AW: Proposed draft of an AI policy

Norm,

Thank you for drafting the proposal. I fully appreciate your initiative and the effort to define rules on how to deal with AI in our group.

We are all flooded with non-human content today. At the same time, it becomes increasingly difficult to tell whether AI was used to refine, revise or create an existing text (ironically, the most reliable way to detect it is often the same technology). The more flawless a text seems, the higher the chances are that AI was used to generate it. Which is a curse, since professionalism is the very essence of what we strive for. Next, content creators themselves often cannot tell anymore how much AI is involved when translators, spell/content checkers, or web search are used. Finally, if we access existing sources, including scientific papers, we cannot tell how they came into existence either.

In short, AI is here to stay, its quality is improving rapidly, and it will become impossible to assess whether it was part of a creative process. The interesting practical questions for me would be whether and how its influence harms our daily work. I can only think of very basic guidelines, as follows:

• With regard to contributions, as the origin becomes fuzzier, we should continue to expect each contributor to be able to explain every single bit of their work.
• With regard to reviews and testing, new technologies should in no way cause extra effort.

Incidentally, both questions do not really revolve around AI. Thus, my (very personal) conclusion tends to be that the growing predominance of AI should in no way push us to become more permissive regarding contributions. A work shouldn’t expect to be consumed, let alone taken seriously, simply because it exists. We should continue to reject everything we are not fully convinced is an improvement over the status quo.

Having said this, I share your general concern that AI will have a vast effect on all of us. Pushing the limits of what is possible has always driven science and technology – for better (we might otherwise never have met electronically) and for worse. The open and important question is how much we want to be part of an ongoing and potentially destructive process. I think that a policy on AI is definitely something from which we will benefit, even if it does not establish fixed rules.

However we decide, I am sure it will be an improvement. Even if we decide to be as strict as possible, it will still help us to rethink our position later, once AI has evolved further.

Thanks,
Christian
________________________________________

Von: Norm Tovey-Walsh <norm@saxonica.com>
Gesendet: Montag, 25. Mai 2026 13:51
An: public-xslt-40@w3.org
Betreff: Proposed draft of an AI policy

Per my action, QT4CG-0165-01, here’s an attempt at an AI policy:

Contributions must not include content generated by large language models (LLMs) or other probabilistic tools (including tools which are often colloquially categorized as "AI").

This policy applies to the specifications, tests, issues, comments, pull requests, and any other contributions to QT4CG.

Although we acknowledge the many widely discussed ethical issues with LLMs, we will not refer to those issues as a primary justification for this policy. Instead, the specific justifications for the policy are practical.

For most members of the community group, our efforts are focused on document review. Reviewing technical specifications requires comparing what the specification says with an understanding of what was proposed by the group and determining if they are aligned. This is a time-consuming process.

LLMs are exceptionally good at producing large volumes of very /plausible/ text, but that plausibility is a statistical parlour trick: the LLM does not understand and cannot evaluate the output against any objective criteria.

Any process that increases the volume of material that has to be reviewed must be held to the highest standard of intellectual contribution.

A secondary practical concern is that LLMs have been trained on an enormous corpus of material, sometimes obtained in violation of copyright and other laws. Their output often includes verbatim samples of the training data. Incorporating such prose into the specifications may unknowingly violate licenses of copyrighted works.

For these practical reasons, QT4CG does not accept content that has been generated or substantially constructed using LLMs, "AI", or any similar probabilistic tools.

If you are uncertain whether the tool you wish to use, or the way in which you wish to use it, is covered by this policy, you are very welcome to discuss it with the CG's Chair.

                                        Be seeing you,
                                          norm

--
Norm Tovey-Walsh
CEO, Saxonica

Received on Wednesday, 27 May 2026 19:43:31 UTC