As far as the ethical arguments are concerned, I don't feel it is appropriate for the working group to attempt to control how contributors do their work: we are only concerned with the quality of the inputs that they submit
In most cases, when people say this kind of thing, they don't really believe that it's inappropriate to set ethical limits; they just think that *their* ethical limits are so obvious that they can be taken for granted.
Are there no limits, other than the letter of the law, on what we would accept from contributors? If a contributor were paid to make their contributions, with funding coming from a white nationalist organization, would you still argue that only the quality of their contributions matters? What if a contributor publicly refused to engage with any female or LGBT members of the CG, ignoring their questions and suggestions, but otherwise made decent contributions?
Of course these are extreme and unlikely scenarios (I mean, female members of the CG? Craziness!). I'm just asking us all to consider whether we genuinely see *no* ethical boundaries for our work, or whether it's actually that we think some ethical boundaries are "obvious" while others are "extreme". I like to think that the group would be unwilling to accept contributions directly funded by white nationalists, regardless of the technical quality of those contributions.
assistive technologies are greatly benefiting from AI and are helping to level the playing field for participation
Well. Let me pull out the "as a disabled person" card, and put it on the table next to the "as a speaker of a minority language" card.
LLMs are being used to create the illusion of accessibility, but disabled people have also been speaking quite passionately about how LLM-generated assistance is often misleading and of poor quality. Sure, writing alt-text and making accessible websites is extra work, but it's work that should be done - and it's work that we should be learning how to do, not putting into a pile of stuff that doesn't matter enough for human effort.
Using LLMs in this way offloads the burden of accessibility, shifting it onto disabled individuals. Can all disabled people afford to use LLMs in this way? Not a chance. Is this a massive reduplication of effort? Of course it is. Make something accessible, and it's accessible to anyone who uses it. If you instead require people to use LLMs, the work of making accessible happens over and over again, for each individual.
AI systems in general have been quietly multiplying within public services, used to decide who is worthy of medical intervention (and bringing all the usual biases to bear, with sometimes disastrous results), used to decide who is defrauding the state when claiming benefits (with some devastating consequences for massive numbers of innocent people). The roots of this kind of statistical intervention in our public lives are in eugenics. That isn't hyperbole - the historical record supports the claim.
Claiming accessibility as a selling point for LLMs ignores the actual experience of disabled people, as well as the ideological background of the technology, and the ways it is actually being used to the detriment of disabled people.
As for language access, there's quite a bit of research showing that LLMs are simply deepening the divide between English (and a few other dominant languages) and the world's thousands of other languages. LLMs ignore information from most languages when answering questions. Say you ask a question in German, and there are no German training data relating to your query - an LLM will mostly default to answering you using information from English sources, ignoring anything written, say, in Hindi or Welsh. Most of us who aren't monoglot Anglophones are having our perspectives and our languages erased from the dominant digital record. And when we interact with an LLM, the information we get will be more relevant to our cultural context if we're an Anglophone American than if we're a German, or a Welsh-speaker, or whatever.
So does using LLMs enhance access for people whose first language isn't English? Maybe - but at the cost of forcing that person to discard their cultural context. At the cost of making their own language or languages less and less relevant. Given the digital invisibility of most languages, the problem of lack of training data will likely occur much sooner for anything other than English, Chinese, maybe a few other languages. The LLM isn't going to keep up with German; it's certainly not going to keep up with Welsh. Sooner rather than later, the LLM you're relying on for linguistic help will start to degrade in quality for your native language, and you'll be forced into using English instead.
When I need to write about technical subjects in Welsh, it's hard work. Welsh is my first language, but most of my technical life is lived in English, so I don't have Welsh vocabulary and idioms easily available. I understand the frustration and the horrible slowness of expression. But using LLMs to make that easier on myself would be kicking the problem down the road to other Welsh speakers, and to the language as a whole. If, instead, I engage with actual language resources, I am making it worthwhile for authors to keep writing those resources, and for publishers to keep publishing them. If I engage with other Welsh speakers and ask them for help and advice, I'm maintaining my own linguistic community, and also increasing all of our knowledge by encouraging us to think about expanding and standardizing our vocabulary.
Using LLMs for language help is fine if you aren't worried about the quality of the output, or the cultural biases, or the maintenance of non-English languages, or supporting true language experts and authors, or the growing divide between those who can afford to access LLMs and those who can't, or any of the other ethical and practical issues. But the gains in individual accessibility are, I'd argue, massively overshadowed by the loss of linguistic capacity and diversity over time, and the homogenization of cultural and linguistic perspective.
Essentially, I think arguments that praise LLMs for increasing accessibility are generally just a convenient way to leverage the genuine struggles of disabled people and non-Anglophones in support of a technology which is actively causing significant harm to both groups.
BTW
****************************************************
Dr. Bethan Tovey-Walsh
linguacelta.com
Golygydd | Editor geirfan.cymru
Croeso i chi ysgrifennu ataf yn y Gymraeg
On 29 May 2026, at 23:33, Michael Kay <mike@saxonica.com> wrote:
I personally have been a slow adopter of the LLM flavour of AI, largely because I'm generally a late adopter of most technologies; I like to wait until the hype has settled.
And I've only really used it as a "smart Google" - to get answers to technical questions where conventional approaches weren't successful. In that context, I've found the answers extremely impressive, especially as my questions are typically on very obscure topics.
Also, of course, I use whatever coding assistants are embedded in my IDE (IntelliJ). There too the technology has clearly been advancing rapidly and often delivers very impressive suggestions for code improvement (which I never accept without careful manual checking).
Our key mission is to produce high quality specifications, and to achieve that, I think the key factor is that we can't accept any proposal to change the specs unless it is practically feasible to check it by hand. I'm sure, for example, that an LLM could do a pretty good job of of re-arranging the XQuery spec in a more logical order, but there's no way we would be able to check that it hadn't inadvertently changed anything. Our rationale for rejecting such a proposal would be our inability to check it, not the way that it was produced.
I might note that even without AI, there are proposals where human verifiability is hard to achieve: this is particularly true of the grammar, where throughout the development of XQuery we have had a handful of experts who have understood how to apply technology to the task of defining a bug-free grammar. I think we value these tools because they are deterministic and follow provably correct algorithms, which is a very far cry from the kind of reasoning used by LLMs.
As far as the ethical arguments are concerned, I don't feel it is appropriate for the working group to attempt to control how contributors do their work: we are only concerned with the quality of the inputs that they submit (we also have to consider whether we have the resources to handle the quantity of inputs - something that is affecting many open source GitHub projects). In any case, the boundary between AI tools and non-AI tools is pretty fuzzy, especially when it comes to things like spellling and grammar correction. We can't really go telling people that they are allowed to use a grammar checker so long as it's a stupid one. Christian's point about non-native speakers is also relevant: assistive technologies are greatly benefiting from AI and are helping to level the playing field for participation, a trend which we can only welcome.
Norm and Bethan have marshalled some very persuasive arguments that should influence us as individual contributors. I'm conscious that the LLM services we are being offered are being sold at way below their true cost in order to get us all addicted. I also have fears for the future, of how LLMs will be trained when the internet becomes dominated by slop. There are clear arguments for political regulation of the technology.
But in the meantime, if someone produces a PR that aims to improve our grammatical consistency, or to correct errors in example code, or perhaps to identify changes we have made to the 3.1 specs that we haven't highlighted in the changes appendix, I think we should judge it on its face value (by careful review of the proposal content), and not by the amount of help that the author got from software tools, or the number of kWh that those tools consumed in the process.
Michael Kay