The 10 Things AI Cannot Do Yet

The Hype Gap

There is a wide gap between what AI can do impressively in a demo and what it can do reliably in your product. Understanding the limits is not pessimism -- it is practical risk management. If you know where AI breaks down, you build better systems around it.

Here are the five most important limits of current AI, with concrete examples of each.

It Cannot Reliably Reason About Novel Situations

AI models learn patterns from training data. When a situation closely resembles something in that data, the model often performs brilliantly. When the situation is genuinely new -- a combination of factors it has never seen together -- performance degrades fast.

Example: a legal AI assistant trained on US contract law will sound confident about a cross-border deal involving regulatory frameworks from two smaller jurisdictions it barely encountered during training. It may produce plausible-sounding but incorrect advice without any sign of uncertainty.

This is not stupidity. It is a structural limit of pattern-matching at scale. Novel = outside the distribution = unreliable.

It Cannot Verify Its Own Claims

AI does not have a fact-checking layer separate from its text generation. When it produces an answer, that answer feels confident because the model is optimized to produce fluent, coherent text -- not to separate known facts from interpolated guesses.

Example: ask an AI model who won a particular regional award in a particular year. It may name someone confidently. That person may not exist, or may have won a different award, or the award may not exist. The model does not know it is wrong. It cannot look up sources the way a search engine does.

This property -- confidently wrong -- is called hallucination, and it is a direct consequence of how language models work, not a bug that will be patched away completely.

It Cannot Learn From a Single Conversation

When you have a productive exchange with an AI assistant and close the window, that learning does not persist. The model weights do not update. Tomorrow, you start from scratch.

Example: you spend an hour teaching an AI your company's naming conventions, acronyms, and preferred tone. That context exists only for the duration of that session. The next day, the model has no memory of the conversation unless you give it again (or unless the system stores and re-injects it).

This is a systems problem, not only a model problem. Retrieval systems, memory layers, and fine-tuning can partially address it -- but these require deliberate engineering and do not happen automatically.

It Gets Confused by Negation and Edge Cases

AI models are notoriously bad at handling negation precisely, especially when it appears in complex or nested sentences.

Example: "List all the countries we operate in that are not subject to export controls" sounds simple. But the model may return countries that ARE subject to export controls, or miss the negation entirely, especially when the sentence has multiple clauses.

Edge cases cause similar problems. A model trained to classify customer sentiment handles "great product, terrible support" inconsistently. It handles sarcasm poorly. It handles culturally specific expressions even worse.

If your application depends on precise handling of conditions, negation, or unusual inputs, build explicit validation around the model's output rather than trusting it directly.

It Does Not "Understand" in the Human Sense

This is the most philosophical limit, but it has practical consequences. AI models process statistical relationships between tokens. They do not have goals, beliefs, values, or an internal model of the world the way a person does.

Example: ask an AI to explain why water is wet. It will give a fluent, accurate-sounding explanation. But the model has never touched water, has no sensory experience, and does not "understand" wetness. It has learned that certain words and concepts appear near each other in text about water.

This matters because human reasoning includes common sense, physical intuition, social awareness, and causal thinking that we often take for granted. AI can simulate many of these, but the simulation breaks at edges.

It Cannot Reliably Tell Instructions From Content

When you ask an AI to read something (a document, an email, a web page) and then do something with it, the model sees your request and the content as one continuous stream of text. If the content contains instructions, the model may follow them as if they came from you. This is called prompt injection, and as of now nobody has a reliable fix for it, because it is a consequence of how these models process text rather than a bug that can be patched.

It matters most for AI agents, the newer kind of AI that can take actions: browse sites, send messages, edit files, make purchases. An agent that reads a malicious page can be steered into doing something you never asked for. The practical consequence is that the safety of an AI agent depends far more on what it is allowed to do than on how smart it is. When you evaluate a tool that acts on your behalf, ask what it can do without asking you first. That list should be short.

What This Means Practically

These limits do not make AI useless -- they make AI a tool that requires thoughtful deployment. Use it where the cost of an error is low or easily caught. Add verification layers where errors are costly. Keep humans in the loop for novel, high-stakes, or legally sensitive decisions.

What to Explore Next

• /posts/why-ai-makes-mistakes -- the five mechanisms behind AI errors, with examples you can act on
• /posts/responsible-ai-for-practitioners -- how to think about risk and bias when building or buying AI products
• /posts/prompt-engineering-intro -- how better inputs can reduce (but not eliminate) these failure modes

What to Practice Next

• Find a news headline that overstates an AI capability (e.g., "AI can now diagnose cancer better than doctors") and write a one-paragraph rebuttal that cites the actual benchmark conditions, dataset scope, and what the headline left out.
• Ask a current LLM a question you know it should struggle with (multi-step spatial reasoning, a precise arithmetic problem, or a question requiring knowledge after its training cutoff) and document exactly where and how it fails.
• Sketch the boundaries of a real AI system you interact with: list three things it does reliably, three things it does inconsistently, and one failure mode that would be catastrophic in a high-stakes deployment.