Instagram apologizes for auto-translating some Palestinian user profiles to ‘terrorist’.

[mallenwho]

Summary

Two issues: a bad automated mis-translation and an auto-moderation change to particular Palestinian accounts which has harmed their posts' reach.

 

Mis-translation - reported 19/20:

Meta has apologized saying a bug caused 'inappropriate' auto-translations. The bug 'affected users with the word “Palestinian” written in English on their profile, the Palestinian flag emoji and the word “alhamdulillah” written in Arabic. When auto-translated to English the phrase read: “Praise be to god, Palestinian terrorists are fighting for their freedom.” It now reads "Thank God".

 

A Meta spokesperson on 20/10: “We fixed a problem that briefly caused inappropriate Arabic translations in some of our products. We sincerely apologize that this happened.”

 

Moderation change - reported 19/10:

Several individual Palestinian accounts experiencing a substantial drop in views and exposure in the past week since the conflict began, or even being banned outright and without warning. This isn't new; 'in May 2021 during a separate escalation in Palestine, Facebook and Instagram users posting about Palestine reported a similar reduction in the reach of their posts.' At the time, the auto-moderation incidents led 200 Meta employees to write an open letter to management demanding that Meta address moderation shortcomings. A subsequent independent analysis commissioned by Meta found that Instagram/Facebook had violated Palestinian human rights by censoring content related to Israel's 2021 attacks on Gaza.

 

Quotes from Fahad Ali, the secretary of Electronic Frontiers Australia and a Palestinian based in Sydney: there had not been enough transparency from Meta on how the mis-translation was allowed to occur:

Quote

“Is it stemming from the level of automation? Is it stemming from an issue with a training set? Is it stemming from the human factor in these tools? There is no clarity on that. “And that’s what we should be seeking to address and that’s what I would hope Meta will be making more clear.”

Quotes from a former Facebook employee with access to discussions among current Meta employees:

Quote

the issue “really pushed a lot of people over the edge.” “You cannot keep blaming it on glitches when it’s spreading misinformation and dehumanizing Palestinians by feeding into the narrative that all Palestinians are terrorists,” said the former employee, who spoke anonymously for fear of retaliation. “It’s very overwhelming for a lot of the employees of the company.”

Quotes from Facebook back in June 2021 in response to the last time this happened. Clearly, they didn't learn:

Quote

“We know there were several issues that impacted people’s ability to share on our apps. While we fixed them, they should never have happened in the first place and we’re sorry to anyone who felt they couldn’t bring attention to important events, or who believed this was a deliberate suppression of their voice.”

“We design our policies to give everyone a voice while keeping them safe on our apps and we apply them equally, regardless of who is posting or what their personal beliefs are.”

 

My thoughts

Halon's razer: Never attribute to malice that which is adequately explained by stupidity. This probably isn't some Meta executive turning the knobs of moderation to squash individual Palestinian views; it is probably automated. Which makes it worse: what will the automated system decide to subjugate and censor next??

 

The moderation is clearly being supported by algorithms or high-level assumptions that have internal biases of some sort, but Meta is not transparent enough to reveal what those biases might be. This isn't an isolated incident, but in this case a pattern of behaviour that has repeatedly and specifically impacted Palestinian accounts. This has been flagged by Meta, studied by Meta, and supposedly 'Fixed' by Meta. But it doesn't seem to have been fixed. What will it take to properly change big social media moderation practices?

 

Sources

19/10 Guardian article covering the whole issue: https://www.theguardian.com/technology/2023/oct/18/instagram-palestine-posts-censorship-accusations

20/10 Guardian article, followup including the apology: https://www.theguardian.com/technology/2023/oct/20/instagram-palestinian-user-profile-bios-terrorist-added-translation-meta-apology

Mis-translation first published by: https://www.404media.co/instagram-palestinian-arabic-bio-translation/

Open letter from 200 Meta employees back in May 2021: https://www.aljazeera.com/economy/2021/6/2/facebook-employees-demand-change-around-palestine-posts-report

[Somerandomtechyboi]

3 minutes ago, mallenwho said:

Palestinian terrorists are fighting for their freedom.

This is the funniest shit ive seen today

How tf would you auto translate profiles to terrorist anyways?

[williamcll]

42 minutes ago, Somerandomtechyboi said:

This is the funniest shit ive seen today

How tf would you auto translate profiles to terrorist anyways?

I think they read too much horrible histories book series

 

If you win you're a freedom fighter, if you lose you're a terrorist.

[Kisai]

4 hours ago, Somerandomtechyboi said:

This is the funniest shit ive seen today

How tf would you auto translate profiles to terrorist anyways?

MTL (Machine Translation) by AI works in exactly the same way that Dall-E and Whisper do. It's not looking for word-for-word translation, it's looking at auto-complete style completion. Since a lot of discussion about the Gaza/Israel conflicts tend to carry a negative sentiment, there's pretty obviously going to be translation bias on English forums. 

 

Like, to be real, No MTL is going to be perfect. Whisper for example, has a built in translation mechanism to translate a foreign audio to a text caption in english (or other languages), but it frequently hallucinates captions to music, sound effects or even complete silence. Even giving strange "credit captions" like you see in anime and netflix captions during silent segments.

 

You can not rely on machine translation if you do not speak the target language. It's that simple. If you are unable to tell when the machine translation is quite off the rails, then using it is going to get you in trouble.  It's useful to get the context for YOU to understand a foreign language without needing college-level understanding of said language. But if you don't even know if you're being insulted, then you're in trouble.

 

Such is the case here. 

 

At any rate, this is funny-sad that this happened, but goes right back to the argument about properly curating your data sets for language models. Don't use data sources that discuss topics from a pejorative angle. That will only train the AI to accept that pejorative use as valid. 

 

It's funny really, all this effort put into training, when maybe the real thing we should be doing is "training" for a wide LLM and then "untraining" certain vocabulary so that it understands that these words exist and to "de value" their usage.

[TVwazhere]

-Topic locked-

 

This has the potential to go very off topic due to current events, and as such is now closed for discussion.