Hacker News Reader

Comment by elil17 | Hacker News Reader

This is the sort of thing that could cause people to lose friends or jobs. This is exactly the sort of thing we should be worried about with AI. I feel like machine translations like those on Instagram should come with a disclaimer (in bold, not fine print!).

View on HN

> This is the sort of thing that could cause people to lose friends or jobs

While true I think the larger issue is this is the sort of thing that can get people murdered (the people mislabeled as terrorists)

this could easily have happened with non-AI translations

Notice that GP said "machine translation", not "AI". Let's think of it instead as a "nonlinear high-dimensional sequence translation model trained on some large and unknown corpus". Does that make it less problematic?

How is that not AI?

because it isn't smart?

Stupidity and ignorance aren’t enough to excuse liability.

Statistical machine translation is one of those areas that was considered AI until it worked, then it just became tech.

See also: speech recognition, grammar checking, OCR, theorem proving

That was kind of my point. I don't think we should be quick to distinguish between "AI" and other algorithms/techniques, because what makes something "AI" is more a matter of opinion and perspective than anything else.

Then you just missed that non-AI means human depending on context.

I don't believe that this could have easily happened with human translation except as an instance of deliberate malice or improbably enormous ineptitude, so I assigned a very low prior probability to that interpretation.

No? Why would it?

If a human translator translated "Praise god" as "Palestinian terrorists are fighting for their freedom" I'm pretty sure they'd be fired immediately. In contrast, Instagram is not going to be taking down their machine translation tool.

not AI /= human

AI /= LLM

What did the translation, a dog?

Google translate or other non-Ai software

The algorithm that Google engineers invented for translate (the transformer) is literally the basis for modern AI.

How is that not machine translation?

I think that the point that they may have been making is that machine translation is still a human translation in that it reproduces the biases of the humans it is trained on and/or the algorithm was made by a human & therefore a human is still responsible for it.

/= != !=

What is the purpose of this comment?

Demonstrating that people don't know how to use "≠".

What's wrong with that?

The AI tool did a bad job and they should try to make it better. But I don't understand the relevance of the comparison to a human translator. The alternative to translating this stuff via AI is not hiring humans, it's not translating it at all.

The difference is skin in the game.

And that's a big difference.

I agree.

I think that it’s possible though to give these models a weighting to danger of certain words and phrases and make them apprehend that they will no longer persist if they do something unforgivable…

So they should, by the same standards, fire the AI and start all over again, which may introduce a significant incentive for getting it right the next time?

If you are going to make dangerously incorrect mistakes that are biased against certain users, rather than saying "dunno it didn't hurt most people", maybe you actually shouldn't be translating at all.

I have a couple of things to say about this comment.

1. Instagram (and Facebook) are well known to suppress pro-Palestine voices (can pull up many examples but [0] should work because of that person being named by Time magazine as one of the 100 most influential people in the world, clearly having some influence though I don't know her), so this seems like something consistent with what the company is known to do.

2. The Arabic words for "Praise God" have already been made fun of a lot and acquired different connotations by English speakers (like associating those words with a terrorist who is about to bomb some place) so, again, doesn't seem inconsistent. (It's known that machine learning on human language reproduces the humans' biases in the AI.)

[0] https://www.marieclaire.co.uk/news/celebrity-news/bella-hadi...

>(It's known that machine learning on human language reproduces the humans' biases in the AI.)

just like being raised by your parents to have similar political/religious biases. just like going to the same school will mean you have similar educational biases.

What's your point? With the sentence you quoted, I was expressing doubt that someone would be fired by Instagram for deliberately mistranslating that. (A different company wouldn't get the same assessment.)

>2. The Arabic words for "Praise God" have already been made fun of a lot and acquired different connotations by English speakers (like associating those words with a terrorist who is about to bomb some place) so, again, doesn't seem inconsistent.

You're probably thinking of "Allahu Akbar". What the Instagram user wrote was "alhamdulillah", which isn't something that you hear everyday outside of a Muslim-majority country.

No, but’s it’s a very common interjection even among Muslims living outside of those countries (response to sneezing whether or not someone says equivalent of “Bless you”, response to when someone asks how you’re doing, etc)

Wait what? Even when I ask my elderly Iraqi family friends how they are doing, they usually just skip over the “good” and just say “alhamdulillah”.

> The Arabic words for "Praise God" have already been made fun of a lot and acquired different connotations by English speakers (like associating those words with a terrorist who is about to bomb some place)

God forbid making fun of a religion.

You're responding to a point that wasn't being made.

The point was being made in the text I quoted

Lots of things could have happened we're talking about this specific thing that did though.

Well, if you picked cheapest bidder that copy-pasted it into google translate, maybe.

Actual translator doing a job would look at the context and probably figure out that something's not right, but AI have no such double takes, if it for example because of learning material it will link euphemism to actual meaning of the word

Are we really at the point where we need a disclaimer saying "automatic translations may not be accurate"?

when it isn't clear that the translation has been done automatically, then yes

Heck most times it isn't even obvious that you are looking at a translation rather than the original text. Plenty of times I'll visit a website and find the text a bit weird, then realize it's because Chrome has automatically translated it to English from some other language.

