EPO Likely Breaking the Law Yet Again, This Time by Using Slop for Patents (to Lower Costs While Producing Monopolies That Cause Ruinous Lawsuits)
Nobody authorised this. The EPO is just a dictatorship of the corrupt - people who repeatedly buy votes and then do illegal things, exploiting diplomatic immunity like Benoît Battistelli and António Campinos have done for 15 years.
Yesterday we published EPO leaks related to a secret and most likely illegal policy.
It is clear that something must happen.
The serious crimes and corruption of the EPO aren't stopping. The only thing that's stopping is the media, which the EPO kept threatening (it did this to us several times) while doing illegal things behind closed doors. It's not limited to European software patents - i.e. patents which are both illegal and undesirable. The EPO does loads and loads of illegal things, in the process disgracing the EU and trying to cover things up, displacing real patent courts (seeking to replace them with illegal and unconstitutional kangaroo courts where Nokia staff is "judges"). "I hope the EPO can be cleaned up or shut down instead," one person told us. "It is very problematic that Microsoft killed Nokia and left the dying corpse with a pile of useless, invalid software patents because now Nokia management seems to be maneuvering to act as if the software patents have value in any other aspect than killing European software development across the board."
Today we release some commentary connected to the two leaks. Those leaks are imperative if the primary goal is to show how the EPO breaks rules, laws, and conventions. The Central Staff Committee wrote to staff:
"Leveraging AI" in classification to 90% by 2028
Dear Colleagues,
Since 2024, the EPO has considerably increased the proportion of classification tasks that no longer require human intellectual classification.
In the DG1 Town Hall meeting of 2 April 2025, the COO (and CTO) announced his intention to continue in this direction with a target of 90% according to the Strategic Plan Progress Dashboard 2028. The project bypasses classifiers and gérants who are for the vast majority dissatisfied with the quality of the classification provided with AI-autoclose.
This paper raises questions on the sustainability of the patent classification in the future and its usability by humans.
Consider what the EPO is doing in the name of "hey hi" (AI, buzzwords), seeking to replace professionals and experts with unproven algorithms without seeking any authorisation/approval from a legislative branch. Here's the paper:
Zentraler Personalausschuss
Central Staff Committee
Le Comité Central du PersonnelMunich, 11-04-2025
sc25024cpTowards humanless patent classification?
Since 2024, the EPO has considerably increased the proportion of classification tasks that no longer require human intellectual classification. In the DG1 Town Hall meeting of 2 April 2025, the COO (and CTO) announced his intention to continue in this direction with a target of 90% according to the Strategic Plan Progress Dashboard 2028. The project bypasses classifiers and gérants who are for the vast majority dissatisfied with the quality of the classification provided with AI-autoclose. This paper raises questions on the sustainability of the patent classification in the future and its usability by humans.
“Leveraging AI” in classification
Beginning of March 2025, the EPO adopted an AI Policy without any consultation of staff and their representation. To this date, the document remains classified “for internal use” only. In the AC meeting of 19 March, the Office presented in its Quarterly Activities Report containing a slide “summarizing” the EPO AI Policy. The full content of the document itself was however not presented. One slide of the presentation shortly addressed AI-supported classification as follows:
AI supported classification: Increase in “leverage” AI KPI from 15% in Jan 2024 to 47% in Dec 2024 Continue to ensure high-quality, efficient, and accurate AI supported classification while optimising human interventionThis does not reflect what the EPO is actually implementing.
The Key Performance Indicator (KPI) “Leverage AI” was mentioned for the first time in the third quarter of 2024 in CA/105/24 without any clear definition. The definition can only be found in the Strategic Plan 2028 Progress Dashboard. The KPI measures the proportion of classification tasks that no longer require human intellectual classification (thanks to AI support), compared to the total number of classification tasks. The target is 90% by 2028. Instead of a mere “optimisation of human intervention”, it is a (quasi) supplanting of humans which is taking place.
How is the supplanting implemented so far?
In June 2024, DG1 directors presented a document celebrating a “New Era of Classification” and rolling out AI-autoclose. The principle is the following: Allocated classes by human third-parties (other patent offices or non-authorized examiners) are compared against AI suggestions. If the classification groups match and the number of classification groups are matched (regardless of the probability in practice), the classes will then be confirmed as final by AI and the circulation of the file for the purpose of classification is closed without any intervention of the classifiers and gérants authorised in the relevant fields. In practice, for the purpose of circulation, unauthorised classifiers often propose only one classification group.
1
The document “New Era of Classification” came as a surprise to DG1 staff. From that date, authorised classifiers and gérants saw the batch of files they have to review shrink. There is therefore significantly reduced fresh data classified by authorised classifiers that can be used for the purpose of AI training. At the same time, quality checks (“OQC”) were reduced in the same proportion.
