I had the pleasure to be invited to the AI & IP FORUM organised by Forum Heidelberg.
Numerous speakers, from academia and the profession, industry and private practice, were taking part and the whole day was interesting as it gave a good insight to the ins and outs of AI’s influence on the world of IP in general and patents in particular.
The topics discussed were the following
- AI and the evolving landscape of Intellectual Property
- Patenting AI case-law update: US, UK, EP/DE
- Panel discussion: Patenting AI technologies – opportunities and challenges
- Panel discussion: AI’s impact on corporate practices in contracts, collaborations and trade secrets
- IP implications of the AI Act/data privacy
- Panel discussion: Use of AI for IP management
- Panel Discussion: Present issues and the future of AI and IP (navigating the future)
The present recollection is personal and does not claim to be exhaustive or correct in all points as it results from notes taken on the fly. It will also not cover all the topics mentioned above.
AI and the evolving landscape of Intellectual Property
A lady from the MPI reported on Index report on AI established by the Stanford University. Its key take aways set the picture.
All industries are touched by AI, the costs for AI go down, the investments in AI are increasing as does the number of applications in the domain. The “hallucinations” which can be generated by AI systems were also mentioned.
The question of protection of AI and the concurring problems of the protection offered to AI inventions were also broached. It appears clearly that an AI system cannot invent on its own and that an inventor has to be a natural person, cf. Dabus/Thaler.
Some board decisions were also presented: T 1425/21, T 1193/23, commented on the present blog and T 206/22.
The USPTO seems to have recently published a set of Guidelines on patenting AI inventions.
AlphaFold, see IPKat, on protein folding was mentioned as an apparent key achievement in AI. AlphaFold has been subject to a patent application WO2022/112248=EP4 200 854, for which the EPO established the ISR and the IPER. The examination is ongoing at the EPO.
The EU-AI act and its current status were also presented. The whole difficultiy appears to lie in the Code of Pratice which is not yet finalised. There is apparently still a lot of work to be done. Explanations about the risk-based approach in the EU-AI Act were convincing. The Info Soc. Directive and the GDPR were also mentioned.
A quote from Broussard M. and others was given by in which it was regretted that the term AI had in fact been chosen incorrectly. Todays AI “is merely complex and beautiful mathematics”.
Patenting AI case-law update and Panel discussion on Patenting AI technologies – opportunities and challenges
EP/DE
A DE/EPO lady representative considered that grants in the domain increased, but that the boards of appeal of the EPO appeared to be too restrictive in the matter as they stick on the technicity in the application, cf. Comvik.
Some case law was also mentioned: T 1669/21, T 161/18, T 1191/19, T 702/20, T 1998/22, T 1952/21, T1425/21.
As far as training of AI systems is concerned, a decision from the Hamburg Regional Court was quoted (310 O 227/23) was mentioned.
There was a need for a common understanding and a balance to be found between the divulgation of the AI training data and the requirement of sufficiency. Some other forms of protection could be needed for AI inventions.
A distinction was made between Applied AI which is non-technical and Core AI which is technical. The influence of OSI models was also mentioned.
UK
The Aerotel decision was mentioned. According this decision it is not enough to have “any machine” to escape the exclusion. This decision is in clear contrast with EPO case law, cf. T 154/04.
There will be a discussion in a British Court towards the end of this month and it could well be that the UK takes a different path than the continent when it comes to patenting AI. The discussion when too quickly to be able to take notes.
US
Alice was still the leading decision, but a new tendency is emerging.
The USPTO has issued inventorship guidance and examples for AI-inventions, according to which an claim for an AI is not purely abstract. Some US patents were also mentioned, inter alia, US 9 106 609 and US 6 307 576.
There are a series of law suits gong on in the USA, with a clear emphasis on copyright problems when using AI.
Panel Discussion
In the following panel discussion, the chairman of board 3.5.06 reminded that there CII are only excluded as such and that it was not possible to make a distinction between CII and AI. The distinction between technical/non-technical was not only to the disadvantage of applicants. Any technical effect has to be achieved in the real world. G 1/19 is the leading decision in the field.
The interveners from the profession or the industry insisted on a more realistic approach and the technical effect should not be the only aspect, but also what the AI achieves. What happens if AI is a mere layer in a system? Training with AI should actually be patentable as well.
Patent offices have the tendency to only grant AI patents which have an impact on the real world. AI should also not be considered as a mere black box.
As AI inventions are rather short lived, the necessity of a protection for 20 years was questioned. A shorter protection could be envisaged, or even a mere declaratory system, where examination only occurs in case of litigation.
