The question BBC Tech Life put to William Tunstall-Pedoe this week was: when AI gives you an answer, do you actually trust it?

It’s a timely question. As AI development accelerates and these tools become part of how we work, communicate and make decisions every day, you’d expect confidence in them to be growing. And yet the honest answer is: not really.

UnlikelyAI’s latest research found that employees now spend almost as much time checking AI’s answers as they do using the tools themselves. Among business leaders, only 41% say they trust AI completely and this is coming from the people deploying AI directly.

Two types of software, one unsolved problem remains

The tension at the heart of modern AI is easier to understand than the industry makes it sound.

Traditional software, such as spreadsheets, databases, algorithms has proven itself to be reliable and explainable. You trust it because when it calculates something, it’s right, you know what you have put in and what the system knows and when it can’t calculate something, it tells you.

The second type of software is machine learning: the statistical models behind image recognition, language understanding, the breakthroughs of the last few years that have genuinely changed what computers can do. They operate differently in principle and you never know how it arrived to the output it gives you.

“”It’s statistical. It’s wrong some of the time. It can’t explain how it gets results in any way that makes sense to people.””

These models are always making a sophisticated guess. Even the best are right 87%, 92%, 99% of the time, never 100%.

Which raises the question William put to the BBC audience directly:

“”Have we accidentally rewarded AI for sounding confident rather than being correct?””

Trained on data that is itself expressed with confidence, these models learned to perform certainty whether they’ve earned it or not. That’s the origin of hallucinations: invented facts delivered like established truth. and every time it happens, trust erodes a little more.

What happens next in AI?

This is the problem UnlikelyAI was built to solve using Neurosymbolic AI technology.

Neurosymbolic AI is the idea of blending both worlds, the neuro (the capable, statistical, machine learning layer) with the symbolic (the reliable, explainable, traditional software layer). The goal isn’t to choose between them, it’s to get the best of both: the remarkable capabilities of modern AI, combined with the integrity of software that can show its work and say I don’t know when it genuinely doesn’t.

In the short-term it is important to understand the limits, and design around them. In the BBC interview, William mentions that the real solution is technological, and that it’s achievable.

“”Long term it’s about making this software work really robustly and really reliably and really explainably — so it can be used in a way that’s completely trustworthy.””

That’s what we’re working toward, AI that doesn’t just perform well in controlled conditions but that businesses can stake real decisions on.

The work is hard, the problem is real, and the timing has never been more right.

Want to understand how neurosymbolic AI could change what’s possible for your business? Get in touch.