and they can tell the difference between a firm using these tools under real supervision and a firm outsourcing judgment to a model. Over time I expect the durable advantage to be trust: faster work, clearly bounded, with a named professional standing behind the result.