Journal

USPTO: The impact of the proliferation of artificial intelligence on prior art

On 29 July 2024, CIPA submitted a response to the USPTO’s request for comments on the ‘Impact of the Proliferation of Artificial Intelligence on Prior Art, the Knowledge of a Person Having Ordinary Skill in the Art, and Determinations of Patentability Made in View of the Foregoing’. The response did not represent CIPA’s formal position but provided various comments and observations relating to the issues identified in the request for comments. The response was collated by Simon Davies chair of the Computer Technology Committee.

CIPA’s response is primarily concerned with two aspects regarding artificial intelligence (‘AI’) and intellectual property:

• How any proposals might impact CIPA clients who file for and obtain US These inputs are flagged as ‘Comments’).
• How the questions raised in the survey might be answered under EP law rather than US law (c.f. question 14 below). CIPA and many other organisations regard a consistent approach to AI and IP across multiple jurisdictions as a desirable outcome (where feasible). These inputs are flagged as ‘Comparisons’.

As a general point, the survey appears to focus on a clear division between human authorship and non-human authorship. However, there is a strong likelihood of producing material having mixed authorship, both human and AI, in which the respective contributions cannot be separated. By way of example, a computer program may be a mix of human and AI-generated software, or a translation might be performed by a human working with a machine translation system.

At present, we think the patent system is generally coping with AI in various contexts. Any changes should only be made in respect of specific issues that are causing difficulties now, rather than addressing more speculative concerns. (Examples of topics which might potentially be considered for review now are Question 4c, relating to a presumption of validity for AI documents cited in patents, and Question 9a relating to ‘analogous art’.)

At present patent literature is a corpus of human-curated data that is valuable simply because it is human-curated rather than AI generated. It seems that it might be beneficial to retain the human-curated nature of patent literature.

A.  The impact of AI on prior art

1.  In what manner, if any, does 35 U.S.C. 102 presume or require that a prior art disclosure be authored and/or published by humans? In what manner, if any, does non-human authorship of a disclosure affect its availability as prior art under 35 U.S.C. 102?

Comparison: Under the EPC, prior art status is not affected by human/non-human authorship, but applies to any public disclosure prior to the priority date of a patent application (which is enabling as explained in more detail below).

(The EPC has a limited number of additional exceptions for non-prejudicial disclosures, for example in relation to certain international conferences; however, these do not appear to have specific relevance to AI).

If AI-generated prior art is differentiated, this leads to the question of whether we should differentiate other types of prior art – e.g. is a blog post to be handled in a different manner from a published patent application).

2.  What types of AI-generated disclosures, if any, would be pertinent to patentability determinations made by the USPTO? How are such disclosures currently being made available to the public? In what other ways, if any, should such disclosures be made available to the public?

Comparison: Under the EPC, any type of public AI disclosure represents prior art and so is generally pertinent to a determination of patentability. By way of example, such a disclosure might comprise:

• published written material – such as a patent application, academic paper, source code, social media postings, etc;
• publications in other media – such as audio recordings, video recordings, etc;
• use of a computer program or analysis of the use of such a computer program (without breaking any licensing restrictions associated with the computer program);
• physical models produced by an AI program such as by using a 3-D printer; and

(vi) spoken disclosures – conference presentations, lectures, etc.; for example, such a presentation may utilise AI-generated material

[See also our response to question 13 for further considerations on this topic.]

3.  If a party submits to the Office a printed publication or other evidence that the party knows was AI generated, should that party notify the USPTO of this fact, and if so, how? What duty, if any, should the party have to determine whether a disclosure was AI generated?

Comment: We would generally not support requiring a party making a submission to notify the USPTO if the submission is known to be AI-generated. Many parties may not know if a submission is AI generated (and hence not make a submission), since the party making the submission may be different from the author of the submission. It is unclear whether it would be helpful for the USPTO (or the public at large) to receive a notification indicating an unspecified subset of AI-generated material.

