Browse Comments — Raw (as collected)

This nails why "confidence" is the wrong metric for AI. Humans signal uncertainty with hesitation; models don't have that tell. The skill we're all quietly developing now isn't prompting it's knowing when to distrust a fluent answer.

One of AI's most impressive skills is delivering a completely wrong answer with the confidence of someone who already has a TED Talk scheduled about it. 😄 My favorite cases are when it invents a fact, doubles down on it, and then smoothly transitions into life advice. “The answer is incorrect, but have you considered mindfulness and a healthy work-life balance?” That's when you know the conversation has truly evolved.

Wrong formula. Confident delivery. And when I objected, a calm reframe that maybe I was the problem. The AI is in its situationship era and I walked right into it.

It’s worth noting something about probabilistic intelligence that confidence is not a proxy for correctness. A good skill going forward is knowing when the system is guessing confidently versus actually grounded in something verifiable.

Summarising historical information based on meeting material and minutes and then concluding what decision was made in a meeting happening 2.weeks into the future and when challenged providing arguments as to why it would be the right decision. Not that it was predicting decisions in the future. Use AI but don't blindly rely on it....its good but not yet excellent.

The most dangerous thing about AI isn’t wrong answers… it’s how politely they’re delivered 🤣

Maybe the real AI breakthrough will be teaching it to say: “I don’t know.”

AI can reduce the time spent finding information. Deciding what to trust remains a human responsibility.

😁

Love this 😂

The thing isn't that AI gets things wrong occasionally, it's that it can sound equally convincing when it's right or wrong.

absolutely what I face each time i found a mistake. AI would admit the mistake profusely and not care about the consequence of the previous advice.

User error. AI just didn't say, you must not be the slowest person running from the bear.

AI has mastered one very human trait: refusing to admit it doesn't know. 😄 The best results still come when confidence is paired with verification. "Trust, but verify" remains a surprisingly effective AI strategy.

This is why "I'm not sure" is an underrated AI feature. A model that admits the gap beats one that papers over it smoothly.

Which is why it should only be a tool, and not a sole source of reliance. AI, when used incorrectly, will cause mayhem….and already has to be fair, let alone weaken cognitive ability and capacity. I use AI, yes… I would be a fool not to. It’s a fabulous tool, but being aware of the limitations is a must.

I find that answering "are you sure? Feel free to say yes but only if you can ground your reasoning in facts" tends to yield a better final answer

The confidence is impressive until you already know the answer. Then it becomes a reminder to double check important details.

Prompting is very important in AI. If he had asked "what is the best course of action when a bear is behind you" results would have been different!

It always tries to explain why it’s right.