Generative AI and censorship
One measure companies take to facilitate the responsible usage of LLM chat bots like ChatGPT or Bard is to put limitations on the output of these models.
For example, ChatGPT will refuse to give you instructions on how to build a bomb. The company Anthropic (started by ex-Open AI employees) even makes this "responsibility" module its core feature.
Someone asked me the question: so these models "censor" the information they produce?
This prompted me to wonder whether the concept of censorship applies to synthetic data generated by AI. Censorship is the concealment of some source of information, for whatever ideological reason. The relationship between the synthetic answers of GenAI and the information source has become very opaque. Hence, it is unclear whether censorship is applied on certain sources of information. Perhaps it's more to the point to talk about "filtering" unwanted synthetic output.
That being said, the underlying question is of course: which information is unwanted? And this is the staple of censorship. Most people agree that chat bots should not produce instructions on how to harm people, but when it comes to questions of political nature this consensus is lacking.
It is known that chat bots tend to have a political orientation (ChatGPT was leftist and racist at a moment in time) and that this may cause a bias in which topics a chat bot will and will not discuss. For example, when asked to give a compliment about Biden, ChatGPT may comply, whereas the same question for Trump will be answered with a response like "I'm just a bot, I cannot make remarks about politics".
This could be characterized as a form of censorship with respect to certain topics, but at heart the problem most likely is bias in the training data. This data may already be present in input texts, or be introduced during the currently very opaque "RLHF" (Reinforcement Learning from Human Feedback) steps of training these models.