Comment: It may be useful to consider this question separately for patent publications and for non-patent literature. If the printed publication is a US patent publication item, then the US attorney who filed the US patent application may have carefully checked for any AI-generated content in the patent publication. However, this may not be the case for patent items filed in other jurisdictions.

Comparison: There is a separate issue under the EPC, in that it may be required for a party to provide evidence regarding a public disclosure, such as confirmation of the date, content, and public access. While such a showing is often straightforward, for example in relation to a cited patent application, establishing the precise, date, content and public access for an AI- generated citation may be more difficult in practice.

4.  Should an AI-generated disclosure be treated differently than a non-AI- generated disclosure for prior art purposes? For example:

a. Should the treatment of an AI-generated disclosure as prior art depend on the extent of human contribution to the AI-generated disclosure?

Comment: We disagree with this suggestion. For example, it would seem to require that whenever anyone (patent examiners, patent attorneys, etc) cites a patent document as prior art, there is an additional requirement to confirm that the patent document is not created using AI – but in many cases, such information may not be publicly available.

b.  How should the fact that an AI-generated disclosure could include incorrect information (e.g., hallucinations) affect its consideration as a prior art disclosure?

Comparison: The EPC generally supports free evaluation of evidence. If a document includes incorrect information, this might potentially lead to the document being disregarded as prior art for reasons such as: the incorrect information renders the overall disclosure unclear (and hence more limited in teaching); and/or the incorrect information may impact enablement (see (c) below) so that the document does not become part of the prior art.

Comment: It may be sensible for the weight given to evidence to vary according to the level of certainty that the evidence is human-curated. Non-patent literature from peer-reviewed journals may carry more weight than non-patent literature which is not peer-reviewed, even where both contain some AI-generated content.

One area in which the treatment of AI-generated prior art may need further consideration relates to proof of publication, for example in the following situation:

AI work generated by a user on 1 January Priority filing date of patent on 2 January

AI work shared (disclosed) by user on 3 January

The AI work would only qualify as prior art if providing the AI work to the user represents a public disclosure. In general terms, we think providing the AI work to the use would not form a public disclosure, although there may be additional circumstances to consider – for example, in a SaaS environment, a third-party operator might have access to AI disclosures (without confidentiality restriction).

c.  How does the fact that a disclosure is AI-generated impact other prior art considerations, such as operability, enablement, and public accessibility?

Comparison: The EPC already requires a prior disclosure to be (i) enabled, and (ii) publicly accessible. We think that these existing requirements provide a viable framework for handling AI- generated disclosures.

In existing EPC practice, the bar for establishing the above may vary according to the type of disclosure – for example, the bar for accepting a patent as publicly accessible is low, but the bar for accepting a product sale as publicly accessible is higher, and may require additional corroboration. This discretion in setting the bar for public disclosure may have increased relevance in a world with AI-generated material (the same might also apply for enablement).

The US presumption that any prior art cited in a patent is enabling might be revisited since there is a concern this has made it easier for certain parties to enforce flawed patents. The presumption that cited prior art is enabling puts a burden on applicants to provide evidence that speculative disclosures are non-enabling. A rapid increase in AI-generated publications might be addressed not by having notification requirements, but instead by reviewing the presumption that all prior art cited in a patent is enabling.

In particular, an AI-generated publication and any other publication could, if cited by an Examiner, be presumed to be an enabling disclosure unless it ‘appears not to be enabling on its face’. The burden would then shift to the applicant to provide evidence of non- enablement. We have an impression that for some US attorneys, labelling of publications as AI-generated may be enough to argue that the publication is ‘not enabling on its face’.

However, there is doubt as to whether a labelling requirement for all publications generated with AI assistance is workable.

5.  At what point, if ever, could the volume of AI-generated prior art be sufficient to create an undue barrier to the patentability of inventions? At what point, if ever, could the volume of AI-generated prior art be sufficient to detract from the public accessibility of prior art (i.e., if a PHOSITA exercising reasonable diligence may not be able to locate relevant disclosures)?