Panel discussion: AI’s impact on corporate practices in contracts, collaborations and trade secrets
That AI can also have a dear impact in contracts between an AI developer and its customers.
One important point is the respect of the GDPR.
It is also important to decide were the AI data is stored. On the client or server or on the server of the AI provider?
One aspect which appeared important is how to deal with the staff so that is does not migrate to competitors with its knowledge. One way is to treat staff well, so that do not have the desire to migrate. Another way is to block any attempt to download information, especially before a staff member leaves the company.
One thing is however clear: a staff member cannot take with him information on any form of support, but it is impossible to control what it takes with him in his head.
Once it appears that trade secrets have migrated, it is in general too late. The information is lost to third parties and even legal action will not allow to recover the lost information.
Use of AI in IP management
This was also a very important part of the conference.
Interestingly, it appears that AI is used in industries having in-house patent departments, but not in private practice.
Classical AI systems were becoming better and better, but they were still not exempt of hallucinations. It appears necessary that the output of an AI system is checked by a human being.
In a number of companies, the use of AI in IP management appears quite advanced. It was considered that the use of AI increased the productivity of staff in IP departments.
It is used not only for drafting applications and claims, but also for replying to office actions.
It appeared however necessary to properly train people in the use of AI. The AI as such is important, but also the way of questioning the AI system, the “prompts”, need careful and continuous training. In some companies people were encouraged when using AI to even develop their own code.
One industry representative went even as far as to allege that claim drafting was merely “repetitive”, so that AI could well be used for drafting claims.
That AI can be used for scheduling jobs does not come as a surprise, but confidentiality in using a general purpose AI can be a matter of concern.
In spite of using AI, it was also stressed that time should be given to people to get acquainted by the system. AI should not be used by newcomers as they have to learn from their mistakes.
Present issues and the future of AI and IP
The general view was that patents are useful in protecting AI, but other, for instance shorter, forms of protection could well be envisaged. It could even go as far as issuing defensive publications instead of patents. Trade secrets are also a possibility, especially for smaller actors in the field.
Training data are valuable data, and it does not appear to exist any compensation for the divulgation of training data.
One big challenge remains the inventorship of AI inventions.
It was also criticised that the approach to AI inventions is not uniform among examiners at the EPO. The number of clarity objections appears to be on the rise and the standard of disclosure required by the EPO is not sufficiently clear.
FTO and enforcement of AI patents are also not straightforward.
More surveys like that of the Stanford University were needed.
As far as the link between AI and copyright is concerned, it seems that governments are rather dithering.
Comments
My thanks go to Jean-Claude Alexandre Ho from Forum Heidelberg for inviting me to this event.
My conclusion is that AI is actually a hype. We previously had big hypes like with “big data” or biotech. We may well see that the hype on AI falls like the preceding hypes. At least it is one way of getting money from governments.
To be cynical, one could consider AI as being clearly artificial and by no means intelligent. For the time being it just spits out what it has been told to do.
AI is certainly useful when it comes to automatising repetitive tasks, like for example analysing X-rays. Here exists a direct link with the real world, cf. G 1/19. Without any relation with the real world and proper sufficiency of disclosure, e.g. by disclosing the training data, it does not appear possible to get patents at the EPO.
If claim drafting is as repetitive as I have heard from an industry representative, then the logo of the EPO should be changed immediately.
As it is not possible to know how the data are aggregated in the system, an AI will, at least presently, not be in a position of defining the common general knowledge going back to the effective date of a claim.
It should also not be forgotten that the output of an AI system is not better than its training data and the subsequently entered data. There is also no guarantee on how the result is achieved, hence the necessity of protection like the EU-AI act, which should come as soon as possible.
This act is criticised on both sides of the Atlantic, but in view of the risks inherent to the use of AI, regulation appears to be an absolute necessity. What is technically possible is not necessarily good for society at large. I remind here about the EU regulations prohibiting the use of human embryos.
If, as we have heard during the conference that in IP, the output of an AI system is to be checked by a human being, what about outside IP?
Creating sui generis rights for AI appears not a good or viable solution, I take for example the protection of semi-conductor topographies. After a big hype in the mid-eighties, ending with the Washington Treaty on Intellectual Property in Respect of Integrated Circuits (IPIC) under the auspices of WIPO and the Directive 87/54/EEC, has anybody ever heard what happened to it?
Last but not least, I have heard from the developer of SIRI, that AI needs much more energy than a search with available search tools like “Bing” or the like. No wonder that Microsoft wants to refurbish a decommissioned nuclear power plant to supply energy to its AI servers. Meta want to even to build a series of nucler power plants to supply energy to its AI servers. In time of global Earth warming, do we need even more energy to be used in order to nurture tools which for lots of people are no more than a black box.
