Patent offices do not examine inventions; they examine patent applications. AI is materially changing the relationship between the two. That could be a positive development; but then again, perhaps not...

By Sharaz Gill

Thus far, significant investment has been required to turn an invention into a patent application. Someone must investigate the prior art, formulate a claim strategy and prepare a supporting specification. The application must then be translated, filed and prosecuted wherever protection is sought. Because invoices have a certain restraining effect, all this serves as a de facto invention filter: not every idea becomes an application and not every application is pursued everywhere.

All this could be about to change. Commercial AI tools already claim to identify patentable features, search prior art, generate claims, produce credible first drafts and draft responses to patent office actions. A company faced with 10 variations of a technical concept may now file all of them. Researchers using AI to explore thousands of possible solutions may generate applications alongside each one. The increase in applications may therefore be considerably greater than the increase in patentable inventions.

The consequences extend beyond heavier workloads. AI could change how examination is paid for, which filing routes applicants choose and where institutional power resides. The offices and organisations that survive these changes will also have to preserve the evidence and independent judgement on which their authority depends.

Patent offices: who pays for examination?

More applications might appear to equate to more income for patent offices. But patent office accounts are less intuitive than that. Many offices set filing, search and examination fees below cost, helping individual inventors, universities and smaller businesses to enter the system. Much of the deficit is recovered through renewal fees for commercially valuable patents.

This model is effectively a wager that enough applications will mature into patents that are worth renewing. AI could put pressure on both sides of this wager. A larger volume of applications will increase examination costs. If more marginal inventions are filed, a smaller proportion may reach grant. Separately, even patents that survive examination may have too little commercial value to justify years of renewal fees. Neither outcome is inevitable, but offices cannot assume that application growth will produce a correspondingly larger stream of renewal income.

The obvious response is to change who pays for examination and when. If AI reduces drafting costs, offices could raise filing, search and examination fees so that each application bears more of its own cost.

Fees could also follow complexity. An application with hundreds of claims imposes a different burden from one directed concisely to a single invention. Existing charges for excess claims, pages and lack of unity already point the way.

Patent offices are not profit-making businesses. They can recover costs, but they cannot impose scarcity rents merely to suppress demand. Nor are indiscriminately higher fees a satisfactory filter. A fee set to deter a multinational will exclude private inventors and universities first. General subsidies may therefore give way to targeted subsidies, with larger and more frequent users bearing the full cost of their work.

Renewal fees will still be important. Their increasing price forces owners to decide periodically whether a patent is worth preserving. Without that discipline, cheap AI-generated applications could become patents that remain in force long after their owners have any commercial use for them.

The fee-funded office would therefore likely survive, but in a different balance. More cost would be recovered during examination, subsidies would be targeted and renewal fees would clear away rights that no longer justified their place in the system.

Patent offices join the arms race

It is also likely that patent offices will not stand still while applicants deploy AI. The same technology that generates claims and responses can help examiners to search prior art, compare claim features and map documents against them, leaving more time for matters requiring technical or legal judgement.

Applicants will respond in kind. Their systems will analyse cited documents, propose amendments and generate arguments explaining why the invention is novel and inventive. Left to themselves, two tireless machines might spend eternity explaining why the other has misunderstood Claim 1. The procedural challenge is how to bring that exchange to a decision.

Fortunately, procedure has a power that AI lacks: it can say, “Enough!” At the EPO, later amendments require the Examining Division’s consent. The EPO can charge additional search fees for lack of unity, prevent a shift to unsearched subject matter and refuse unmanageable auxiliary requests. It can then summon the applicant to oral proceedings and refuse the application.

The USPTO reaches a similar destination by a recognisably American route: finality, followed by another fee. A second office action may normally be made final. Substantive amendments then are not admitted as of right, leaving the applicant to appeal, abandon the application or pay for a request for continued examination. Restriction requirements can force different inventions into separately funded divisional applications, while excess-claim fees discourage sprawling claim sets.

The greater pressure may arise across patent families. AI will make continuations and divisional applications cheaper to devise. Offices may face more applications, each complying with the rules and requiring its own decision. Fees and procedure will need to reflect that work.

Patent offices: a question of class

AI is often described as a democratising technology. Computing budgets, however, are stubbornly aristocratic. The EPO, the USPTO and the major Asian offices possess enormous collections of patent data, substantial expertise and the resources to develop sophisticated examination tools. Many smaller national offices do not.

Those offices may have to acquire commercial products, rely on searches performed elsewhere or use systems developed through WIPO or regional organisations. Shared tools can reduce duplicated searching and allow the focus to be placed on the application of national law.

