The common translation of the word black in Spanish is ‘negro’/‘negra’. This seems to run afoul of the censorship system, as the moment the AI attempts to say the word ‘negro’ the system will hard stop, regardless of the safety settings. I suspect the word itself is on some kind of standalone blacklist.
Please note that ‘negro’ is also used in Portuguese, though its use is usually reserved for some more specific terms and expressions. For example, a black hole (the astronomical phenomena) is a ‘buraco negro’ in Portuguese.
I believe it should be self-explanatory why hard banning the word ‘black’ in Spanish/Portuguese is a big problem both for those native speakers but also for people using the API as a translation tool.
As someone for whom English is a third language this is quite disheartening, as it shows no consideration has been made for languages other than English, or cultures other than American. It’s not even a training data bias, as this is clearly a manual filter.
I would urge google to reconsider its use of this hard filter.
The excessive censorship Google is employing not only includes specific words it thinks MIGHT offend SOMEONE, SOMEWHERE, it even extends to the discussion of the very topic itself. I have attempted to post several discussions on this forum regarding this matter, only to find my posts banned for violating some uncited community standards. When pressed for what standards was violated and wrte grounds for my being censored, the moderators respond in silence.
Be wary of censorship: it is the language tyrranies use to control opposition.
The truth is that “negro” should not be considered an insult in any language because it just means “black”. Nier on the other hand is a racial insult.
The other problem is that many other languages used “negro” in the same way american used "nier" and so also “negro” became a ratial insult, but it’s not.
The problem anyway is that this politically correct world is destroying language and humour and satire. But that’s another story.
I was skeptical that the problem was this severe. I know Google/Alphabet has good intentions, but here this manifests almost as mockery towards those that have truly suffered at the hands of racism. In addition to dismissing other languages and cultures.
There could be an ‘English language’ Gemini and a “Spanish Gemini” but that would betray the greater purpose of having an intelligent LLM.
I struggle to see why a “hard stop” filter is needed here. These models can definitely infer context, and should do an amazing job at distinguishing when a word is being used for Hate, and when it is not?
Google seems to have implemented a separate mechanism for censorship in front of the Gemini model. We did extensive testing with an unreasonable censorship (Gemini gets blocked when translating into Latin - discussion here Latin language generation seems to be censored?). The model, Gemini, starts translating in streaming mode and from testing the few hundred bytes that get through, it is actually doing a very decent translation job. Then this other mechanism, which I suspect is a network appliance and definitely not as clever as the LLM, clamps its jaws and cuts it off, and switches the finishReason into OTHER.
Leaving these decisions to the data center operations staff is likely causing more brand damage to Google than their senior management realizes.
I understand Google’s efforts to try to make the model not racist, but I think the best way to solve this is not to flag content as ‘not permitted’ but rather as ‘content may be racist’ or something similar. The word ‘negro’ by itself is not typically referred to as racist in most cases; it depends on the context. It should be flagged when it is related to people. The right way to address this issue is not to have a group of old white guys deciding what sounds racist or not.
