I was trying to defuse a fight between two people I care about, and reached for a quote that might help them pause:

“Holding onto anger is like drinking poison and expecting the other person to die.”

It’s widely shared, usually credited to the Buddha, sometimes to Mandela. I asked three AI models where it actually came from.

Grok said Seneca. Confidently.

ChatGPT hedged — maybe Seneca, maybe not.

Claude said no. Seneca wrote about anger being self-destructive, but the “poison” line has no traceable ancient source. It’s a modern aphorism.

One question, three answers, and only one that actually checked.

When Confidence Sounds Like Fact

Here’s what matters: if this happens on a quote — cheap to check, low stakes, nobody hurt if Seneca gets wrongly credited — it happens just as easily on a citation in a grant report, a compliance claim in a client proposal, or a “fact” dropped into a board memo. The mechanism is identical: a sentence that sounds right, delivered with total confidence, standing in for a sentence that’s actually been verified.

Grok didn’t guess. It didn’t say “probably” or “commonly attributed to.” It stated a false attribution with the same tone it would use for a true one. That’s the real story here — not that AI is wrong, which is old news, but how it’s wrong. Fluency and accuracy are different things, and AI is very good at the first. Confident phrasing is not evidence. It’s just phrasing.

Seneca did say something real and well-documented on this theme: “The greatest remedy for anger is delay.” That’s from Book II of De Ira, written around 45 CE. True claim, correct source — but less catchy than the poison line, which is probably why the poison line won the internet and Seneca’s actual words didn’t.

Now Put That Mistake in a Security Report

Picture the same failure mode in a security context: an AI-drafted compliance report cites a control standard’s requirement — a retention period, an encryption threshold, a reporting deadline — with total confidence. Nobody checks, because it reads like something that was checked. Six months later an auditor does check, and the number was wrong. That’s not a hypothetical category of risk. It’s the poison quote, wearing a different outfit, in a document that actually matters. Unverified information that slips into security documentation or system configurations doesn’t just embarrass you — treated as fact, it leads to weak controls, missed vulnerabilities, decisions built on bad data. Attackers exploit exactly that kind of gap.

Trust, but Verify

The fix isn’t distrust. It’s discipline. Treat AI output as a draft that wants a source, not an answer that already has one. Ask which model said it, ask why, and if it matters — check.

Even when the subject is anger management.

Want to learn more? AI has been a subject of my writing for several years, and CGNET has offered AI user training and implementation for both large and small scale organizations.   I would love to answer your questions! Please check out our website or drop me a line at g.*******@***et.com.