Can We Trust AI: It Sounds Confident But Is It Right?

From writing reports to summarising meetings and building persuasive arguments, AI’s capabilities keep growing. Yet one question persists: can we trust AI?

That’s the million-dollar question, and it’s not as simple as a “yes” or “no”.

AI is persuasive, it’s fluent, and it’s fast. And one of AI’s superpower is sounding convincing. The problem is, it sounds just as confident even when it’s wrong.

I asked ChatGPT for guest ideas for a podcast on virtual training. The list was full of impressive names… it’s a shame they were all imaginary. Even ChatGPT admitted it had just made them up as ‘examples’ when I challenged it.

It also insisted that I’m married to Simon Nicholson…  my husband, John, found that one half hilarious and half worrying 🤣

Why does AI sound so convincing?

Because that’s literally what it’s built to do.

Large language models like ChatGPT and Copilot don’t actually “know” things. They string together the most statistically likely next words based on patterns in data. The output feels natural and authoritative but it’s really more of an advanced autocomplete.

According to a recent MIT article, people tend to over-trust AI precisely because of that fluent confidence. When something looks polished and human-like, we instinctively assume accuracy, even when it’s fabricated.

And those fabrications are getting more sophisticated. We know that generative AI has reduced much of its more obvious “nonsense”. But these much more subtle inaccuracies blend in so well that many of us don’t spot them.

Want to see how easily it happens? Try our Spot-the-Flaw: Your AI Thought Partner tool and test your own AI outputs for accuracy and flaws.

Why do we trust it so easily?

It’s simple: we’re wired for shortcuts. If something sounds logical, we assume it is.

Psychologists call this “cognitive ease”. It’s the bias that makes familiar, fluent information feel truer. Combine that with AI’s authoritative, confident tone and you’ve got a great recipe for misplaced confidence.

KPMG’s Global Study on Trust and AI (2025) found that 66% of respondents in the study reported using AI output without evaluating the accuracy.

And let’s be honest: we want to trust it. It saves us time. It feels like progress. Questioning it slows us down…  but that’s exactly why we need to.

What’s the risk of copying without checking?

At best, you’ll end up sharing something that’s just a bit off. At worst, you’ll repeat false or biased information that damages your credibility or worse.

Remember the US lawyers who let ChatGPT write their legal brief? It confidently invented judges, rulings – the lot. It looked fine until the court fact-checked it. Painfully avoidable (and deeply awkward!).

Research from Harvard’s Misinformation Review: “The Hallucination Challenge in Generative AI” explains why this happens. AI fills gaps in its knowledge with the most statistically “fitting” answer. It’s not lying – it’s guessing with confidence.

At work, the same risk applies to every AI-generated slide deck, report, or client email. One dodgy “fact” can snowball into a brand headache or credibility nightmare.

So, can we trust AI? Not automatically. We need to check it before we trust it.

How can we sense-check what AI gives us?

That’s where our LEAP model here at VTT comes in. This is a four-step framework that helps you separate solid insights from smooth nonsense and keeping human judgement in the loop.  It works by testing any AI outputs against these four areas:

• Logic – Does this actually make sense?
• Evidence – What’s backing it up?
• Attribution – Where’s it from? Can you trust the source?
• Perspective – What might be missing or biased?

This approach echoes findings from Frontiers in Psychology’s “Developing Trustworthy Artificial Intelligence”, which highlights that genuine trust in AI depends on transparency, traceability, and human oversight – exactly what LEAP helps people apply day-to-day!

Run any AI output through those four lenses and you’ll catch most of the problems before they reach anyone else. It’s quick, practical, and easy to use across teams.

You can explore LEAP in our Critical Thinking for AI workshop, which is designed to help people work with AI rather than outsourcing their thinking to it.

How can you build this habit at work?

Start small. Pick one thing this week that you have created using input from AI. It could be a report, an email, a presentation etc, and put it through LEAP before you share it. You’ll be surprised what you spot when slowing down for thirty seconds and thinking.

If you want a quick shortcut, paste your AI output into Spot-the-Flaw GPT and see what it throws back. Treat it like a sparring partner, not a safety net.

With a bit of practice, that sense-check becomes instinctive. You won’t need a framework because you’ll just feel when something’s off.

Finally, consider asking your team the question:  “Can we trust AI?”. This can help to drive a conversation that raises awareness of the dangers of not checking our outputs, as well as providing space to share ideas on how to overcome this (hopefully using the ideas you’ve gained from this article).

So, what changes when we think with AI, rather than outsourcing our thinking to AI?

When we stop seeing AI as a master and start treating it as a thought partner, everything gets better:

• Our work becomes faster and more accurate.
• Our ideas stay human but are fuelled by new perspectives.
• Our teams gain confidence in challenging what the tech suggests.

The smartest professionals aren’t the ones who trust AI completely…. they’re the ones who verify, adapt, and own the final output. So next time AI gives you something that looks perfect, don’t just nod and say thank you. Ask, “Does this actually make sense?”

In summary

AI is brilliant, but it can still bluffing with confidence. Trust it by testing it. Partner with it by challenging it. And remember: critical thinking is still our best competitive edge.