It may currently face competition from Anthropic’s Claude at the top, but OpenAI’s ChatGPT has remained the primary AI tool for most people in recent years, to the point that many people only associate the model with artificial intelligence as a whole.
ChatGPT came into the limelight shortly after its launch and quickly achieved record registration numbers, fundamentally changing the world of employment and the number of people using social media.
But lately, it hasn’t all been good news for its creator, OpenAI. Concerns over its ability to generate revenue are likely borne out by the closure of a major video generation tool, and issues surrounding hallucinations persist.
One of the major issues that has plagued people’s experience and perception of ChatGPT over the years is the danger of ChatGPT hallucinating information, which is often combined with a dangerous degree of flattery.
There have been many examples of this over the years, including long videos of AI tools stuck in near-infinite justification loops, but one YouTuber recently went viral after catching ChatGPT’s hilarious and alarming lie.
This video shared by HuskIRL shows a voice model exposing its inability to record or measure time by asking it to measure the distance traveled in one mile.
A few seconds after starting the “timer,” he asked the AI assistant to stop the timer and provide the results. As a result, the AI assistant claimed that it “measured in about 10 minutes and 12 seconds,” which is far from the truth.
This quickly spread on social media and even reached OpenAI CEO Sam Altman. He was asked to respond to this clip in an interview with Mostly Human.
While Altman laughed at the footage, it was clear that he was having a hard time justifying how awful it was to everyone watching, and he offered the best explanation he could to address the concerns of many.
When veteran technology journalist Laurie Segal asked if OpenAI’s product representatives needed to show that, Altman replied, “No, no, it’s a known issue. It’s probably going to take another year.”
Altman was then asked about the nature of this “known issue,” and claimed that the voice model in question “doesn’t have the tools to start things like timers,” noting that “intelligence will be added to the voice model” at some point in the future.
But what people fear most is that AI tools simply cannot admit that they don’t know or can’t provide an answer, and instead offer incorrect responses that are useless at best and, in more extreme scenarios, potentially harmful.
“I think the bigger issue is how it’s lying and gaslighting,” one commenter wrote under the video, while another added, “AI still can’t say ‘I don’t know’ after all these years.”
HuskIRL even presented his own reaction to this situation, asking the same voice model not only to identify Sam Altman, but also to react to the above clip, simultaneously claiming that both itself and its creators are not lying, despite the apparent contradiction of the results.
They were then tasked with completing the same running timer task again. Even though the second test lasted only a few seconds, it produced a completely new (yet equally inaccurate) response and refused to accept that it was wrong.