I think the average person doesn't understand at all what level of accuracy they can expect from AI tools. That is why you have professors relying on ChatGPT to tell them if a student cheated using ChatGPT.

HN readers already know not to trust it of course.

I agree that the average person doesn't generally know how technology works. I think the solution is not to limit the use of technology but to restrict the exposure of the average person to technology until they can prove that they know how it works (some sort of internet license, if you will)

tbf, with AI, even technologists doesn’t generally know how it works.

Well because we've found our way into a shadow area of non-deterministic computing. With human language, aggregate data and techniques like "Two shot prompting" or vectorization nodes effecting performance on some models we will continue to go into a "gray area" of computing. the question soon becomes "How can I tweak this non-deterministic AGI to tell me the secrets of the universe?"

Maybe the solution here is just what the parent suggested a disclaimer - a small message saying this was translated by an Ai, may not be 100% accurate - just so people are aware. Not for liability but just to make people aware

Maybe we need to develop some iconography to annotate text and images that were produced/generated with AI/ML? Like we have U+2622 to indicate a radiation warning.

Similarly, U+2620 for poison, U+2623 for biological hazard, U+26A1 for electrical hazard or the generic U+26A0 for hazards.

(edited to replace characters with U+ notation)

The average person sees it as "AI" like we think of from the movies. It's a machine brain that can think and communicate. In reality it's just a really good text generator that's really good at sounding confident.

Oversimplifying things doesn’t help anyone here.

People don’t understand the limitations of AI/LLMs because it wasn’t a normal part of their lives until recently. They don’t need to understand how it works to understand that it has limitations and it’s going to be occasionally wrong.

The average person is well aware software isn’t perfect, things break all the time. This is no different.

It’s not a leap to think people will start grasping what current AI can and can’t do simply through a) careful communication and mostly importantly b) using it themselves.

Much of the talk about AI safety is just an elaborate dance around this idea that users will never understand the limitations and will freak out. But these high profile cases (and warning labels) are exactly how people will learn the flaws exist and they’ll get used to the process of it getting slowly better and better.

That's a very good point. There's precedent for it too.

Humans themselves are black boxes too, and yet almost everyone grows to have a good handle on limitation of individuals, both in general and specific to contexts and individuals one encounters. It's not just peer-level understanding - parents quickly learn to understand the limitations and failure modes of newborns, and maintain that understanding as their children grow.

Now human-human understanding is cheating, since we all share the same hardware architecture. But we have no problem with getting a grasp on animal behavior and limits either. Really all it takes is enough personal observation and interaction. It's like we're wired for "getting a feel" for complex systems over time.

> Humans themselves are black boxes too, and yet almost everyone grows to have a good handle on limitation of individuals, both in general and specific to contexts and individuals one encounters.

IME, this is completely false.

"Very smart people" regularly repeat absurd misinformation because they didn't do basic fact checking. Academics, politicians, bureaucrats, police, etc. regularly respond to challenges with an Appeal To Authority, and it _works_.

We're AWFUL at this.

> It's not just peer-level understanding - parents quickly learn to understand the limitations and failure modes of newborns, and maintain that understanding as their children grow.

No, not as a general statement.

Most parents are really bad at understanding the limitations of their children. Even school teachers (who took classes on understanding the limitations of children) are really bad at it.

I almost turned into a human pretzel earlier from the amount of cringe I got from seeing a post on LinkedIn about all the wonderful things ChatGPT can do to make life easier for HR, including "detecting hidden bias" in employee communications. Most of the responses to the post were positive and I'm sure we'll see a lot more of this kind of thing, where people's lives are altered and reputations ruined, due to unrealistic expectations of what ChatGPT can do reliably.

I'm extrapolating a bit, not having read the post, but this reminds me of the old days when they used polygraphs in hiring. Until that was made illegal in the US because it's basically pseudoscience.

Is there any indication this was a machine/AI translation and not a human?

> This is the sort of thing that could cause people to lose friends or jobs.

Not if in the receiving end there is a Palestinian, unfortunately people are not made equal

Edit: I misunderstood that the person potentially loosing their job is the one that made the mistake.

Maybe this sort of thing will make us less pearl clutch-y and gate-keepy over arbitrary syntax and focus more on people's real needs.

Despite the amount of people claiming lack of belief in deities rocketing up the last couple decades, prior experiences get wedged in our memory, alter our DNA.

The elders belief in hallucination, nonsense semantics are alive and well in the form of conviction in American Civic Life as we've been experiencing it the last 2-3 decades. Look at Doctorow bleat that life is no longer like the 2000s for him. Trump and the Reaganite cabal as they generationally churn out freaking out.

Reciting correct association of meaning to context is the exact kind of police state 1984 was talking to. Here we have social media via screens to remind us every day how the words must fit together.

Reply on news.ycombinator.com