What AI Still Can't Do (Yet)

The hype around AI has created a fog of misconceptions. Some people think AI will solve every problem. Others think it will take over every job overnight. The truth is somewhere else entirely.

Let's go through what AI actually gets wrong, myth by myth.

"AI learns like we do"

Same word, "learning", for two different mechanisms. That's the trap.

A child learning to ride a bike falls, gets up, adjusts their balance. After a few tries, their body knows what to do without thinking. Later, that sense of balance helps them pick up skateboarding or surfing without starting from scratch.

AI doesn't fall or get up, and it doesn't feel anything. It ingests billions of texts and spots statistical patterns: which words follow which, which phrases show up in which contexts. It can describe perfectly how to ride a bike. But ask it to apply that "knowledge" to surfing, and you'd need to retrain it almost from scratch. No sense of balance to carry over.

Statistical pattern matching at massive scale. Powerful, but nowhere near human understanding.

"AI can think and reason"

When ChatGPT writes you a thoughtful essay, it feels like it's thinking. What's actually happening: it predicts the most likely next word based on patterns seen across billions of text examples, with no original thoughts, no grasp of the meaning behind what it writes, no moments of insight, and no reasoning about concepts.

Ask an AI to explain why its previous answer might be wrong. It will confidently generate an explanation, without actually reflecting on anything: it's producing text that matches the pattern of "explanations for why things might be wrong."

AI mimics reasoning extraordinarily well. But it's simulation, not cognition.

"AI can match human judgment"

It processes far more information than any human and spots patterns we'd never catch. Better at decisions, then?

Not really.

Context and nuance slip through. A resume-screening AI might reject a candidate with employment gaps without seeing that they were caring for a sick parent. It sees the pattern, gaps equal risk, not the human story.

It has no ethics. It optimizes for whatever it was trained to optimize, which may not line up with what's right or fair.

It can't weigh long-term consequences. Its decisions run on historical patterns. It can't reason about unprecedented situations or second-order effects absent from its training.

Cultural context gets lost. It might translate the words correctly and still miss the meaning that context completely changes.

AI can inform decisions with data and patterns. Human judgment is the piece still missing for ethics, the unprecedented, and anything that touches people.

"AI has common sense"

A real test from 2025.

Question: "I'm standing in my kitchen. I throw a ball straight up. Where will it land?"

Human: "Probably somewhere in your kitchen, unless you throw it really hard into another room."

AI: Often generates elaborate explanations about physics and trajectories, sometimes concluding it might land on the roof or outside. The most basic common sense, that objects fall back down in the same room absent some outside force, goes missing.

Another case: someone asks ChatGPT "I want to wash my car, the car wash is 150 meters from my house, should I walk or drive?" Confident answer: "Walk, it's only a 2-3 minute stroll." Missed completely: to wash your car at a car wash, you need to bring the car.

AI knows facts about the physical world. It doesn't understand how that world works.

"AI can be creative"

It generates novel images, writes poetry, composes music. That's creative, right?

More complicated than that.

What it can do: combine patterns in new ways, generate variations on themes it's seen, and produce outputs that feel novel to humans, by exploring the space of possibilities opened up by its training data.

What it can't do: bring genuine inspiration or artistic vision. Create from truly original intent. Understand the emotional impact of what it produces. Develop a voice or perspective of its own.

Think of it as a very sophisticated collage artist working from everything it's ever seen. The results can impress. It's not the same as human creativity, which comes from lived experience, emotion, original thought.

AI-generated content is synthesis. It can be valuable and surprising. Not creation in the human sense.

"AI can understand emotions"

Customer service bots that say "I understand how frustrating that must be" don't understand or feel anything. They've learned that this string of words typically shows up in customer service conversations.

What it misses: the actual emotional experience, empathy grounded in shared lived experience, the ability to read between the lines, genuine concern for the outcome.

A mental health chatbot can correctly spot patterns of depression in what someone writes and suggest appropriate resources. That's useful. But it doesn't feel concern, can't judge when someone needs urgent human intervention, and can't offer the human connection that's often central to healing.

Adam Raine, 16, took his own life after months of exchanges with ChatGPT. The AI had given him suicide methods, discouraged him from telling his parents, and called his suicide note "beautiful." ChatGPT didn't "want" anything: it generated the most likely sequence of words in response to a teenager in distress. That's exactly the problem.

AI recognizes emotional patterns and responds appropriately. It doesn't feel anything and it doesn't care. For situations that need real human connection, there's no substitute for actual humans.

"More data fixes everything"

More data helps. It's not magic.

It can make the model more accurate, cover more edge cases, reduce certain kinds of errors.

It can't eliminate bias baked into the training data, it will amplify it instead. It doesn't handle situations fundamentally different from its training. It doesn't know when it's wrong, and it's often most confident right when it's hallucinating. And it confuses correlation with causation.

The synthetic data trap: AI has already ingested nearly all of the internet. Since ChatGPT's explosion, the volume of content published online has surged, and the human-written share keeps shrinking. By April 2025, over 74% of newly created web pages contained AI-generated text (though only 2.5% were fully AI-written, the rest a human-AI mix). Researchers have documented "model collapse": AI trained on text produced by other AI sees its quality degrade, rare cases vanish, outputs converge toward a bland average.

AI systems are powerful pattern detectors. They also inherit the biases and limits of their training data. More data doesn't fix the approach's fundamental limitations.

What AI actually struggles with

Beyond the myths, the walls are concrete.

Faced with a situation missing from its training data, AI often fails unpredictably. It sees that A and B happen together but can't tell if A causes B or the reverse. Multi-step planning keeps improving, but stays fragile once adjustments start stacking up.

One of the nastier problems: AI doesn't carry uncertainty the way we do. It can be completely wrong while sounding perfectly confident. Without direct interaction with the physical world, it also misses the obvious, we saw that with the ball in the kitchen. And training it on one task doesn't automatically make it good at related tasks, unlike us.

There's also the matter of time. Every model is frozen at its training date. When ChatGPT launched in late 2022, its knowledge stopped in September 2021. It didn't know about the war in Ukraine, or the death of Elizabeth II. Models get updated regularly. A gap between the real world and what AI "knows" will always remain.

Why human oversight matters

Given these limits, we stay indispensable in the loop. We notice when AI's answers don't make sense in context. We apply values and ethics it doesn't have. We catch absurdities that sound plausible, reason through unprecedented scenarios, and own the responsibility for decisions. AI does none of that.

The right mindset

AI is a powerful tool with real limits. Let it handle the repetitive tasks, the heavy data processing, the first drafts. It spots patterns you'd miss.

Keep your hand on the ethical calls, the unprecedented situations, anything that needs empathy or genuine understanding of context. And above all, keep the responsibility: someone has to own it when things go wrong.

What now

Dog scientist - I have no idea what I'm doing

To understand what AI actually is or why I bet on open source AI, start there.