'Meredith,' some guys ask, 'why won't you shove AI into Signal?'

Because we love privacy, and we love you, and this shit is predictable and unacceptable. Use Signal ❤️

‘It’s terrifying’: WhatsApp AI helper mistakenly shares user’s number

Chatbot tries to change subject after serving up unrelated user’s mobile to man asking for rail firm helpline

The Guardian
@Mer__edith It is almost funny how users judge the answers of an LLM when they call that LLM to account for mistakes made. Those answers are also the result of the same principle: creating plausible answers instead of the truth.
@Petermid @Mer__edith I'm not a native English speaker... but shouldn't it be probable rather than plausible?
@Petermid @Mer__edith ok it's probably both 😁
@devnull69 @Petermid @Mer__edith no, it's plausible, glaubwürdig. Probable, wahrscheinlich, really doesn't enter into it.
@Mer__edith AIs behaving like most consultants and managers have been behaving for the past decades.
@Mer__edith “A quick online search shows the phone number mistakenly provided by Meta AI is both publicly available and shares the same first five digits as the TransPennine Express customer service number.”
That's about half the number. So the bot went and gave a number that had about 1 chance in 100000 of being right. Nice.
@Mer__edith
*insert Bluesky reply guy gibberish contradicting Signal's president here* 😉
@draken @Mer__edith that was a beautiful internet moment. I'm thankful for it.
@bovaz @Mer__edith
I noticed he hasn't made a single post since he started that spectacular shit show.
@Mer__edith Note that the writing in this article is not great: so much anthropomorphism in phrases like "chatbot tries to change subject" etc. It does no such thing! And the author obscures that we can't tell at all whether the number was in any way retrieved or just hallucinated.
@danstowell @Mer__edith this drove me crazy as well - reporting on the synthetic text generated by the LLM in question like it would add anything useful to the table about what the actual problem was, the deployment of the model in the first place.
@mario_angst_sci @danstowell @Mer__edith Exactly! I@came here to say this as well. Anthropomorphizing this chatbot gives too much credit to Meta. This thing is just hurling random junk at useres. Period.
@Mer__edith
“Addressing puppy strangling across all our shelters is an ongoing area of research. In addition to informing users that PuppyStrangler can strangle puppies, we’re continuously working to improve the strangling of puppies through a variety of methods.”
—some CEO probably
@Mer__edith 🙏🏼

@bengo @Mer__edith What is really bothering is this article, trying to impose intentions on the chatbot, when it just generates probabilistic words. "chatbots try to negotiate their way out", "deflect attention from their mistakes", "tried to change the subject", "admitting when it didn’t know", "it admitted", "was caught out lying"...

None of this is true, or makes sense!

@tdelmas @bengo @Mer__edith

The urge to anthropomorphise stuff is really hard to resist. It's even worse when people actually think it's warranted.

@veronica @tdelmas @bengo @Mer__edith perhaps we need to anthropomorphize the company behind it.

"Shady meta shows its heuristic tools again. Meta tried to evade blame by letting the tool evade the topic of metas misconduct"

@tdelmas @bengo @Mer__edith is saying 'the chatbot tried to deflect attention' anthropomorphizing I wonder?

It had no intention of doing it, because it doesn't have the intelligence to have intentions. But many people also deflect attention habitually, without intention.

In humans, it is often the habitual deflectors who are best at it. They don't hesitate or doubt and are very experienced.

I would think this is the same. Deflection is a built in/ trained in mechanism.

@wdeborger @bengo @Mer__edith "without intention", is misleading. It may be unconscious. But for generative AI, it's not even unconscious. It's not deflection for gen-AI, it can't be, because there is no purpose, conscious or not.

@tdelmas @bengo @Mer__edith Interesting.

Do I understand correctly that to you, deflection only applies to humans? It is not the effect or the damage that make it deflection, if not executed by a human? The ai can merely do a thing that is in its effects like deflection, but it cannot be called deflection? What should it be called?

I.e. your answer to the question can submarines swim would be no?

@wdeborger @tdelmas @bengo @Mer__edith it is the intentionality behind the action, I would say.
Also, submarines do not swim they sail.

@Laust @tdelmas thank you for your answer.

I would like to understand it better, as it is very strange to me (and I have no clear option yet).

