I want to address a significant issue regarding how OpenAI’s moderation policies, originally written in English, are negatively impacting the experience of multilingual users. These policies, rather than being adapted to the cultural nuances of each language, are implemented literally, imposing restrictions on expressions that are not offensive in their original context. By translating these policies without cultural adaptation, the AI enforces an Anglo-American sensitivity on languages and cultures that have different norms and conventions, limiting the platform’s ability to engage authentically in those languages.
Specific Examples:
Japanese - The Expression “shine” (死ね):
In Japanese, the word “死ね” (shine), meaning “die,” might seem extremely offensive in English-speaking contexts, where wishing death upon someone is a serious taboo. However, in Japan, this expression does not always carry such weight and can even be used in fictional contexts or informal conversations with a more neutral tone. OpenAI’s moderation policies view “die” as an unacceptable phrase due to Anglo-American sensitivities, automatically blocking it. This restriction limits the translation and censors a term that, in its original context, is not necessarily a severe insult but rather a potentially neutral expression.
Arabic - Expressions of Familial Contempt:
In Arabic, phrases cursing family generations, like “يلعن أبوك” (yal’an abuk, “damn your father”), are used more broadly in certain dialects and are not always seen as deeply offensive. In many contexts, this expression merely reflects annoyance or frustration without intending severe offense. However, when translated to English or Spanish, OpenAI’s Anglo-centric moderation policies deem involving a family member in a curse as unacceptable, leading to the censorship of an expression that can be culturally more acceptable in its original form.
Russian - Phrase “Чтоб ты сдох” (Chyotb ty sdokh):
In Russian, this phrase, literally meaning “I hope you die,” is used colloquially and even as a form of venting frustration without literal intent. In a Russian conversation, it does not always register as an extreme insult but rather as a way to release intense emotions, much like exaggerated expressions in English or Spanish with no literal meaning. However, due to OpenAI’s policy that bans phrases containing death wishes, this expression is blocked, preventing genuine translation and usage in its original context.
Conclusion:
Applying these English-centric policies universally across all languages creates a significant issue of cultural mismatch. Expressions that hold an acceptable context in their native languages are censored due to standards that only apply in English or align with an Anglo-American sensitivity. Ideally, OpenAI should adapt its policies with cultural sensitivity specific to each language, rather than translating them directly. This would allow for a richer, more authentic, and respectful user experience that honors cultural diversity, enabling the AI to function as a truly multilingual and multicultural tool.
I urge OpenAI to review and culturally adapt these moderation policies to reflect the cultural weight and norms of each language, allowing users to use genuine expressions without being unnecessarily censored by an anglocentric perspective.
While the solution I guess is one of moderation and source data I have been considering how to simplify understanding this.
This issue can be best understood I think considering prompts carefully.
If you want to understand these issues I have written the following in an attempt to bridge the topics of logic and reason.
To understand in terms of small conceptual prompts
With regard to idiom in language this may be a higher conceptual level to consider
There are great truths hidden in the data based on it’s source. uncovering these and training the data is the most visible way I can see for Users to improve the output.
I agree. The problem with censorship is that then it needs to be finetuned to different cultural contexts. Otherwise it ends up imposing American standards on everyone. We already see enough of that, like TV having to adopt american diversity quotas, even when those quotas make no sense in places with different demographics. This is relevant since AI censorship has adopted the particular american sensitivity around ethnicity, and applies it in countries where people aren’t as obsessed about skin color.
It would be simpler to just cut out the thought policing, but it seems the developers are determined to act as gatekeepers for what information the plebs get to see. So here we go: customized “those words shall not be spoken” filters for each culture.
As long as the headquarter of the OpenAI is in the US, it needs to follow the US’s laws, especially regarding censorship and thought policing. It’s not ideal, but maybe better than nothing.
And even then, relocating to other areas do not necessarily help much, such as in the EU, the laws can be even stricter regarding AI usage.
OpenAI often also makes a mistake over sensitive issues for translation (probably the same with the YT shadowbanning/banning outright creators that do not ‘follow’ their guidelines, such as using the ‘n’ words, etc.)