Comment: We think it remains to be seen as to whether AI-generated material will create a much greater volume of prior art. However, even if this does happen, it is also possible to use AI for searching. Consider for example searching prior art in a foreign language. For many years this has nominally been part of the patent examination process, but has been somewhat limited in practice. More recently however, with machine translation (which typically used various forms of AI), it is now much easier to review prior art in a foreign language. In a similar vein, AI might be used to find (pre-screen) citations of potential relevance, so that a set of documents to be reviewed by an Examiner remains realistic.

B.  The impact of AI on a PHOSITA

6.  Does the term ‘person’ in the PHOSITA assessment presume or require that the ‘person’ is a natural person, i.e., a human? How, if at all, does the availability of AI as a tool affect the level of skill of a PHOSITA as AI becomes more prevalent? For example, how does the availability of AI affect the analysis of the PHOSITA factors, such as the rapidity with which innovations are made and the sophistication of the technology?

Comment: We think PHOSITA should remain a human, but a human who is capable of using AI-generated material, such as for making routine developments to a product. Under UK practice, the PHOSITA may also compromise a team of humans having complementary skills, and again, such a team may use AI-generated material.

Since AI is a tool, just like CAD or wiki or a textbook, we don’t think an AI tool should change what is general knowledge (but of course, general knowledge will typically develop in time with the advance of technology). Since AI is a tool, it may lead a PHOSITA to analogous prior art in different fields, but the person skilled in the art would still have to make a determination to apply this art (this may have more relevance when using the EPO problem-solution approach).

7.  How, if at all, should the USPTO determine which AI tools are in common use and whether these tools are presumed to be known and used by a PHOSITA in a particular art?

Comment: This is a factual question to be determined by evidence in a relatively routine fashion. By way of example, if a particular AI system (e.g. for detecting bone fractures in X- ray images) is disclosed in a range of academic papers and/or some other widely available material (e.g. Wikipedia), then this would support a determination that a PHOSITA is familiar with the use of such an AI system.

The situation may potentially change if AI systems are developed to the extent that they become capable of creating inventive subject matter. However, such a development appears to go beyond the scope of the present consultation and we do not consider such a circumstance in our response.

8.  How, if at all, does the availability to a PHOSITA of AI as a tool impact:

a. Whether something is well-known or common knowledge in the art?

Comment: This situation does not appear too different from existing situations in which the PHOSITA may rely on specific tools – computers, medical imaging systems, and so on. The question of whether something is well-known is factual. If the use of an AI tool is routine in a particular field, then such a tool can likewise be considered a part of the common knowledge in the art.

b.  How a PHOSITA would understand the meaning of claim terms?

Comment: It is unclear to what extent a PHOSITA needs to understand claim terms – for example, in a prior art attack, the PHOSITA is assumed to want to develop technology according to motivation from existing art (rather than to produce a specific entity that falls within the scope of a claimed invention).

If enablement was contested on the basis that the PHOSITA did not understand the claim terms, then in theory the PHOSPHITA should access a dictionary, or discuss with a colleague, for example, potentially another PHOSITA from the AI field.

As for the previous question 8(a), we do not regard this point as a concern.

9.  In view of the availability to a PHOSITA of AI as a tool, how, if at all, is an obviousness determination affected, including when:

a. Determining whether art is analogous to the claimed invention, given AI’s ability to search across art fields? Does the ‘analogous’ art standard still make sense in view of AI’s capabilities?

Comment: This may be an area for further consideration, since as noted in the question, an AI system can search non-analogous art as readily as analogous art.

There may still be an issue as to the searching efficiency of the AI system. If a normal search across a single field yields say five citations to review, and the AI system likewise returns five citations across a wide range of fields, that is fine. However, if the AI system yields 100 citations across the wider range, this may start to become more difficult for human review.