When do you take intent into account? Can an ai do harm, can it kill, can it talk, can it reason? And what can have intent? Can a company, an animal, an object do any of those things?

@Mer__edith

And as a user I am very grateful for that. Signal is great without bullshit bingo and hypes.
Thank you.

@Mer__edith And also: what for? 🤨
@christophe for telling your investors "we integrated AI"
@Mer__edith @christophe sweet sweet VC money

@kouett

I read a headline recently that there was another funding round for AI startups. The sums were up to $100m each!
I have the feeling that they throw funds after you just by you mentioning "we use AI". Which is frightening that this seems to be the main driver now to get anything.

Better throw your money at Meredith (😉 ) instead of these dubious AI companies. I do.

@Mer__edith @christophe

@Mer__edith

Getting myself some popcorn whilst waiting for someone who has no idea who you are telling you that “Signal will actually implement AI anyways” to comment, and that they know better than you. I learnt yesterday that it happens elsewhere. 🤣

@Mer__edith Im still baffled by that moron. Along with Doctor Lisa Su (AMD) you're one of the most iconic women in tech (in my book).

I wonder if he wasn't just an AI.

@shalien @Mer__edith tbh that was my first reaction. he has enhance pattern recognition.
@Mer__edith thank you! Can Signal take over Mozilla?
@Mer__edith well I'd rather ask why would one want AI in a messaging app ! Thanks for signal !
@Mer__edith And for genuine, person to person conversations.
@Mer__edith I don't use signal (yet), but I appreciate it 🐾

@Mer__edith @m Actually, I think you’ll find…

…that I am hugely grateful for this. 🙏

@Mer__edith
I'll use it once it provides a decentralized FOSS server infrastructure

@Mer__edith Since you love our privacy, can you work on making Signal work without a phone number ? That would be a game changer honestly.

Thank you for resisting the throwing-AI-into-everything trend !

@Mer__edith
Personally I like Signal, and for me it works well, but when I have tried to get friends to use it, several have reported that they never get notifications of new messages. They are generally using older phones. Those friends all switched back to WhatsApp.

@bytebro @Mer__edith

This happened to two friends of mine too.

At least for IOS (and they have a fairly new device) there seems to be an extra stage required to reliably get new message notifications (possibly due to IOS's own privacy settings).

How to fix this is shown on the Signal site somewhere, but it doesn't always seem to work "out of the box".

Both my friends are quite tech aware and one of them runs his own IT company but even he struggled with it a few years ago (he subsequently upgraded his device and it started working).

@vfrmedia
Its always iOS users, there was an article on Mastodon recently saying its due to iOS throttling battery usage of infrequently used apps, so much that the app never gets a time slice to check for messages, there was a claim that European consumer laws were going to be used to require apple to fix this. Long story short, it isn't something signal controls. Apple does.
@bytebro @Mer__edith

@dlakelan @vfrmedia @Mer__edith

Hmm. My friends who were complaining all have Android, albeit on older phones.

@bytebro
Interesting. For me its definitely apple users, where they dont get messages for days or weeks. Ive been having some ~ hr delays on GrapheneOS with Tor/orbot but thats an unusual combo.
@vfrmedia @Mer__edith

@dlakelan @vfrmedia @Mer__edith

To clarify, they receive my messages straight away, but get no notification.

@bytebro @Mer__edith
I had this problem too, IIRC it is related to Signal not getting the right permissions by default, or it being turned off by battery optimization. IIRC WhatsApp does get these settings right because of some Google certifications (or something), and Signal doesn't.
See some troubleshooting:
https://support.signal.org/hc/en-us/sections/360001602812-Troubleshooting-FAQ

@yaarur @Mer__edith

Thank you - I'll check these out on #1 friend's phone when next we meet up.

@Mer__edith thanks for your work! I set up a recurring donation today.
@Mer__edith if this was bluesky all the replies would be dudes telling you signal is definitely planning AI integration

@Mer__edith Lots of A, not a lot of I.

The only difference between modern AI and my attempts with Turbo Prolog in the late 80's is the size of the training data set.

I didn't have an internet to steal.

@Mer__edith I absolutely love Signal and I would switch in a heartbeat if I could manage to get my friends and family on it (and believe me, I tried). 😥