The risk is that cooperation becomes dependence. An examiner may not know which databases a foreign or commercial system searches well, which languages it handles poorly or which assumptions shape its rankings. Prior art from smaller countries may receive less attention. The standards of the largest offices could spread beyond their jurisdictions, not through treaty but through infrastructure.

The legal decision would still sit with the national examiner. In practice, however, an office without the resources to test or replace the system may find its freedom to apply national law constrained by choices made elsewhere. Patent sovereignty could remain intact on paper while much of the substantive analysis is controlled abroad.

Shared tools could give smaller offices capabilities that they cannot develop alone. The institutional questions are who controls that infrastructure, on what terms offices can use it and whether they retain a practical alternative. WIPO and regional granting organisations could help to ensure that cooperation strengthens national capacity rather than replacing it with dependence.

WIPO: when the world is a bit too much

WIPO occupies a different position from patent offices because it does not grant patents. The PCT is, among other things, a highly elaborate way of buying time: an applicant can file one international application, obtain an international search and defer the much greater expense of entering national and regional systems. The decision on whether to grant still rests with those authorities.

WIPO’s finances are nevertheless tied closely to patent filings. In its budget for 2026/27, WIPO estimates that 96% of its income will come from fees generated by its global intellectual property systems, with PCT fees alone accounting for 76%. In practical terms, patent applicants finance not only the administration of the PCT but also much of WIPO’s norm-setting, development and international cooperation work.

For an institution that is financed by traffic, more traffic sounds like good news. Formalities checking, classification, translation and publication are also amenable to automation, so WIPO might process more applications without a corresponding increase in cost. As a non-profit organisation, however, it could not simply accumulate the surplus; pressure would grow to reduce its fees or expand its services.

There is also a less comfortable possibility. The PCT’s optionality is most valuable when the applicant is uncertain about where protection will be needed. Repeat filers may already know the comparatively small group of commercially and strategically important jurisdictions that they intend to pursue. This is particularly apparent in the SEP world.

An AI-powered global filing platform could aggregate national requirements and practices, prepare forms and translations, coordinate local representatives and recommend a filing strategy according to commercial objectives. A sophisticated user could affordably assemble a tailored bundle of national and regional filings without passing through a single international application.

For those users, an AI platform would compete with the convenience of a familiar filing route. The PCT would retain its international search and the important advantage of deferring national costs and decisions. Applicants that were uncertain about markets or commercial value would still find this option useful. The competitive question is how much predictable filers would continue to pay for those benefits once coordinating their chosen filings became cheaper.

WIPO could respond by itself becoming the platform rather than allowing itself to be bypassed. Alongside the PCT where deferral is valuable, it could explore a coordinated direct-filing service where it is not, with translations, deadline monitoring and links to national offices. The advantages would be trust, neutrality and direct relationships with member states.

Its wider coordinating role might also grow. The smaller-office problem suggests a need for standards for AI-assisted searching, interoperable databases and transparent examination tools. WIPO could help to organise that infrastructure. AI may therefore weaken the filing route that finances WIPO while making international coordination more important than ever. The institution established to manage the complexity of national filings may find itself competing with technology that manages this complexity directly.

The EPO and UPC: till revocation do us part

The EPO faces a related challenge, but it has a stronger defence than the PCT route. It offers more than coordination and deferred decisions: it conducts a single substantive examination that can secure protection across Europe.

For applicants targeting only a few EPC states, the direct-filing platform described above could nevertheless make separate national applications a practical alternative. The calculation would depend on the cost and speed of the relevant national procedures, as well as the value of a common examination.

This strategy might offer more than lower cost. National applications could be tailored to local law and would produce independently granted rights. The loss of one would not necessarily threaten the others. AI could make fragmentation attractive by making it considerably less burdensome. This would be consistent with greater dependence on shared examination tools: filing strategies could become more dispersed even as the infrastructure used to assess applications became more concentrated.

The EPO would still retain important advantages: one examination, consistent claims and access to the unitary patent and the UPC. In fact, the UPC may become its strongest defence. AI can offer many administrative benefits of centralised filing, but it cannot reproduce the legal effect of a single court that can decide on infringement and validity across several countries.

The choice may therefore become more sharply divided. An applicant interested in only a few jurisdictions may prefer direct national filings, while one seeking broader protection and centralised enforcement may continue through the EPO. For SEP owners, the UPC offers substantial cross-border leverage from a single action, but also the risk of central revocation. National filing offers the opposite bargain: independently enforceable rights, with greater resilience but without a single multinational remedy.

The number of countries needed to justify the regional route could therefore rise, affecting EPO income. If more applicants chose national filings, the UPC’s future caseload could be smaller than it would otherwise have been, even as the court became more strategically important for those opting for centralised enforcement. This would also have implications for the court’s funding. The value of its jurisdiction and the volume of patents entering it need not move in lockstep.

