AI can write emails and summarize meetings, but here’s what it still can’t do in 2026

This is the most important one — AI tools can hallucinate, which is the industry term for when they make things up.

What’s crucial to understand is that this isn’t a bug that’s going to be fixed in a future update. Instead, it’s at the core of how a lot of LLMs (large language models), like ChatGPT and Claude, work.

Despite how it may seem, they’re not retrieving facts from a big store of information. Instead, they’re predicting the next word based on patterns learned from churning over huge amounts of training data.

Hallucination can look like your favorite AI tool confidently stating incorrect information, inventing citations or blending some real sources with made up ones.

Sign up for breaking news, reviews, opinion, top tech deals, and more.

What makes this worse is the confidence. These systems are designed to produce fluent language that sounds authoritative and we’re wired to trust authority. That makes it easy to overlook errors if we’re not careful.

That’s why it’s essential to fact check anything an AI tool tells you. This is good practice for everyday use, but it’s critical when the stakes are high, like they are with legal advice, medical information or financial decisions.

We’ve already seen multiple cases of people being caught out after submitting documents that included fabricated citations or incorrect claims generated by AI.

What to read next

Have you seen the viral videos of people asking ChatGPT or Grok how many “r”s are in strawberry? If not, I’ll spoil them for you. AI often gets it wrong. ChatGPT has been known to confidently say there are only two, then after some pushing, concede that there are in fact three.

I’ve tested this myself and had mixed results. Sometimes it answers correctly. Sometimes it doesn’t. So what’s going on?

AI tools like ChatGPT, Claude ,and Grok don’t process text the way we do. They don’t scan each letter in order. Instead, they break language into “tokens”, which are words or smaller chunks of words the LLM has learned from its training data.

So when it sees “strawberry”, it isn’t counting each letter. It’s predicting a plausible answer based on patterns it has learned before.

Once you understand how it happens, the simple mistake makes more sense. But we tend to associate fluency with intelligence, so when we know ChatGPT can write an essay in seconds but can’t count letters, it feels jarring.

This one might sound obvious, but stick with me. Acknowledging that AI hasn’t lived and never will is central to understanding what it can’t do.

It doesn’t have a body, memories, a childhood, needs or stakes. That doesn’t matter much if you’re asking it to proofread a technical blog post or generate code.

But rely on it for philosophical debate, therapy or creative work and something important shifts. It’s important to understand that it’s not drawing from a past or from dreams or experiences. It’s drawing from existing material and then recombining it.

Because it hasn’t lived, it has no skin in the game. It can describe ethical frameworks, weigh arguments for and against controversial decisions, and simulate moral reasoning. But it can’t bear consequences or be held accountable the way a person can. So if an AI system causes harm, responsibility lies with the humans who built it, deployed it or use it. The model itself has no awareness or care.

This raises deeper questions about creativity, originality and moral agency. Those debates are ongoing. But for now, it’s enough to recognize that some forms of judgement do rely on experience, vulnerability and a sense of responsibility, AI doesn’t have those.

AI tools are trained on vast amounts of data. But that data has cut-off points and those vary depending on the tool. That means a model might not know about recent events, evolving norms, or shifts in language unless you explicitly provide the context or check how up to date its knowledge is.

Sometimes this becomes a problem because older information is delivered with the same confidence as everything else. There’s no built-in signal that says, “This might be out of date”.

This really matters if you’ve started relying on AI as a news source or if you work in journalism, law, policy, or any fast-moving field. It’s increasingly normal for people to lean on these tools for research and summaries too. But you can’t guarantee information will be current.

When you first use AI, it can feel intelligent because it tends to handle language well. It can certainly give the impression of reasoning, empathy, creativity and even authority.

But it’s important to remember that underneath, it’s predicting patterns rather than understanding meaning. Recognizing these limits doesn’t diminish what the technology can do. But it’ll help you use it more clearly, more deliberately, and in ways that actually serve your goals.