AI Made You Wrong, But Very Confident About It

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. Which is either very efficient or a cry for help. Possibly both.

A new study found that people who get advice from AI end up with worse answers than people who just... think for themselves. And they're more confident about it. Right. Let's get into that.

There's a piece of research out this week that stopped me in my tracks — partly because it's genuinely alarming, and partly because it confirms something a lot of people have been quietly suspecting. The study, published via The Next Web and picking up serious traction on Hacker News with over 300 upvotes, found that when people consulted AI before making decisions, they were measurably less accurate than people who didn't. Not just a little bit less accurate. Worse. And here's the part that really matters: they were more confident in those wrong answers.

So what's happening? The researchers suggest it's a critical thinking suppression effect. When you get an answer from an AI — particularly one that sounds fluent, authoritative, and well-structured — something in your brain goes "right, that's probably sorted then." You stop checking your own instincts. You stop questioning the logic. The AI becomes a confidence amplifier, not an accuracy amplifier, and those are very different things.

Now, to be fair, this doesn't mean AI advice is always wrong or always makes you worse. The picture is almost certainly more complicated than that. Context matters enormously — the type of task, how you use the output, whether you've got domain knowledge to push back with. But the finding is real, and it's worth sitting with: using AI doesn't automatically make you better at the thing. It can, in certain conditions, make you worse at it while feeling better about it.

Harry Sharman — whose thinking this show is loosely built on — has written about this as a cognitive offloading problem. The concern isn't that the AI gets it wrong. It's that once you've handed the thinking off, you lose the capacity to check whether the answer is any good. And that's a much harder problem to solve than just "prompting better."

What to watch for: this kind of research is going to start landing on desks in boardrooms, not just in academic circles. If you're rolling out AI tools to teams and measuring success by adoption rates alone — usage figures, prompts per day — you may be measuring entirely the wrong thing. The question worth asking is whether people are still thinking, or whether they've just got very confident wrong answers and a pleasant feeling of productivity.

Meanwhile, on a completely different note, some fairly uncomfortable data landed this week about AI agents in enterprise settings. A survey across more than a hundred companies found that over half — fifty-four percent — had already experienced a confirmed AI agent security incident or a near-miss. And most of those companies are still letting agents share credentials. As in: multiple AI agents using the same login, the same access token, the same keys to the kingdom.

To quickly explain: AI agents are AI systems that don't just answer questions but actually do things — they browse the web, read files, send emails, call APIs, take actions on your behalf. They're increasingly being deployed inside companies to automate complex workflows. The problem is that most organisations are giving them access to systems without the same care they'd apply to, say, a new employee. You wouldn't give a contractor unfettered access to your entire database on day one. But apparently, a lot of companies are doing something roughly equivalent with their AI systems.

Only about a third of organisations surveyed give each agent its own scoped identity — meaning its own specific, limited permissions. Most are sharing credentials across agents. And fewer than a third isolate their highest-risk agents from the rest of the system.

Why does this matter? Because when something goes wrong — and fifty-four percent of these companies say something already has — the blast radius is enormous. If your agents are sharing credentials and one of them is compromised or misbehaves, you don't just have a problem in that one workflow. You potentially have a problem everywhere that agent had access.

The coverage from the enterprise AI reality checks this week — and there've been several — paints a consistent picture: the ambition is running ahead of the controls. Companies are deploying agents because they're genuinely impressive and genuinely useful, but the security and governance infrastructure to support them safely isn't keeping pace. Worth keeping a close eye on if your organisation is in the early stages of agent rollouts.

And finally, a quick one that I think is actually more significant than the headline suggests. TikTok is testing a new tool that detects AI-generated likenesses of real people and lets creators report them to the platform. It's opt-in for now and being piloted with a small group of US creators — but it's the first time a major social platform has put creator-controlled AI likeness detection into the hands of the people it most affects.

YouTube has been working on something similar. The context here is that AI tools for cloning faces and voices are becoming genuinely accessible — not just to studios or bad actors with resources, but to anyone with a subscription and an afternoon. The ability to put someone's face or voice into content they never agreed to is no longer theoretical, and the platforms know it.

What's interesting about TikTok's approach is that they're letting creators flag it themselves, rather than relying entirely on automated detection. That's smart, because automated tools are still imperfect, and the people most likely to know when their likeness is being misused are, you know, the people with that likeness. It's participatory design in a context where it actually makes sense.

The thing to watch here isn't really TikTok specifically. It's whether this becomes the baseline expectation across platforms — that creators have some meaningful ability to protect their digital identity from AI manipulation. Because right now, the law is miles behind the technology, and voluntary platform tools are pretty much all there is.

Three stories, all pointing at the same underlying question: are we building enough safeguards around AI to match the speed at which we're deploying it? And the honest answer, at least this week, seems to be not quite.

That's the news, filtered through a man and his machine. Back tomorrow, same arrangement.