Who Do You Trust?

There is a particularly comforting argument circulating among those who remain suspicious of artificial intelligence. It goes something like this: AI cannot be trusted because it makes mistakes, invents facts, fabricates references, and occasionally produces complete nonsense with breathtaking confidence. Therefore, we should be extremely cautious about using it.

Fair enough. But I have a question.

Have these people ever met another human being?

Apparently, somewhere along the way, we decided that unreliability was a uniquely technological problem. Humans, particularly those equipped with impressive degrees, professional titles, and sufficiently expensive suits, are presumed to operate at considerably higher standards. We call them experts. We invite them onto television. We pay them handsome consulting fees. We occasionally elect them to public office, although that particular experiment deserves an entirely separate discussion.

And when they get things spectacularly wrong? Well, nobody's perfect.

When AI gets something wrong, however, we have apparently discovered a fundamental threat to civilization.

I find this distinction fascinating.

Let's examine the evidence, something we humans are notoriously reluctant to do when it threatens our cherished beliefs.

A 2025 systematic review published in npj Digital Medicine examined 83 studies comparing generative AI with physicians on diagnostic tasks. The researchers found that AI achieved an overall diagnostic accuracy of 52.1%. More interestingly, the difference between AI and physicians overall was not statistically significant. Specialist physicians did outperform AI by a statistically significant margin.

Now, before anyone rushes off to replace their family doctor with a laptop, these were specific diagnostic exercises, not a demonstration that AI can safely practice medicine. The studies also had important methodological limitations.

Nevertheless, the findings raise an uncomfortable question. If a machine can perform comparably to physicians on certain diagnostic tasks, perhaps our traditional assumptions about the exclusive superiority of human expertise deserve reconsideration.

It gets better.

In a randomized clinical trial published in JAMA Network Open in 2024, researchers studied 50 physicians to determine whether access to GPT-4 improved their diagnostic reasoning. Surprisingly, physicians using the AI system did not perform significantly better than physicians relying on conventional resources.

Think about that for a moment. We developed a powerful tool capable of analyzing enormous quantities of information, placed it in the hands of trained professionals, and discovered that simply giving people access to better technology does not necessarily improve their decisions.

Perhaps the problem is not always the tool.

Perhaps the person holding it deserves a little scrutiny.

Of course, the critics are correct about AI hallucinations. These systems can fabricate information with astonishing fluency. A 2023 study published in Scientific Reports examined 636 bibliographic citations produced by GPT-3.5 and GPT-4. Researchers found that 55% of the citations generated by GPT-3.5 were fabricated, compared with 18% for GPT-4.

Those are alarming numbers, even allowing for the fact that they concern older systems rather than today's models.

But consider the delicious irony. We condemn AI for inventing references while humans have spent centuries producing dubious scholarship, promoting unsupported theories, exaggerating findings, and defending conclusions that subsequently collapsed under scrutiny.

We even have a respectable academic term for some of this behaviour: confirmation bias. It sounds much more sophisticated than admitting we are desperately searching for evidence that proves we were right all along.

Psychologist Philip Tetlock spent years studying expert political predictions. His research demonstrated that impressive credentials and confident pronouncements are remarkably poor guarantees of forecasting accuracy. Being an expert does not automatically make someone a reliable prophet, particularly when dealing with complex, uncertain events.

Naturally, this has not discouraged the endless parade of commentators confidently explaining what will happen next year, next decade, or next Tuesday.

And this brings me to what I consider the most neglected aspect of the entire debate: the reliability of the person using AI.

We are constantly warned that AI can hallucinate. But what about the human asking the questions? What about the individual who lacks sufficient knowledge to recognize an incorrect answer? Or the person who asks a carefully constructed question designed to produce precisely the conclusion they wanted?

What about the individual who receives an answer that challenges a deeply held belief and immediately dismisses it, not because the evidence is weak, but because the conclusion is inconvenient?

We have all encountered these people. Some of us encounter one every morning while shaving.

Research by psychologists Daniel Kahneman and Gary Klein has shown that genuine intuitive expertise depends on appropriate conditions, including environments where meaningful patterns can be learned and feedback is available. Confidence alone is not evidence of competence.

Yet confidence is practically a national pastime. Everyone has opinions about education, medicine, economics, climate change, and international affairs. Social media has generously provided each of us with a platform from which to broadcast our expertise, preferably without the inconvenience of acquiring any.

Artificial intelligence has now entered this magnificent intellectual circus, and somehow, it is the machine that is being singled out for being unreliable.

I am not suggesting that we should trust AI unconditionally. Quite the opposite. We should question its answers, demand credible sources, examine competing explanations, and recognize its limitations. Particularly in medicine, law, and other consequential areas, independent verification remains essential.

But why stop there?

Why shouldn't we apply precisely the same standards to our teachers, doctors, journalists, politicians, researchers, consultants, and, most importantly, ourselves?

Why do we demand that AI demonstrate near-perfect reliability while tolerating human assertions supported by little more than confidence, credentials, and an impressive ability to speak without pausing?

The objective should never be to replace human judgment with artificial intelligence. It should be to improve human judgment through the intelligent use of artificial intelligence. That requires curiosity, intellectual humility, critical thinking, and the willingness to discover that we might be wrong.

Unfortunately, these qualities are considerably harder to acquire than a ChatGPT subscription.

Perhaps the greatest danger of artificial intelligence is not that it occasionally produces nonsense. Human beings have been doing that for centuries, frequently with impressive credentials attached.

The real danger is that we now possess a technology capable of challenging our assumptions, expanding our knowledge, and exposing our ignorance, only to use it as another instrument for confirming what we already believe.

The question was never simply whether we could trust the machine. It was whether we could trust ourselves to think.

And judging by much of what passes for public discourse these days, the jury is still out.