The ‘analogous’ art standard may still apply even though AI systems can be used to find semantically similar art from different fields. This is because the PHOSITA is a human and the human is limited in their ability to transfer learning from one field to another.

b.  Determining whether there is a rationale to modify the prior art, including the example rationales suggested by KSR (MPEP 2143, subsection I) ( e.g., ‘obvious to try’) or the scientific principle or legal precedent rationales (MPEP 2144)?

No comment.

c.  Determining whether the modification yields predictable results with a reasonable expectation of success (e.g., how to evaluate the predictability of results in view of the stochasticity (or lack of predictability) of an AI system)?

Comment: The use of an AI system is now an obvious possibility in many situations – even though the results are not necessarily predictable. In order to show an inventive step, this may require the patentee to go beyond merely showing that the results are not fully predictable. For example, claim scope might be limited to some specifics of the implementation that produced the desired results.

d.  Evaluating objective indicia of obviousness or nonobviousness (e.g., commercial success, long felt but unsolved needs, failure of others, simultaneous invention, unexpected results, copying, etc.)?

Comment: In general terms, these indicia remain relevant. Depending on the specifics of a given case, the indicia may have to be applied having regard for factors such as the use of AI tools.

10.  How, if at all, does the recency of the information used to train an AI model or that ingested by an AI model impact the PHOSITA assessment when that assessment may focus on an earlier point in time ( e.g., the effective filing date of the claimed invention for an application examined under the First-Inventor-to-File provisions of the America Invents Act)?

Comparison: In respect of Internet disclosures, EPO practice is more cautious in the assessment of the date available to the public. For example, a date marked on a document may represent the date the document was completed, not when it was uploaded to the Internet.

In the context of generative AI, the EPO therefore would be more focused on the date at which information from the AI system becomes publicly available (rather than the recency of training data, which is more of an internal process).

This also leads into the issue set out in our response to Question 13 below.

11.  How, if at all, does the availability to a PHOSITA of AI as a tool impact the enablement determination under 35 U.S.C. 112(a)? Specifically, how does it impact the consideration of the In re Wands factors (MPEP 2164.01(a)) in ascertaining whether the experimentation required to enable the full scope of the claimed invention is reasonable or undue?

No comment.

C.  The implications of AI that could require updated examination guidance and/or legislative change

12.  What guidance from the USPTO on the impact of AI on prior art and on the knowledge of a PHOSITA, in connection with patentability determinations made by the Office, would be helpful?

No comment

13.  In addition to the considerations discussed above, in what other ways, if any, does the proliferation of AI impact patentability determinations made by the Office (e.g., under 35 U.S.C. 101, 102, 103, 112, etc.)?

Comment: One unanswered question is the extent to which the entire contents, including potential outputs, of an AI-generative system are disclosed when the AI generative system first becomes available online. Possibly this is dependent on the nature of the input to the AI system.

For example, if the input to the AI system is ‘make me a better mousetrap’, then the input has no inventive contribution. In this case, it could be argued that all relevant information for the ‘better mousetrap’ must be already present in the AI generative system and available to the public – which is the EPC requirement to be prior art. In this case, the output of the AI system could be considered as disclosed at the point of availability, i.e. prior to the query actually being entered.

On the other hand, there is a strong argument that the output of the AI system is only disclosed at the point of use of the AI generative system for inference in response to the prompt ‘make me a better mousetrap’. The AI generative system only generates an answer to the query ‘make me a better mousetrap’ at the point the question is given to the AI generative system. Because the process whereby the AI generative system generates the answer is probabilistic, the answer may vary when the same question is given at different times.

14.  Are there any laws or practices in other countries that effectively address any of the questions above? If so, please identify them and explain how they can be adapted to fit within the framework of US patent law.

See ‘Comparison’ entries in answer to preceding questions.

15.  Should title 35 of the US Code be amended to account for any of the considerations set forth in this notice, and if so, what specific amendments do you propose, and why?

No comment.