We do have community guidelines for this forum and for the conference (Community Guidelines - SWIB Forum) but we do not have an explicit policy for the use of the newest generation of AI methods (in particular what I like to call “Large AI” which is mostly generative AI).
There are various areas and aspects to the organization of a conference and forum where the use of AI-based tools and the conditions for that use can be put up for discussion. There is the generation of abstracts by authors, for example, or the generation of reviews for submissions (we are obviously not going to automate that!), and then of course there is the content of proposals itself.
In a sense, SWIB has always been about AI since semantic technologies are a subfield of AI (symbolic AI) which has been somewhat eclipsed recently by the current hype surrounding deep learning (which is a subfield of machine learning which in turn is another subfield of AI: subsymbolic AI). We did add machine learning to the list of topics some years back because combining the two approaches seemed worthwhile exploring. Also note that SWIB has always been about open source technologies, so SWIB has never given and will not give a platform to commercial solutions, and that includes generative AI tools.
Large AI is putting an enormous strain both on the climate and on society and that is obviously at odds with SWIB’s stance on sustainability. However, approaches to make that footprint smaller and to work with AI models that are as open source as possible, can be run locally, and adhere to ethical conditions could very well fall into the range of SWIB topics in the future.
In any case, we would want to be as transparent as possible concerning our AI policy for SWIB and that starts with the process how to get to an explicit AI policy – we want to involve you, the community. So get in touch with us and tell us what you think and what would be important to you!
Looking forward to a lively discussion here in the forum!
Example AI policies
We will be collecting some other AI policies here and made this post a wiki so that all logged-in users can contribute.
Thanks, Argie, for starting this discussion. I would very much welcome it if SWIB were to establish a policy re. “AI”. The FluConf policy, to which you linked, could provide an excellent starting point. I’d suggest two additions: Since many institutions in our field are hit hard by reckless LLM crawlers, that should be included explicitely among the negative impacts. And among the strongly discouraged content, perhaps ai-generated images on slides could be mentioned, too.
The larger issue - how the forced proliferation of genAI hurts science, education and libraries in general - should also be dealt with in a forum such as SWIB. Perhaps that could be the subject for a SWIB keynote in future years?
I’m so glad to see SWIB thinking seriously about this! Straw-persons for consideration:
Any submission documenting use of commercial generative AI should receive an immediate rejection without benefit of peer review. This includes text, image, and code generators. Obviously this stricture should be enshrined in policy and documented in submitter instructions.
Slides, text, and (if any?) abstracts/papers should not use generative AI in any way.
Machine learning that is not generative AI is in-bounds, but submitters should address questions of ethics (e.g. data sources for training, minimal computing, use of undercompensated labor) in their abstracts and presentations. Reviewers should be alerted to downgrade scores on submissions with questionable ethics.
I would be delighted to see that keynote, Joachim. I’d even give it, if I could do so remotely, but you can likely do better – I’d suggest Alex Hanna or Emily Bender.
Thanks for the straw-persons @dsalo! I think it’s good to have proposed starting points like these.
I’ve previously given two SWIB talks which involved what you describe as “commercial generative AI”. The first one was a lightning talk at SWIB23 (Berlin) titled "Extracting metadata from grey literature using large language models” (see recording) - GPT-3 was used for the initial experiments. A follow-up talk that I gave together with Pierre Beauguitte at SWIB24 (online) was Automating metadata extraction and cataloguing: experiences from the National Libraries of Norway and Finland - those experiments involved several fine-tuned local LLMs (Llama, Gemma, Mistral, StableLM, Qwen etc.). Whether those are considered commercial is debatable; but they are created and published mainly by for-profit corporations, even if they are provided free of charge with relatively liberal “open weight” licenses.
@dsalo are you saying that these talks should have been rejected without review? Can you elaborate why? Or if not, why should they be exempt from this straw-person rule?
I have added a table to the original post to collect other examples of AI policies. Also, the original post is now a wiki to be edited by all logged-in users. We invite you to add to the list.
The discussion seems to have stalled, but let’s try to restart, because the topic is important and timely. With the train wreck of the AI industry heading towards a (possible) crash, it would be really important for SWIB to stake its position.
I added the new Fedora AI contribution policy to the table in the OP. I think it is very clearly formulated and has some good ideas on what should be permitted, what must/should be disclosed and what to avoid entirely. Of course Fedora is a Linux distribution and not a conference, so not everything applies in this context. But what I like about this policy is:
the use of MAY, MUST, SHOULD i.e. RFC2119 terms; very nerdy but very clear
Accountability: you are responsible, duh!
Transparency: drawing a line between “significant contribution” that MUST be disclosed, and otherwise SHOULD disclose “when it might be useful”
also the practical examples of disclosure in git commit messages; maybe something similar could be done in the SWIB submission template - have one or more fields asking about the use of AI
that AI tools MUST NOT be used to make substantive or subjective judgments on contributions
I think that SWIB should say that AI tools MUST NOT be used to review contributions (they are very short abstracts anyway, so I don’t think there is a big temptation to do this currently, but it’s important to state the principles)
I added the CEUR-WS GenAI policy as well. It is only concerned with use of generative AI for writing workshop papers themselves, not about the research or other activities described in those papers. But I think it is very clear and contains nice examples of acceptable and non-acceptable uses of genAI for drafting papers. This could be a basis for SWIB policy on drafting abstracts and possibly also the presentation slides or other presented materials.
Some points I noted:
generation of large sections of text (e.g. paragraphs) is not acceptable
small scale sentence polishing is OK, but e.g. generation of prose from short snippets is not acceptable
using genAI to generate diagrams or illustrations is not acceptable, unless the paper’s core topic concerns these methods
any genAI use must be disclosed and categorized according to a 13-class taxonomy