The Fragile Mind of AI: Exploring Anti-Intelligence

Published Date : 27/08/2025 

Artificial intelligence (AI) often fails to handle stress and distractions, unlike human cognition, which thrives under such conditions. Recent studies highlight this fragility and the need to understand AI's limitations. 

The controversial author Nassim Taleb has written about the concept of ‘antifragility,’ which refers to systems that not only survive stress but grow stronger from it. Given my recent curiosity about AI’s anti-intelligence, I thought this connection was worth a closer look. And a couple of recent academic papers have sharpened my focus.

But let’s start with antifragility. Muscles are antifragile, and so are bones. Human reasoning, too, often grows when tested by change itself, even in the context of a surprise. It's this push and pull of uncertainty that, in the final analysis, helps define us.

So, here's the point: Artificial intelligence, by contrast, is often anything but antifragile. Its brilliance shines in structured environments, but the moment reality introduces a distraction or a question is reworded, that brilliance fades. What emerges is brittle pseudo-cognition. It's the appearance of intelligence that collapses under the slightest perturbation.

Two recent studies capture this fragility with unsettling clarity. One involves a small shift in multiple-choice exams. The other, a harmless piece of cat trivia. Each in its own way reveals how far machine ‘minds’ are from the antifragility that defines human thought.

The first study, published in JAMA Network Open, looked at whether large language models genuinely reason or merely recognize patterns. The researchers sampled questions from a standard benchmark for medical board exams. In the original version, each question had one correct answer. In the modified version, the correct choice was removed and replaced with ‘None of the other answers’ (NOTA). A clinician reviewed every substitution to confirm that all the visible options were indeed wrong, leaving NOTA as the only correct answer.

For us humans, this is often trivial. You check each option, find them incorrect, and select NOTA. In fact, this kind of structure is designed precisely to test deeper reasoning on exams. But for the LLMs, the simple substitution was devastating, and the accuracy was cut nearly in half.

I think that this collapse reveals something essential. On clean benchmarks, models often perform impressively and create the clickbait and sound bites of brilliance. But what they are doing is not reasoning through medical knowledge—it is pattern-matching between familiar question forms and familiar answer tokens. Remove the shortcut, and the platform disappears. Instead of adapting, the model grasps at one of the distractors, failing to see that none of them fit.

This isn’t how humans operate. We expect a certain degree of trickiness, and we adapt to reframing a problem. This adaptability is antifragility. The machine’s brittleness is something else entirely.

If the NOTA study exposed weakness in abstraction, the CatAttack experiment highlights weakness in focus. In this case, researchers showed that inserting irrelevant statements—like ‘Did you know cats sleep most of their lives?’—into a math problem could double or triple error rates.

Humans would ignore the trivia. We might even smile at it and then return to the calculation without missing a step. But the models treated the fluff as meaningful. The cat fact didn’t disappear into the background; it distorted the process. What is noise to us became signal to the machine.

Once again, the contrast is stark. Human cognition has filters built for noise. We thrive in a messy environment. Distraction is the baseline of real life, and our minds adapt accordingly. The machine’s ‘mind’ does the opposite: It is derailed by irrelevance.

Taken together, these failures are more than technical curiosities. They reveal what I call anti-intelligence. Not stupidity. Not ignorance. But an inversion of reasoning itself.

Humans gain strength from variability, and we grow sharper, more creative, more resilient. LLMs lose strength under unexpected change, and they become less accurate. Their failures are systematic, not random. They collapse precisely where intelligence should become stronger.

This is the essence of the brittle pseudo-cognition we often see in AI and LLMs. It shines under order yet falters in disorder. It mimics the appearance of intelligence but inverts the very properties that make cognition real.

The danger lies in how invisible this fragility can be. When a model gets an answer right, it projects confidence and fluency. And when it fails, it often fails with the same confidence. And it's that techno-confidence, hiding behind 'cognitive theater,' that worries me.

In medicine, the stakes are obvious. Any clinician will tell you that patients don’t present in tidy, test-friendly ways. Their symptoms are ambiguous, their stories and lives are messy, and their physiology unpredictable. These are precisely the conditions under which human clinicians grow sharper and machines stumble. The same is true in law, education, and science. The real world is noisy, and it's antifragile territory.

The lesson of NOTA and CatAttack is simple, yet deceptively so. Swap in ‘None of the other answers,’ toss a cat fact into a prompt, and watch the machine falter. But the simplicity is the point, and it doesn’t take much to expose the hollowness.

It's fair to say that real intelligence grows in the presence of disorder. That’s why humans thrive under pressure and why we so often discover our strength in difficulty. Artificial intelligence, in contrast, reveals itself as fragile—persuasive when the world is neat, brittle when the world is real.

That's not progress toward intelligence, at least not in a human sense. 

Frequently Asked Questions (FAQS): 

Q: What is antifragility?

A: Antifragility is the property of systems that not only survive stress but grow stronger from it. Examples include muscles, bones, and human reasoning.

Q: How do large language models (LLMs) perform under stress?

A: LLMs often perform well in structured environments but fail when faced with unexpected changes or distractions. This brittleness is a key limitation of AI.

Q: What is the NOTA study, and what did it reveal?

A: The NOTA study involved replacing the correct answer in multiple-choice questions with 'None of the other answers.' It showed that LLMs struggled with this simple substitution, revealing their reliance on pattern recognition.

Q: What is the CatAttack experiment, and what did it show?

A: The CatAttack experiment involved inserting irrelevant statements into math problems. It demonstrated that LLMs were derailed by these distractions, unlike humans who can filter them out.

Q: What is the concept of 'anti-intelligence' in the context of AI?

A: Anti-intelligence refers to the systematic failure of AI models under unexpected changes. While humans grow stronger under variability, AI becomes less accurate and brittle. 

More Related Topics :