If Your Employee Acted Like AI, You’d Have Sacked Them by Now

Imagine you’ve just hired someone new, and most days they’re efficient, knowledgeable, articulate and confident. They’re everything you hoped they’d be, demonstrating many of the qualities of a great employee.

Now imagine that every now and then, when faced with a question they don’t know the answer to, they just make something up. It’s fluent, convincing, so at first you take it at face value. But when it has consequences for the business, you come back to that employee, challenge them – and they can’t explain their thought process or share their workings. You might even explain to them that they’re wrong, to be met with insistence to the contrary and more fabrications to try to back it up. After the second or third time this happens, would you likely have fired them, or at the very least given that employee a warning? The answer for most employers is almost definitely yes.

I find it baffling that businesses now hand critical workflows and decision-making over to an employee that is known to exhibit this exact behaviour. The only reason it goes unchecked is because this ‘employee’ is software rather than a person.

Why AI Doesn’t Fail the Way People Do

There’s a common, incorrect assumption among businesses deploying AI that if the technology has an error rate, it’s roughly the same problem as human error and can be managed the same way. But checks, reviews, and escalation don’t always work with large language models (LLMs). They’re made to sound like humans, but look at what’s under the hood, and you’ll see they don’t operate like us at all.

If a star employee makes a mistake, we’d assume that the error was in earnest – they were doing their best and slipped up. It happens. If they deliberately invented an answer to get away with not knowing something, and they did this repeatedly, we would see it as misconduct. Managers know where to focus their oversight, because difficulty and error margins coincide. They can ask the employee to explain their reasoning and question them on it. The employee usually knows when they’re out of their depth, and says so. And after one error, they learn, so the same mistake doesn’t repeat in a constant cycle.

AI systems invert every one of these assumptions. Their default response to failure is fluent, confident invention. They’ll produce a plausible answer that is simply untrue, which is the exact behaviour that would end a human career. LLMs are made to produce the most statistically likely continuation of a piece of text, and when they don’t know the answer, they frequently produce something that sounds like it could be right anyway. OpenAI’s own researchers have acknowledged that some level of factual error is a statistical inevitability of how these models work, and that training rewards a confident guess over an honest admission of uncertainty.

Where Neurosymbolic Approaches Come In

This is precisely what neurosymbolic approaches are designed to fix. They pair neural networks, which are great at understanding messy, real-world language and data, with deterministic, rule-based reasoning. This means that a system’s conclusions follow from explicit logic that can be inspected, rather than emerging from a process no one can see inside of.

Without this added layer of logic that neurosymbolic technology provides, AI’s errors don’t correlate with the difficulty of the work the way those of a person do – they’re unpredictable. They can appear on trivial tasks while difficult ones sail through, leaving it difficult to discern where best to focus human oversight. Their explanations, when offered, are generated after the fact rather than providing a true account of how they reached a conclusion. This same mistake can recur indefinitely, because there’s no one inside the system to learn from it, or to be held accountable for it.

Trust Without Understanding

Researchers have documented that when people are new to a system, they have a tendency to abandon it after a single visible error; but when it’s a familiar one they swing to the other extreme – accepting its outputs simply because they have become accustomed to them, not because those outputs have become more reliable. What’s rarely seen is the right level of trust, because that depends on people understanding when an AI system is likely to be wrong and being able to recognise when it is.

That doesn’t mean AI shouldn’t be trusted with important work. It means we’ve been willing to give it a kind of trust we wouldn’t extend to a person, partly because there is no obvious individual to hold accountable when it fails.

If AI is going to take on judgement-based work in the way a valued colleague would, it needs to meet the same basic standards. Answers need to be verifiable, with reasoning visible, and accept the fact that sometimes, the best thing it can do is say “I don’t know”.