Shhhh. Not at the table Aunty.

You’re like, ‘not to brag, but I have 4 THOUSAND followers on BlueSky.’
Ummm, aunty?

🤣

Ok thank you. This is exactly what I needed. ❤️

Wait, why? Is there an actual link?

Thanks for clearing this up for me - that he’s known for years and his latest is a PR move to save face. I’d heard the scandal, but only knew his side. ☺️

When my friend was a teen he was on Teen Week on Wheel of Fortune. He got all the way to the final round and came in 2nd. Pat Sajak was like ‘Don’t worry, you win a free trip to Anaheim.’ Not Disney tickets mind you, just Anaheim. On air my friend was like ‘ Yeah I live in Anaheim.’ 🤣

Oh wow. I really like this.

So no, the AI you and I use isn’t “just” a Clever Hans—but the story is a good reminder that humans, horses, and machines can pick up the right answers for the wrong reasons, especially early in the learning process.

But here’s the important distinction: unlike Hans, modern AI can generate entirely new outputs rather than simply mimic what it has seen. And now that researchers understand the Clever Hans effect, they design tests to prevent models from learning the wrong cues.

There’s even a term for it in psychology and AI research: the Clever Hans effect. One famous example is an AI trained to detect pneumonia from X-rays that actually learned to key in on chest tubes and labels that just happened to appear in many pneumonia cases.

So how does this apply to AI? Since LLMs and other models learn by detecting patterns in massive amounts of data, some argue they’re doing what Hans did—getting the right answer for the wrong reason. And the truth is… sometimes they are.

The mystery wasn’t solved until psychologist Oskar Pfungst discovered that Hans wasn’t doing math at all. He was reading subtle shifts in his questioners’ posture and expressions. In other words, he was clever, but at reading human body language, not numbers.

Scholars, philosophers, business leaders—everyone tested him—and he answered close to 90% of their questions correctly. Square roots, dates, distances… he’d tap his hoof the correct number of times, like tapping three times for the square root of nine.

Hahaha

Smooth, bald head, flamboyant suit jackets.

I came here to say something similar. In consulting, the customer will just continue asking for more until you say no. How do you say no? Point to the scope.

Bottom line, Musk’s little maneuver exposed the rot we all knew was there. Tech companies absolutely can fix this, they just don’t want to.

How do we get rid of them? Musk cut human moderation to the bone. Bring back a small team to deal with state propaganda, CSAM, threats of violence. Community notes is a good start but incentivize good behavior and honest accts. You’re already paying influencers, why not the notes team?