Concerns from classifiers and gérants
While the EPO continues to roll out AI-automated classification, the staff representation conducted a poll to gather feedback from the gérants directly affected by the change.
From 12 February to 13 March 2025, the staff representation gathered 247 responses from gérants with the following results:
- 67,61% replied they are dissatisfied with the quality of classification provided by AI- autoclose,
- 19,03% replied that they are unaware of the AI-autoclose procedure,
- 13,36% replied that they are satisfied with the quality of classification provided by AI- autoclose.
An exemplary feedback from gérants is that out of 44 quality control (“OQC”) samples of AI- autoclosed circulations, 26 were satisfactory, but 18 were not. In other words, 40% were
_____
1 In the right section, the decision diamond should read: "The list contains a symbol" (i.e. has more than 0 members)
considered wrong to the point that the document was not retrievable according to the applicable criterion.
If the final decision to close circulation is made by AI, the process cannot be considered to be “human centric”, even if the original input of a class is provided by a human non-authorised classifier. We observe that when a non-authorised classifier proposes a classification group, they not only tend to propose only one classification group, they also tend to propose a hierarchically higher (i.e. broader) classification group, for lack of knowledge. AI typically generates a rather huge inventory of suggestions (up to 100 classification groups) thereby improving chances for a match between the two. As explained above, one match is enough to close circulation, if only one classification group is proposed by the non-authorised classifier.
“AI is no expert but AI is consistent”?/
On 2 April 2025, the DG1 COO invited staff to a Town Hall meeting. He explained that feedback loops are important and tried to reassure staff by declaring that “AI is not going to rule the Office” and that “AI is there to support us and not to replace us”.
In the Q&A, the COO was confronted with questions on the AI-based automation of classification currently rolled out. Indeed, while the EPO AI Policy presents as key principle a “human-centric approach” promising that final decisions will be taken by humans at the EPO, the general perception is that classifiers and gérants become supplanted and bypassed.
The COO replied that a classification decision is not a final decision, and that “human-centric” does not mean that a human will check everything AI does”. He also added that “AI is no expert but AI is consistent” and that “classification is only necessary for the routing of files and for the search”.
In our view, consistency does not necessarily imply accuracy, notably in areas with special rules of classification. An accurate classification is essential for search examiners making use of a single class or combination of classes in their search queries. Even if AI consistently classifies according to the probabilities and rules it deems correct and not where it should according to authorised classifiers and gérants, then the classification cannot be considered as directly usable by search examiners for the purpose of retrieving information.
Does the COO consider that patent classification is meant to be used only for AI-automated search in the future?
Routing of files and search are not the only use cases of patent classification. Patent classification is necessary for retrieving patent information and performing statistics, e.g. for users of the EPO’s Espacenet tool. Users of the patent system search by using patent classification to identify which technical areas are dynamic in terms of innovation, which fields are worth investing into and which fields might be subject to litigation.
Scientific literature exposes risks
In the DG1 Town Hall meeting of 2 April 2025, the COO announced his intention to continue in the direction of AI-automated classification. The plan is that by 2028, 90% of classification tasks will no longer require human intervention. We are not aware of any risk assessment study which took place to support this way forward.
A possible risk in this context could be a so called “model autophagy disorder“ (“MAD”), which was exposed for instance in 2023 by a research group of the Cornell University. The paper strongly warned against refeeding “synthetic” data generated by AI for training and advocated for sufficient fresh real data:
“Our primary conclusion across all scenarios is that without enough fresh real data in each generation of an autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease.”
Similarly, the journal nature explained in its article of 11 December 2024 that:
“The problem with synthetic data is that recursive loops might entrench falsehoods, magnify misconceptions and generally degrade the quality of learning. A 2023 study coined the phrase Model Autophagy Disorder to describe how an AI model might “go MAD” in this way. A face- generating AI model trained in part on synthetic data, for example, started to draw faces embedded with strange hash markings.”
Will 10% of human intellectual classification be sufficient to prevent AI-automated patent classification from “model collapse” or “going MAD”?
Conclusion
The EPO has relied over the last decades on the expertise developed by its classifiers and gérants who analysed patent (and non-patent) literature documents to classify them and followed technology trends to redesign classification schemes.
Since 2024, the Office has started to supplant review tasks of authorized classifiers and gérants by AI. Which added-value has the EPO for patent classification compared to any external provider?
Furthermore, it cannot be assessed, how AI-automated patent classification with only 10% human intellectual involvement will be sustainable in the future and remain usable by humans (e.g. search examiners, patent information officials or users of the patent system).
EPO staff, their knowledge and experience should be adequately involved in any development, risk assessment and project for AI-automation for the benefit of all.
The Central Staff Committee
It's not hard to see where this is going. After classification they'll try to leverage slop to automate more things irrespective of compliance, legality, quality etc. The EPO strives to just "print money" without any accountability. Who will pay for this? Everybody. █