When the world's most-used AI confidently hallucinates a dead comedian sexting a millennial writer, "learning to use AI" starts to sound less like empowerment and more like damage control.
The Summary
- ChatGPT falsely claimed Don Rickles tried to sext Lena Dunham in 2012 when asked to identify an unnamed celebrity from her new book, exposing how confidently AI fabricates answers to questions it can't actually answer.
- The incident highlights the gap between "learn to use AI" advice from figures like Reese Witherspoon and the practical reality that most users can't distinguish confident AI hallucinations from fact.
- As AI becomes the default research tool, the ability to verify answers matters more than the ability to write better prompts.
The Signal
A Business Insider writer asked ChatGPT to identify which celebrity allegedly tried to text Lena Dunham late at night in 2012, a detail mentioned but left anonymous in Dunham's recent book. ChatGPT responded with total confidence: Don Rickles. The comedian who died in 2017 at age 90, known for roasting Frank Sinatra in the 1960s, supposedly sent flirty texts to a 26-year-old in 2012 when he was 86 years old.
The writer knew this was wrong. But here's what matters: she had to already know it was wrong to catch it. The AI didn't hedge. It didn't say "I don't have access to that information." It manufactured a specific, plausible-sounding answer and delivered it with the same authoritative tone it uses for actual facts.
"What level of 'using AI' is expected here to stave off falling behind in the workforce and life in general?"
This happened in the context of Reese Witherspoon's viral moment telling women to "learn to use AI," which sparked enough backlash that she issued a follow-up explanation. The advice itself is standard corporate futurism: adapt or fall behind, AI is the new literacy, etc. But the Rickles hallucination exposes the practical problem with that framing.
Most AI training focuses on prompt engineering, writing better queries, using AI to draft emails or summarize documents. Almost none of it focuses on the harder skill: knowing when the AI is lying to you. Because that requires domain knowledge the AI is supposed to replace.
The breakdown:
- ChatGPT doesn't distinguish between "I know this" and "I can generate a plausible answer"
- Users increasingly treat AI outputs like Google results, factual until proven otherwise
- The gap between AI capability and user verification skills is widening, not closing
The writer was using ChatGPT as "a sort of super Google" to find information she vaguely remembered. This is exactly how millions of people use these tools now. And it works great until it doesn't. The problem is you can't tell which category any given answer falls into without external verification.
The Implication
The real AI literacy isn't learning to write better prompts. It's developing a nose for hallucinations and the discipline to verify before you internalize an answer. That's harder than prompt engineering and scales worse. You can teach prompt structure in an afternoon. Teaching someone to spot confident fabrications requires domain expertise in whatever they're asking about.
Watch for companies building verification layers on top of LLMs. Citation systems that link to actual sources, confidence scores that admit uncertainty, tools that flag potential hallucinations. The current model treats all outputs as equally valid, which works fine for creative writing and terrible for everything else. The market will correct this, but until it does, Don Rickles' reputation is apparently fair game.