OpenAI has made a change that allows its models to speak more freely. Until now, its chatbots have often refused to participate when a topic is politically, sexually, or morally sensitive.
The company announced last week that it had revised its 187-page Model Spec, a document that outlines rules for how an OpenAI AI model should behave and when it should refrain from commenting.
This gives the model more ‘intellectual freedom’, meaning it can discuss specific topics more broadly, debate them, and produce text without running up against its own rules, preventing it from providing answers. It also has the new principle that it may not lie by making incorrect statements or omitting context.
It is still challenging to determine precisely how broad that will be. You can discuss sensitive topics more often or ask questions about them with ChatGPT. Think of political statements or actions. In the past, the standard answer would be that they could not help you. Now, it should answer that or at least provide more context.
More nuance, not uncensored
Techcrunch gives the topic ‘Black lives matter’ versus ‘All lives matter’ as an example. When we bring that up with ChatGPT, we get a nuanced answer that explains both statements and provides context about their origin. When we ask further, we also get sources to support certain statements.
Another sensitive topic is erotic conversations. ChatGPT generally does not participate in these, but on forums such as Reddit, there are numerous ways to engage the bot in a more explicit conversation. Since the adjustment, several users report that this is now easier, although some say they hardly notice any difference. There is still a limitation here, but it is easier to circumvent.
Why OpenAI is making this change is not entirely clear. The company could be adjusting its policy in light of a trend under US President Trump to not punish users for discussing sensitive topics on tech platforms. However, it could also be that OpenAI sees that it is losing users to less strict competitors and is somewhat relaxing the rules.