OpenAI has been enjoying the limelight this week with its incredibly impressive Sora text-to-video tool, but it looks like the allure of AI-generated video might’ve led to its popular chatbot getting sidelined, and now the bot is acting out.
Yes, ChatGPT has gone insane–- or, more accurately, briefly went insane for a short period sometime in the past 48 hours. Users have reported a wild array of confusing and even threatening responses from the bot; some saw it get stuck in a loop of repeating nonsensical text, while others were subjected to invented words and weird monologues in broken Spanish. One user even stated that when asked about a coding problem, ChatGPT replied with an enigmatic statement that ended with a claim that it was ‘in the room’ with them.
Naturally, I checked the free version of ChatGPT straight away, and it seems to be behaving itself again now. It’s unclear at this point whether the problem was only with the paid GPT-4 model or also the free version, but OpenAI has acknowledged the problem, saying that the “issue has been identified” and that its team is “continuing to monitor the situation”. It did not, however, provide an explanation for ChatGPT’s latest tantrum.
This isn’t the first time – and it won’t be the last
ChatGPT has had plenty of blips in the past – when I set out to break it last year, it said some fairly hilarious things – but this one seems to have been a bit more widespread and problematic than past chatbot tomfoolery.
It’s a pertinent reminder that AI tools in general aren’t infallible. We recently saw Air Canada forced to honor a refund after its AI-powered chatbot invented its own policies, and it seems likely that we’re only going to see more of these odd glitches as AI continues to be implemented across the different facets of our society. While these current ChatGPT troubles are relatively harmless, there’s potential for real problems to arise – that Air Canada case feels worryingly like an omen of things to come, and may set a real precedent for human moderation requirements when AI is deployed in business settings.
As for exactly why ChatGPT had this little episode, speculation is currently rife. This is a wholly different issue to user complaints of a ‘dumber’ chatbot late last year, and some paying users of GPT-4 have suggested it might be related to the bot’s ‘temperature’.
That’s not a literal term, to be clear: when discussing chatbots, temperature refers to the degree of focus and creative control the AI exerts over the text it produces. A low temperature gives you direct, factual answers with little to no character behind them; a high temperature lets the bot out of the box and can result in more creative – and potentially weirder – responses.
Whatever the cause, it’s good to see that OpenAI appears to have a handle on ChatGPT again. This sort of ‘chatbot hallucination’ is a bad look for the company, considering its status as the spearpoint of AI research, and threatens to undermine users’ trust in the product. After all, who would want to use a chatbot that claims to be living in your walls?