In view of the basic laws of thermodynamics the cooling of AI servers necessarily goes on a par with heating of the atmosphere. It is a fallacy to think that only the reduction of CO2 emissions by using nuclear power plants will reduce global warming.
Comments
13 replies on “AI & IP FORUM – Munich – 08.07.2025”
Thank you for this recap. I fully agree with the conclusion that IA is, for now, more a hype than an actual game changer, except for bad quality stuff that can be produced much more quickly.
I’ve one question and one remark :
1/ I am not sure to understand the following link between the repetitive nature of claim drafting and the EPO logo “If claim drafting is as repetitive as I have heard from an industry representative, then the logo of the EPO should be changed immediately.”
Maybe a joke that I missed ?
2/ Regarding “It should also not be forgotten that the output of an AI system is not better than its training data and the subsequently entered data.” I mostly agree with a small caveat.
Currently, it is true that most of the models are not better than the sum of their training data (and ability to call some specialized algorithm when a specific well-known task is identified).
However, this year we can see “sparks” of general intelligence, or in other words some abilities of some AI models to be able to infere and resolve some trully new problems.
You can read about ARC-AGI, which is a benchmark of tasks that are easy — or at least doable — for human, but quite hard or even impossible for AI. This benchark relies on simple tasks implying basic underlying logic ; these task are doable for a human with a bit of logic (maybe not all humans, but every human with a bachelor in science for example). But until this year, even the best AI were unable to perform better than 5% of the tasks. And, since this year, the best OpenAI models are able to perform near the human limit — but at a very high cost.
There exist another, much harder, benchmark which is still doable for human (even if it’s more challenging) and in which most of AI models fails (<5%), but the best models of openAI with unlimited token (i.e. computer power) can do 25% of the tasks.
It is not human-like intelligence (intelligence means in this case the ability to perform an unknown task, something where the AI models are trully bad, and humans are between mediocre and good). But this year it is an actual breaktrough which could be relevant to follows.
@ cloth,
Thanks for your comments.
My comment about the change of the EPO logo was with a bit of tongue in cheek. If something is repetitive, the result should in principle be the same. The EPO logo has been designed by the Kilkenny Design Workshops, Ireland. It represents a stylized fingerprint, which is the universally recognized mark of individuality and identification. If claim drafting is repetitive, the result should be in principle the same. Where is then the individuality and identification?
I have no doubts that we are only at the beginning of AI and the domain will improve with time. The only question is at what cost, not only in terms of money, but also in terms of computational power and hence energy.
Thanks also for drawing my attention to the ARC-AGI benchmark. It looks very interesting and worth following.
I know that I will not make friends, but societal control appears to be a necessity. Not everything doable should be done or left in the hands of private companies.
Question: are you from Turin? There is a famous cloth there! 😉.
Thank you for the explanation regarding the logo. It is actually fun, and very subtile (I’ve never realized that this logo was a fingerprint). It reminds me that lot of people outside of the patent world, especially at the border of patent practice, do not perceive the nuances regarding claims and their actual scope. Which is why they think that claim drafting is repetitive.
Even us, as patent attorneys, can struggle with the actual scope of claims without read them with deep attention.
Regarding AI, I 100% (ok, maybe 99% but it does not make any difference here) with your opinion. First, I do not believe that AI can (at least now) be even 1% as good as a proper and well-formed patent engineer.
And I think, like you, than it is not a sustainable technology, at least now. The energy consumption is gigantic, absolutely not perennial especially considering our needs for electrifiying and decarbonating our transports and industries).
And the damages that it will (and already does) inflict to the cognition of people using it (especially youngsters, but adults too at a smaller not not negligible scale) is, imho, maybe worst than the environmental problems.
I believe that by controlling the use of AI, we can avoid to sink into a society where AI majors control our minds. In a way deeper, more proprietary way than social networks and big media plateforms already does (with the political consequences that we can observe nowadays).
I’m not from Turin but from France ! It’s just another pseudo on the internet.
Dear “cloth”,
I am aware that for people having come more recently in the patent business, the EPO logo does mean a lot. That’s the advantage of having been in patents since 1971 and at the EPO from its opening.
When even IP departments make clear that a human control of an AI output is necessary, then even your 1% looks very optimistic. Actually, AI can only be trusted when one knows how it has been trained, and even senn it can produce “hallucinations”.
