AI tools like ChatGPT, Claude, and Gemini have gotten scarily good at writing, coding, and analysis. But the hype machine has convinced a lot of people that these models can do anything—which leads to frustration when they fall flat on tasks that seem simple.
Here are five things AI models still genuinely can’t do well in 2026, despite what you might have heard.
1. Accurate Math Beyond Basic Arithmetic
Large language models are text prediction engines, not calculators. They can handle simple arithmetic (“What’s 237 + 489?”), but ask them to compute compound interest over 15 years, solve a multi-step algebra problem, or work through calculus, and they’ll often get it wrong.
ChatGPT and Claude have integrated plugins and code execution environments that can run Python to solve math problems—but the model itself doesn’t “do” math. It writes code that does. If the code is wrong, the answer is wrong.
What to use instead: Wolfram Alpha, a scientific calculator, or ask the AI to write Python code you can verify yourself.
2. Real-Time Information or Live Data
AI models are trained on static datasets with knowledge cutoffs. Even models with web search plugins (like ChatGPT with Bing or Gemini with Google Search) are retrieving and summarizing web results—they’re not “aware” of breaking news in real time.
If you ask “What’s the weather right now?” or “Did the stock market close up or down today?”, the model is guessing, searching, or pulling outdated training data. It’s not connected to live APIs unless explicitly integrated by the developer.
What to use instead: Google, dedicated news apps, stock tickers, or weather services.
3. Reliable Fact-Checking or Source Verification
AI models are confident liars. They’ll cite nonexistent studies, invent book titles, and fabricate URLs with perfect grammar and zero hesitation. This isn’t malice—it’s how the models work. They generate plausible text, not verified truth.
Even when models do cite real sources (via web search), they often misrepresent what those sources say. A model might pull a quote from a paper and completely misstate its conclusion.
What to use instead: Your own eyes. If a fact matters, verify it manually. Use Google Scholar, official websites, or primary sources.
4. Understanding Images in Fine Detail
Vision-enabled models like GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro can describe images, read text from screenshots, and identify objects. But they struggle with precision.
Upload a diagram with small labels, a chart with overlapping data points, or a photo of handwritten notes, and the model will miss details, misread text, or hallucinate elements that aren’t there. They’re also terrible at counting objects in complex images.
What to use instead: OCR tools like Adobe Acrobat or Tesseract for text extraction, or manual review for anything that requires accuracy.
5. Long-Term Memory Across Sessions
ChatGPT’s Memory feature and Claude’s Projects can store context across conversations—but these are shallow, fragile systems. They don’t “remember” the way humans do. They store snippets of text and retrieve them inconsistently.
If you’re building a complex project over weeks, don’t rely on the model to remember every detail. It will forget, contradict itself, or retrieve the wrong context. The memory features are useful for preferences and light continuity, not deep state management.
What to use instead: Keep your own notes in a document, use a task manager, or paste critical context into each new session manually.
Why This Matters
Knowing what AI models can’t do is just as important as knowing what they can. It saves you time, prevents errors, and keeps you from blaming yourself when the tool fails at something it was never designed to handle.
Use AI for what it’s great at—drafting, brainstorming, summarizing, coding assistance—and keep human judgment, external tools, and verification in the loop for everything else.
Want one useful AI idea delivered to your inbox every day? Subscribe to the One Two Three AI newsletter and stay sharp on what works (and what doesn’t) in 2026.