Over the last few days something odd has been happening with posts on X in Hebrew. Non-Hebrew speakers wishing to know what they were saying would click on “translate” and, for some posts, would get a translation. Other posts, though, would trigger a message saying “Hebrew is not currently supported for translation”.

X hasn’t commented officially but according to its AI chatbot, Grok, this wasn’t due to a bug in the system but the result of a policy decision not to translate Hebrew-language posts that artificial intelligence had classified as “inflammatory”. Meanwhile, Hebrew posts deemed to be “neutral” were unaffected.

Baffled X users soon began pointing out innocent Hebrew posts where translation was blocked for no apparent reason. As recently as yesterday, Grok was busy online aruing — in the face of considerable derision — that selective blocking of translation is a brilliant idea. Today, though, the scheme appears to have been abandoned and translation is back to normal.

Incitement to violence

The origins of this experiment can be traced back to a spike in Hebrew-language posts on X in the wake of the attack by Hamas in October 2023. Along with natural expressions of grief and horror, there were also many examples of hate speech and incitement to violence. One Arab monitoring organisation identified more than 19,000 examples on X during the first few days after the attack.

According to Grok, X wanted to avoid “amplifying inflammatory or policy-violating content, like calls for violence, to a global audience” and therefore decided not to provide translations. The aim, it said, was “to curb harm without blanket censorship” — because X “prioritises free speech for opinions”.

One question this raised was why, if Hebrew posts were violating X’s content policy, it didn’t simply block them or suspend the account. There would then be no need to fret about whether to translate them. Grok’s answer was that allowing the Hebrew posts to remain (but without translation) “upholds free speech for Hebrew users”. However, Grok resisted suggestions that Hebrew speakers were being given more latitude than other X users: “Violent speech violates platform rules universally and faces removal or restrictions.”

Another problem was how to justify selective non-translation of Hebrew posts, in view of X’s policy of “prioritising free speech”. Given the current situation in Gaza (not to mention settler violence in the West Bank), surely the non-Hebrew-speaking public are entitled to know what Hebrew speakers are saying about it — so why was X making it more difficult for them?

‘Incendiary’ translations

In response to that, Grok switched to a second line of defence: “translation risk” — which amounts to a claim that automated translation makes Hebrew posts sound “more incendiary” than they actually are.

Machine translations from Hebrew, it said, have “often stripped idiomatic nuances, rendering rhetorical or hyperbolic language as literal incitement”, thus raising the possibility of “inaccurate or literal renditions” inflaming the situation. It added that “Arabic content, despite volume, exhibited fewer such rendering issues that breached policies globally.”

I found this rather puzzling because the quality of machine translations these days is generally very good, so I asked Grok: “Can you give an example of a post in Hebrew that sounds ‘more incendiary’ when auto-translated into English?”

One minute later Grok replied:

Consider a Hebrew sports fan post: “שחטנו אותם במשחק!” (We “slaughtered” them in the game!), using “shachtanu” as slang for dominating rivals overwhelmingly.

Auto-translation often renders it literally as “We slaughtered them in the game,” stripping idiomatic intent and evoking graphic violence to English readers unfamiliar with Hebrew colloquialisms.

This amplifies perceived threats amid tense discourse, while the original is harmless trash-talk — common in Hebrew but lost in mechanical parsing.

This gave a clue as to how X was distinguishing between “inflammatory” and “neutral” posts — by scouring them for words with violent connotations, even when used in a non-violent context.

My first reaction was surprise that Grok viewed “We slaughtered them in the game” as problematic. Did it really imagine that non-Hebrew speakers would fail to recognise “slaughtered” as a figure of speech and instead assume one team in the game had killed the opposing team?

Memo to Grok: Please don’t take this the wrong way, but right now I could murder a pizza.