The lack of sustainability is one thing which should also be regulated, at least in Europe. Merely heating student homes with heat dissipated by existing servers (heard on France Inter), will not be enough. We should also not forget that nuclear power plants are also driven by fossil fuels, and the waste problem is by no means solved. Water to cool them is also becoming scarce.
Controlling the general use of data and protecting users is an absolute necessity. That AI needs as well to be controlled is beyond doubt.
It is to be hoped that the EU commission will resist the pressure of the present (and future) president of the US to dismantle all regulations in the IP area and the introduction of the AI act. IP regulation and AI sovereignty should not be sacrificed to the benefit of German industry at large and especially German car manufacturers.
The question about the cloth and Turin was cheek in the tongue. You have certainly heard about “le saint suaire de Turin”.
I agree that AI will shake some things up, but will not be the game-changer some people would like it to be.
In fact, it has already shaken things up in some fields. Medical imagery analysis was mentioned, but there is one more relevant to patents: translations. When I started as a patent attorney, machine translations (which are all output by an AI in a broad sense) were good. Now they are very good or even excellent, at least in the most common language pairs. But who thinks “I am using an AI” when they click the “Patent Translate” button on Espacenet?
When it comes to automatizing tedious, low-value intellectual tasks, like summarizing documents, looking for information in a document, or coming up with a canned response to an e-mail, I am convinced AI will help, if it isn’t already.
When it comes to claim drafting, I am very, very far from convinced. That use case assumes that inventors/clients provide the model with a lot of well-written text to go on. In my experience, that almost never happens. Instead, you have to talk (gasp!) to the inventor/client, first to understand what they meant with the (often few, sometimes cryptic) sentences they actually wrote, then to understand what they want and why. I don’t see that happening anytime soon. And if it ever does, we will have bigger problems than finding new jobs for patent attorneys anyway. The same goes for many “high-level”, high-value intellectual tasks.
Besides, what are clients paying for: work product (patent application, reply to an office action, whatever), or personalized and relevant advice informed by experience? I dare say that some of the clients’ money currently goes to the latter, and will continue to do so in the future.
Also, as has been mentioned, AI uses an astonishing (and ever-increasing) amount of energy. This is all fine and good as long as the world energy supply continues to increase. The problem is, the opposite will happen soon. According to some projections, the top 15 oil and gas producing countries will _all_ see their output sharply decrease between now and 2050. When that happens, what will have priority: train the latest iteration of ChatGPT, or produce the chemicals on which so many people depend to stay healthy, or even to eat? (Per Our World in Data, almost half of the world population would not be here without synthetic fertilizers.)
Of course, everything I say is probably tainted by my own desire, which is to continue working as a patent attorney as long as I can…
Regarding you sentence : “When it comes to automatizing tedious, low-value intellectual tasks, like summarizing documents, looking for information in a document, or coming up with a canned response to an e-mail, I am convinced AI will help, if it isn’t already.”
Are you sure that it can still be relevant when the document is a prior art and pretty technical? From my personal practice, either the document is pretty simple (some basic mechanics for example) and finding the relevant information is quite easy (so no need for an AI to do my job), or the document is quite technically hard (for example high-level physics or telecommunications) and the AI will absolutely not be able to dive into the subtilities of the document.
For emails & co, the AI can be usefull, but only if the mail is well-written and not cryptic.
However, your remark regarding translation is true. DeepL and the likes are super good nowadays, with little supervision. The other thing were AI is not that bad is to transcript visio meetings, which can replace note-taking, letting us more available to interact more and takes notes with schemes.
“Are you sure that it can still be relevant when the document is a prior art and pretty technical?”
Why not? Large language models can sift through textual information. If the document is a 900+ pages A1 publication with a ton of boilerplate (which I have seen, unfortunately), a model can point to the potentially relevant passages, and perhaps summarize the teaching of the document on a high level, faster than I will ever be able to.
To be clear, we agree an LLM can’t replace carefully reading a highly relevant document, and understanding its disclosure completely. But I think it will be another tool in the toolset, like keyword search or OCR, for doing the job faster.
Fully agree for transcripts — I don’t personally use that functionality, but I am amazed at how good subtitles can be on YouTube, and I assume the underlying technology is the same.
My personal interrogation is not “why not”. Obviously, in a hypothetical future made of super-human arificial intelligence, a such AI would be able to read a big document much faster than I would dreamed of.
But is it actually able to do that, now?
From your comment, I understand that proper doc-summarizing AI can be used as a research engine within a document, in a much more flexible manner. Which is usefull.
I just fear that it could “blunt” my own instinct of guessing, in a document, where the relevant information could be. Which is still usefull in the case of new documents submitted for example during OP, or during negociations.
