| One thing that could be mentioned is how OpenAI's content policy flagging system is still unbearably overzealous on the ChatGPT platform (both w/ GPT-3.5 and GPT-4), punishing the user for using "bad words" in their input, often in a completely arbitrary manner. And no, not asking for the model to create anything else except a reply altogether and perhaps a summary. And nothing more "edgy" than your average Youtube transcripts or even pieces of classic literature. It's ridiculous at times, other times downright annoying, as you have to counter the claims in order to keep your account in good standing -- mind that all content policy warnings might lead to further automated, one-sided account warnings over e-mail. And no, it doesn't help if you've poured four-figure sums into their services over both the API as a dev and into ChatGPT Plus and the Team Plan. Quite frankly, in its current form, the flagging system is hampering the whole platform's fundamental usability for anything serious, since you constantly have to be aware of not tripping the flagging system over some trivial crap that someone, somewhere might find remotely objectionable. Sad to say, but unless OpenAI revises their flagging system's inner workings, they're literally turning a tool into a trinket with a thoughtless word-police unit constantly ready to interject any conversation or analysis. It feels unnecessary and downright Orwellian, and I'm pretty sure that new users are going to find it off-putting as it probably serves as proof to some that OpenAI really wants to control the user's basic freedom of expression on their platform, even when it's not even a "public forum" but literally a more or less "secluded" interaction between a human and and an AI. I get it that the model shouldn't be spouting out anything controversial, but the users should nonetheless be able to go through "serious topics" without the alarm bells going off. It's been 1.5 years now and progress has been made in terms of OpenAI fixing the biases in the model and making it overall better, but the automated flagging system still feels like a Philip K. Dick-level automated thought police from hell. I'd seriously recommend OpenAI revising their system so that they'd have a separate A/B comparison and a veto process where the initial flagging would be passed to either GPT-3.5 or GPT-4 for further analysis, and that model could then either revoke or keep the flagging in place. That would save the user time and energy to go through the appeal process over nothing. And/or add a feature where users can opt into a non-playpen version of a service that they're most likely paying for. It's just ridiculous. To add to the confusion, most of the times I've gotten flagged, the model itself (usually GPT-4) has stated that it didn't find anything against the platform's guidelines upon revising what was said. So the flagging system is really not very sophisticated when even GPT-3.5 often gets it that there was no violation of any sorts taking place. One could argue that OpenAI is now a single Silicon Valley company that has such an advantage over their competition worldwide, hence I do find that culturally quite problematic as well -- a single company one-sidedly deciding what type of language a human can input to an AI? OpenAI's stated goals about an AI that "benefits all of humanity", all the talk about diversity and inclusivity etc don't seem to meet up with the reality of how their platform still flags the user input often out of nowhere. How can you benefit anyone when the user can't even go through transcripts without running the risk of getting flagged? Their current guardrails are ripe for a thorough sanity check. |