Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.

@made @firefoxwebdevs There's already lots of work for on-device ML: https://searchfox.org/firefox-main/search?q=toolkit%2Fcomponents%2Fml

Integrating models into a finalized product with the wide spectrum of end-user devices is tricky though, so it has to be done with care.

@firefoxwebdevs As worded, and if we can trust Mozilla, then the acceptable answer should be No for these reasons: ML is not AI, and on-device means nothing is sent out of the device. In exchange you get free translation. Win.

BUT… there’s the trust issue now.

And what we REALLY need is not an AI kill switch but more of a “data transfer/phone-home kill switch”, almost like a firewall, where we know the browser is not taking any data and sending it to a device we don’t control ourselves.

@firefoxwebdevs not trying to split hairs here but how are the ML models doing translation when they are not LLMs? Maybe they are not as huge as ChatGPT but they are transformers probably with all that entails.

(A Killswitch should of course kill all ML/AI functionality and people could then reactivate certain specific features of they want to, it's really not that hard. Just cause you consider a feature"better" than others does not override consent practices.)

Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.

They're not LLMs. They're trained on open data.

Should translation be disabled if the AI 'kill switch' is active?

@firefoxwebdevs It would also be compelling if a team at Mozilla were dedicated to building the best browser translation add-on on the market, for all browsers. To promote the power of add-ons and, at the same time, the Mozilla brand.

@tante it’s a SLM

@zzt I posted this poll after a meeting where we discussed the design of the kill switch, and there was uncertainty around translations. I want to make sure the community's voice is represented in these discussions.

@mdavis folks want to disable 'AI' for more reasons than privacy. Privacy is important of course, but folks are also concerned about the training data, and energy used for the training.

@joepie91 I think a lot of people in the replies would consider this sneaky. It's a tricky UX problem. But yes, granular control needs to be part of the solution, along with a kill switch.

@firefoxwebdevs But if the ML/AI training work is processing on the device and not is shared off device, and it is in support of a feature like translating a page (which should be prompted/selectable) then what’s the issue? You can say no and nothing happens. Or you can say yes and the worse that happens is you chew up some local power on your laptop or PC. Or are you saying that even though the translation happens on the device, the RESULT of that training data is sent back out?

Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.

They're not LLMs. They're trained on open data.

Should translation be disabled if the AI 'kill switch' is active?

@joepie91 agreed.

@firefoxwebdevs we're not in those meetings so we don't know what all is actually included within the AI module suite, or even if that has been fully defined internally at this point, so of course there won't be a clean consensus externally from us on what "it" is and if it should be included or excluded, as it's up to our interpretation.

@firefoxwebdevs Like many others, I have a metric ton of thoughts on this topic. I might even try writing something to consolidate them.

In the meantime: I don't consider the translation models to be part of the major AI hype I loathe so much. Machine translation of language has been happening for a long time and has proven largely useful, and it lacks the stink of desperation which so many of the generative applications of recent times carry.

While I'm already thinking about it: even the name "AI kill switch" feels bad to think about. I know that "AI" is the buzzword that gets upper management giddy and which the untrained public is now used to hearing, but the fact of the matter is that if you can't "sell" a feature without appealing to buzzwords, your feature wasn't worth the time and effort put into it.

@eckes for that usage pattern the results would probably be even worse with more fabrications. So what are we even doing here?

@mdavis I believe it's a moral stance due to how the models were produced.

@angelfeast @twifkak No, I don't think so. It says this (with a takedown compliance process posted afterward)...

License

These data are released under this licensing scheme: PD

We do not own any of the text from which these data has been extracted.
We license the actual packaging of these parallel data under the Creative Commons CC0 license ("no rights reserved").

@firefoxwebdevs Hookay… then this is less about a local feature or data sharing and more about an overall “Made with AI” concern where nothing related to AI *at*all*ever* taints the user’s browser, in or out. In that case, if the user turns on the AI kill switch, it should totally kill anything having to do with AI for those who take that position.

That’s an issue with these polls — too much undisclosed nuance to be able to answer properly.

@firefoxwebdevs I can only speak for myself of course, but I'm someone who is strongly opposed to sneaky approaches, like hiding things in submenus or requiring people to go back later to disable new things, for example. And I'm also strongly opposed to basically everything in the current generation of "AI" (LLMs, GenAI, etc.) - but personally I wouldn't consider this sneaky, as it's immediately visible that there's a second choice to make, at the exact moment you disable "AI".

Of course if that stops being the case and the second option gets hidden behind an "Advanced..." button or foldout for example, it would be sneaky. But in the way it's shown in my mockup, I would consider it fine as it's both proactively presented and immediately actionable.

(I do still think that exploitative "AI" things should be opt-in rather than opt-out, but it doesn't seem like that's within the scope of options that will be considered by Mozilla, so I'm reasoning within the assumption of an opt-out mechanism here)

Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.

They're not LLMs. They're trained on open data.

Should translation be disabled if the AI 'kill switch' is active?