AI, Guardrails, and Anti-Semitism
Yesterday, the Iranian regime posted the above image in one of the squares in some city in Iran. Too cowardly to pick up the phone and speak in their broken Hebrew to rent space on the side of the Prima Park (come on Khomeni! Your Hebrew - as evidenced in this poster - is probably just as good as any Oleh’s and we have to speak broken Hebrew every day!) to post this ad, he posted a message that terrified both us here in Israel and whoever else understands Hebrew and passes through this square.
Anyway, it seems like Khomeni has an account on DALL-E, though he might not be a paying member given that he got a bad translation to Hebrew and there is the tell-tale repeating of the background. More likely, Google Translate gave him the wrong translation and then they photoshopped it on top of this AI slop, because I can never get AI to print me legible text.
Joking aside, I did want to take a moment to discuss this image in the context of all of the other discussions that are taking place around AI today. Every few weeks you will hear that such-and-such an AI company held back a certain release because of safety concerns. Now mostly these have to deal with not being racist, or producing things that can be damaging to the political process. But perhaps anti-semitism is something that isn’t being properly monitored in regards to the outputs that AI is producing?
It is hard to say exactly what the prompt was that generated the image above, though it was probably in English (see Khomeni, I’m sure if you were willing to pay for an ad they would speak any language you wanted!). It probably also included the words “Israeli soldiers”, “fear”, “running in panic” and things of that nature. It probably took several tries (maybe he is a premium member after all), yet somehow the AI allowed him to continue with generating the image.
It also is not for lack of anti-semitic material online that an AI can’t be trained on what is and what isn’t anti-semitism. I’m pretty sure there are enough 4TRAN sites that are anti-semitic, and probably even Reddit pages that AI can get a pretty good idea of what it is.