Patent institutions: custodians of the AI record

Even as AI challenges existing filing routes, it may create institutional duties that do not currently exist. One concerns evidence that might not be needed for years. Inventive step is assessed at the priority date. The patent lawyers of 2037 may therefore have to become software archaeologists, reconstructing what a computational chemist, telecommunications engineer or materials scientist could reasonably have achieved with AI in 2027.

Models are updated, withdrawn and replaced. Their capabilities may change without a change of product name. Access may depend on a subscription, location or computing resource, and results on the prompts, tools and data supplied. A later demonstration that a model can find a solution does not establish that the skilled person would ordinarily have done so. Failure to reproduce a result 10 years later may reveal equally little about an earlier version.

This compounds the danger of hindsight. Once an invention is known, a later and more capable model may readily find a route to it; but this does not show that the invention was obvious at its priority date. The question remains what the skilled person could reasonably have done with the tools then ordinarily available, without knowledge of the answer.

Patent institutions may well need to preserve a new kind of historical evidence: model versions, documentation, benchmarks, access conditions and evidence of industry use. Complete preservation of proprietary systems may be impossible, but independent capability assessments would be more reliable than later recollections or marketing claims. WIPO, the EPO and other major offices would be natural custodians. Establishing the state of the art may no longer be enough; they may also have to reconstruct the state of the artificial mind through which it could have been explored.

That is a duty to preserve the evidence needed to apply the legal standard, rather than to let an archived model define the standard itself. The same distinction matters when institutions use AI to make decisions in the present.

Responsibility: the institutions’ last stand

Technology has always been generous in offering to solve the problems that it helped to create. AI may assist offices in absorbing more applications, allowing smaller players to share resources and making separate filing routes easier to manage. But the changes in funding, institutional power and evidential responsibility all lead to a further question: who is answerable for the decisions that the system produces?

As regular readers will know, Jorge Luis Borges is a literary hero of mine. Although he wrote fiction, he has proved disconcertingly prescient about the world of AI. In his short story “Tlön, Uqbar, Orbis Tertius”, an imaginary world is constructed in extraordinary detail and recorded in a set of encyclopaedias. Its languages, sciences, philosophies and history are all invented. As knowledge of this fabricated world spreads, it proves so complete and attractive that it begins to displace the untidy reality in which it was created. Humanity gradually comes to prefer the invented world to its own.

For patent institutions, the attraction of a coherent artificial account of the world may prove equally difficult to resist.

AI will not merely search prior art and classify applications. It will construct comprehensive models of technical knowledge, mapping relationships between inventions, interpreting claims and assessing whether an invention appears novel or obvious. Similar systems will be used by applicants, offices, courts, licensing organisations and investors.

Initially, these systems will be tools for organising and interpreting the technical record. Eventually, their representation of that record may become the framework through which everybody sees it.

An AI-generated representation of the prior art is not the prior art itself. A machine's interpretation of a claim is not the claim. A patentability score is not a legal conclusion. Each is a model assembled from specific sources, classifications and assumptions.

The danger extends beyond a smaller office’s dependence on an outside provider. If applicants, examiners and courts rely on the same or similar models, each may appear to reach a conclusion independently while inheriting the same defects from a common source. Successive stages of scrutiny could then reinforce an error instead of exposing it. Apparent consensus could be little more than shared dependence.

Preserving a historical record of AI capabilities is one way to maintain the distance between a model and the world that it describes. Institutions must also preserve the ability to question present conclusions. AI-generated assessments must be open to inspection and challenge. Examiners must be able to scrutinise why documents were selected or excluded. Courts must distinguish evidence from an algorithmic assessment of it. Alternative methods must remain possible.

This is not an argument for preserving human employment. It is rather one about institutional legitimacy. Patent examination is an exercise of technical and legal judgement, which determines whether an applicant should receive the right to prevent others from using an invention. That decision does not become legitimate merely because a machine can produce it quickly, consistently and at scale.

AI may make inventions, applications, searches, arguments and even plausible decisions more abundant. However, it cannot make responsibility more abundant. This may be the last scarce resource: a person or institution that is willing to look beyond the model and answer for the world it is helping to create.

Sharaz Gill is Head of Portfolio Management at Sisvel

The opinions expressed within this article are the author’s and do not necessarily reflect the views of Sisvel. The content is for informational purposes and should not be taken as legal advice.

Other articles in this series:

Do we really want transparency in SEP licensing?

AI and the future of (F)RAND negotiations

Patent law is falling behind the AI invention revolution

AI invention’s patent paradox