Ethical LLMs
See: What is an Ethical LLM?, my first attempt at this, which also has some additional context
See also: The sources linked below.
What Makes For an Ethical LLM?
This is by no means an exhaustive list, so if you think I missed anything important, have a good example you think bears inclusion, or know of an “ethical LLM” project you think should have been included, let me know on Mastodon!
Dimensions of an Ethical LLM
- Source data
- Licensing and consent
- Negative examples abound (see: OpenAI tries to steal Scarlett Johansson’s voice)
- “Vegan” models, a Simon Willison term he uses “to describe machine learning models that have been trained in a way that avoids using unlicensed, copyrighted data.”
- Effect on the Web (see: exhibit A)
- Especially with respect to model updates (see: this reply and the linked post)
- Licensing and consent
- Environmental effects (water usage, which gets discussed a lot, but more importantly, IMO, the ways it contributes to Climate Change, and things like Musk’s illegal power plant construction)
- Power gradients, who benefits
- Not as much of an issue in the poll’s hypothetical, but in real life all of the AI companies seem to be run by, basically, the worst people in Tech (sleazy Sam Altman, Dario Amodei who doesn’t seem much better, and actual Nazi Elon Musk)
- Ideological valence of the tools themselves
- See: AI as a Fascist Artifact
- This piece* makes a reference to the concepts of holistic and prescriptive technologies, from Ursula Franklin’s classic Engineering text The Real World of Technology
- *also via @mayintoronto
- Labor practices
- Especially with respect to the kind of underpaid and dangerous data cleaning that American tech companies have tended to outsource to the Global South (see DAIR’s work in this area, or this representative reply)
- See also: Against the protection of stocking frames.
- Privacy, self hosting
- A given in the thread’s hypothetical, but worth mentioning
- How it’s trained, guardrails
- Not really discussed in that thread very much, but covers things like grok’s Nazi and White Supremacist functions (c.f. Thaura’s core principles)
- Sycophancy, ability to say “I don’t know”
So… Are There Any?
I don’t know, but here are a few potential options. I haven’t done a deep dive on any of them yet, I mostly know that they are pitching themselves as ethical alternatives.
Thaura
Thaura is a very interesting one. An actual product, not a standalone model. Built around a non US-produced model (GLM-4.5 Air, at the time of this writing). Started by a couple of brothers from Syria who were sick of using tech that contributed to e.g. the genocide in Palestine. They seem to be claiming they’ve fixed the issues around sycophancy, which I find exciting and intriguing. They also lay out explicitly what “ethical” means to them in a set of core principles that seem more grounded in reality than any similar statements I’ve seen coming out of OpenAI or Anthropic. Shout out to my brother for tipping me off to this one!
Apertus
Apertus, an LLM being offered by the Swiss government under the Apache license. Seems to be based off of Common Crawl? I think the branding looks a bit fascy, but I’m open to the idea that they’re not the only ones who get to use V’s instead of U’s. They too have a document outlining their principles which by my skimming are decidedly non-fascist, so let’s just call the branding different tastes, maybe.
Vintage Vegan
These models aren’t trying to solve the same kinds of ethical issues as Thaura or Apertus. They’re focused primarily on the issue of training data licensing (AKA these are attempts at “vegan” models) and they’re also experimental research into what we can learn by making specific and time-bound language models, in this case Victorian-era text only. Hence the “vintage” models moniker. Both used modern “non-vegan” LLMs in their creation, so there is a bootstrap/contamination issue that hasn’t yet been resolved.
Sources
- An interesting Mastodon poll-and-thread that was the original impetus for finally writing this piece