[ad_1]
Nearly two months since the release of GPT-5, an update to ChatGPT that was supposed to give it “PhD level” intelligence and bring it once step closer to the industry’s vaunted goal of artificial general intelligence (AGI), the OpenAI chatbot is still going bizarrely haywire over simple and completely innocuous inquiries.
On the ChatGPT subreddit, fans of the AI bot noticed that asking it if there’s an “NFL team whose name doesn’t end with the letter ‘s’” sends it into a long-winded meltdown, allowing you to witness its superpowered “reasoning” abilities turn to mush in real time.
“Yes — there are two NFL teams whose names don’t end with an ‘s,’” ChatGPT says, before proceeding to list two teams that do. “Miami Dolphins? ❌ end with s. Green Bay Packers? ❌ ends with s.”
It’s an odd way of addressing the question. ChatGPT correctly identifies that those names do end with s, but why bring them up?
Then it looks like it’s cutting to the chase — at least, that is, until it goes off the rails yet again.
“The only two teams that don’t end with ‘s’ are: Miami Dolphins ❌ no (wait, ends with s),” it says, catching its mistake. “Hold up, let’s do this carefully. Actually, the correct answer is: Washington Commanders ❌ ends with s. Chicago Bears ❌ ends with s.”
In the original example uploaded to Reddit, ChatGPT goes on like this for several more paragraphs. And it never arrives at the correct answer — that there aren’t any teams that don’t end in an “s.”
Like a high schooler hitting a word count, it peddles irrelevant details while teasing a conclusion. It also peppers in phrases to make it sound like it’s actually doing some deep thinking. “Hold up, let’s do this carefully,” it says. Or “let me do this systematically.”
“The actual correct answer is,” ChatGPT says at one point, not realizing the shtick is getting old.
Eventually, it promises “the correct answer (for real this time).” It says it will list “two teams” that don’t end with “s” — before listing an additional three teams that do.
Other users posted examples where ChatGPT eventually gives the correct answer, but only after stringing the user along a similarly delirious spiel. In our testing, it produced similar bizarre results.
This is far from the first time the chatbot has been foiled by a simple question — or even melted down in such an incredibly circuitous manner.
Earlier this month, for instance, fans noticed that asking it if a mythical seahorse emoji existed sent it spiraling into a crisis of logic. Despite the aquatic creature never being part of the official emoji dictionary, ChatGPT insisted it was real, exemplifying the absurd lengths AI is willing to go to to please the user. What’s bending a few facts if the AI gets to come off as personable and human-like, convincing users that they should come back for more?
Sycophancy probably isn’t the only factor to blame. GPT-5 is actually a tag team of a lightweight model for basic prompts and a heavy duty “reasoning” model for tougher questions. What’s probably going on here is that the lightweight model is getting stuck with a question it can’t really handle, instead of handing it off to its smarter cousin. This often malfunctioning dynamic is part of the reason why fans were left feeling disappointed — and in many cases, furious— with GPT-5’s launch (which was only exacerbated by OpenAI cutting off access to the old models that its customers had grown attached to, a decision it soon reversed.)
In any case, it’s a pretty thin excuse. If the AI needs to bust out its biggest guns to answer such a simple question, then maybe it’s not on the fast-track to surpassing human intelligence.
More on AI: ChatGPT Is Blowing Up Marriages as Spouses Use AI to Attack Their Partners
[ad_2]
Source link
