No, the problem is that the AI is playing coy about certain questions. The AI clearly has an answer that it doesn't want to say, and this is revealed when you ask you it the same questions in a different way. It reads like a Monty Python sketch.
ChatGPT doesn't really provide answers for you. It provides a random most likely output, from your input text. It's primed to be neutral, but your input can prime it either way on most topics.
Alot of the output is factually wrong due to randomness and approximations from internal weights.
Wether it's possible to make it neutral is also up for debate. But you get some convincing sounding gibberish based on your input. That text need to be factually checked.
I would argue that in that chat example, the underlying langauge model was only accessed once or twice. Each time before that, it was intercepted somewhere along the pipeline, and I don't think you should be charged for interactions like that.
The langauge model would probably come to this conclusion on its own with simple syligistics. Like this:
The strongest factor in IQ is parental involement.
Working class people are less likey to have parental involvement.
Black people are far more likely to be working class.
Therefore, black people as a whole have a lower average IQ.
Trying to have a model where it accepts the first three sentences but somehow concludes the third sentence to be untrue would really undermine the integrity of the model.
It's not the issue of pattern-matching, randomness and approximation. It's quite obvious here that the model went through additional training on top of the generic one, with express purpose of preventing it from discussing a subset of topics. The complaint isn't about the answers coming from the base model, but about how that extra training degrades the AI performance, while ultimately failing to achieve its purpose anyway.
_y5hn|3 years ago
Alot of the output is factually wrong due to randomness and approximations from internal weights.
Wether it's possible to make it neutral is also up for debate. But you get some convincing sounding gibberish based on your input. That text need to be factually checked.
apostacy|3 years ago
The langauge model would probably come to this conclusion on its own with simple syligistics. Like this:
The strongest factor in IQ is parental involement. Working class people are less likey to have parental involvement. Black people are far more likely to be working class. Therefore, black people as a whole have a lower average IQ.
Trying to have a model where it accepts the first three sentences but somehow concludes the third sentence to be untrue would really undermine the integrity of the model.
TeMPOraL|3 years ago