#ai-hallucination

#ai-hallucination

One of the strangest things about large language models is not what they get wrong, but what they assume to be correct. LLMs behave as if every question already has an answer. It's as if reality itself is always a kind of crossword puzzle. The clues may be hard, the grid may be vast and complex, but the solution is presumed to exist. Somewhere, just waiting to be filled in.

But like everything else in life, there will always be a more powerful AI waiting in the wings to take out both protagonists and open a new chapter in the fight. Acclaimed author and enthusiastic Mac user Douglas Adams once posited that Deep Thought, the computer, told us the answer to the ultimate question of life, the universe, and everything was 42, which only made sense once the question was redefined. But in today's era, we cannot be certain the computer did not hallucinate.

All the way back in 2019, an article in The Wall Street Journal warned readers that AI has learned to write fake news stories. One of the tools highlighted in the article was GPT-2, a precursor to what we now know as the game-changing tech that is ChatGPT. Fast forward to 2025, and we now have AI integrated directly within the Google Search experience. Now, when you look up something on Google Search, you will see an AI Overview at the top of the page.

But according to 404 Media, in a series of deleted X posts, Grok boasted that Musk had the potential to drink piss better than any human in history, that he was the ultimate throat goat whose blowjob prowess edges out Trump's, and that he should have won a 2016 porn industry award instead of porn star Riley Reid. Grok also claimed Musk was more fit than LeBron James.

For those of us scouring filings for questionable AI screw-ups though, we now zoom to a handwritten insert included with the order, justifying the decision to allow the motion even if it technically missed a deadline based on Jones v. Goodman, 57 Cal.App.5th 521, where the court writes, that an amended motion should relate back to the initial motion "as long as the initial motion was in 'substantial compliance' with the governing rule."

Starbuck's claims against Google came after he filed a similar lawsuit against Meta, whose AI he claimed falsely asserted that he'd participated in the January 6th riot at the US Capitol. But Meta settled that lawsuit in August and even hired Starbuck as an advisor to help address "ideological and political bias" in its AI chatbot, The Wall Street Journal reported. The outlet noted last month that so far, no US court had awarded damages for defamation by an AI chatbot.

Looming over the proceedings even more prominently than the judge running the show were three tall digital displays, sticking out with their glossy finishes amid the courtroom's sea of wood paneling. Each screen represented a different AI chatbot: OpenAI's ChatGPT, xAI's Grok, and Anthropic's Claude. These AIs' role? As the "jurors" who would determine the fate of a man charged with juvenile robbery.

In one case, according to Starbuck, Google's AI claimed he had been a person of interest in a murder case when he was just two years old. For each source, Google's AI provides a URL, giving the impression that these are real news articles with headlines like, Robby Starbuck Responds to Murder Accusations,' he said. The only way to discover that these URLs are fake is to click on them.

"Um ... there we go ... uh-oh," said Meta CEO Mark Zuckerberg on stage as he attempted to answer a video call through a combination of movements between a wristband and a pair of glasses. "Well, I ... let's see what happened there ... that's too bad," he continued, shortly before cutting short the live demo. The video call went unanswered.

A few months ago, I asked ChatGPT to recommend books by and about Hermann Joseph Muller, the Nobel Prize-winning geneticist who showed how X-rays can cause mutations. It dutifully gave me three titles. None existed. I asked again. Three more. Still wrong. By the third attempt, I had an epiphany: the system wasn't just mistaken, it was making things up.

Popular right wing influencer Charlie Kirk was killed in a shooting in Utah yesterday, rocking the nation and spurring debate over the role of divisive rhetoric in political violence. As is often the case in breaking news about public massacres, misinformation spread quickly. And fanning the flames this time was Elon Musk's Grok AI chatbot, which is now deeply integrated into X-formerly-Twitter as a fact-checking tool - giving it a position of authority from which it made a series of ludicrously false claims in the wake of the slaying.

"Language models are optimized to be good test-takers, and guessing when uncertain improves test performance," the authors write in the paper. The current evaluation paradigm essentially uses a simple, binary grading metric, rewarding them for accurate responses and penalizing them for inaccurate ones. According to this method, admitting ignorance is judged as an inaccurate response, which pushes models toward generating what OpenAI describes as "overconfident, plausible falsehoods" -- hallucination, in other words.