For YT subtitles, I frequently used them and I don’t find them that accurate. Especially when the lexicon is highly technical (in maths, physics, or even in woodworking or other forms of craftmanship). In this case, for live-generated subtitles, the quality today is not enough for professional use (for my own standards). It can merely help me to understand some accents that I struggle to decrypt.
But some friends told me that specific AI tools are much much better for this task than YT subtitles, so I’m sure I’ll try, one day. At least for curiosity.
As to the reliability and learning speed of AI, It occurred to me to ask Co-Pilot which Patent Offices in the world, besides the EPO, require the originally filed description before grant to be conformed to the (narrower) allowable claim.
Its answer: US, CN, JP.
I replied that I thought that answer wrong. Co-Pilot replied instantaneously, to this effect:
Thank you so much, David. You are correct and I was wrong. The EPO is the only one.
Huh?!? Perhaps somebody else can now ask Co-Pilot my original question, again. I wonder, has it learned already from the exchange it has just had with me?
Its politeness cannot be faulted. But what if the human enquirer does not already know the answer to its question, and accepts from the AI, and confidently acts upon, its initial, seemingly authoritative answer?
I mean, if none of us can ever know how the AI gets to its answer, will it ever be as reliable and trustworthy as, say, Wikipedia? What if, going forward, students believe Co-Pilot in preference to their teachers/mentors/supervisors? Will AI eliminate the need for human teachers/educators/parents?
Dear Max Drei,
Your example shows clearly how little reliable an AI output can be. If you know the answer, you do not need an AI. If you use an AI, you cannot trust the answer.
Where is the improvement? That is the question!
The improvement is in the pockets of those setting up the various AI, and not for the users who do not have a clue on how the output is generated.
To all commenters,
I never thought that my blog on the conference on AI and IP would bring such interesting comments.
Thanks to all having taken the bother to comment.
I would like to sum up the comments by saying that AI will certainly get better and better, but it will at the same time blunt our knowledge.
Without even talking about AI, just look at mental arithmetic, map reading, spelling words correctly, applying the correct grammatical rules, all represent basic knowledge which slowly disappears.
Nothing to say about technical progress, but without some basic knowledge, we might have difficulties in dealing with such simple tasks, and AI, especially without any control, might transform us in zombies incapable of thinking by ourselves.
Any bias in AI should be nipped in the bud. The control envisaged in Europe, e.g. the future AI act and the existing GDPR, should not be given up, even if Trump wants it.
In lots of domains, AI will be a help, but at what cost in electronic chips and energy? The latest NVidia chips costs apparently 70 000 US$, and with one you are by no means done.
I also doubt that AI will ever replace a human being with its capability of nuancing its reply to an arduous problem.
In IP, AI can become a useful tool, provided it is properly trained, but I do not think that we will see tomorrow, or even the day after, AI replacing human examiners, human representatives and human judges, although some in the upper spheres of patent offices or government offices would like it.
Mr Thomas,
I share your concerns.
Another very serious concern is raised by a recent study of the MIT Media Lab, reported at https://time.com/7295195/ai-chatgpt-google-learning-school/
« ChatGPT May Be Eroding Critical Thinking Skills
Does ChatGPT harm critical thinking abilities? A new study from researchers at MIT’s Media Lab has returned some concerning results.
The study divided 54 subjects—18 to 39 year-olds from the Boston area—into three groups, and asked them to write several SAT essays using OpenAI’s ChatGPT, Google’s search engine, and nothing at all, respectively. Researchers used an EEG to record the writers’ brain activity across 32 regions, and found that of the three groups, ChatGPT users had the lowest brain engagement and “consistently underperformed at neural, linguistic, and behavioral levels.” Over the course of several months, ChatGPT users got lazier with each subsequent essay, often resorting to copy-and-paste by the end of the study. »
Mr Hagel,
Thanks for your comment.
What the MIT Media Lab has discovered does not surprise me whatsoever.
I will never forget the first lesson of my electrotechnics main professor at the engineering school. He first said that what is important in life is to apply the law of least effort, the difference being to apply intelligently.
His second say was: if you work somewhere and you find that you deserve a better job held or given to someone else, then don’t complain, just look for another job. It was a very short lesson, but I have never forgotten it.
What the MIT lab study shows that the avid ChatGPT users are not applying the law of least effort in an intelligent way. This applies to most of the users of AI.
AI has very good sides, and those should be promoted, but it should not be used to foster laziness.
I have warned my granddaughter to keep her fingers off AI when doing her homework. The day of the exam, she will not have any AI to help. I hope the penny dropped.