Ask a large language model what comes after the number 17 in the sequence 2, 4, 8, 16, and it will almost certainly get it right. Ask it to tell you how many times the letter “r” appears in the word “strawberry” and there is a decent chance it gets it wrong. These two tasks seem like they should require similar capabilities. They do not. Understanding why reveals something important about both the power and the limits of AI.
The sequence problem works for AI because it resembles patterns the model has seen millions of times in its training data. Mathematical sequences, programming tutorials, math textbooks. The model has seen so many examples of “double the previous number” that pattern matching gets you to the right answer easily.
The letter-counting problem is different. To get it right, you need to actually look at the word character by character and keep a running count. That is a sequential procedural task, not a pattern-matching task. Large language models do not read words as sequences of individual characters. They process text in chunks called tokens, and the internal structure of a word is not something the model has direct access to in the way a human scanning a page does. So it guesses. And it often guesses wrong.
This pattern shows up across many types of tasks. AI models tend to fail at anything that requires strict step-by-step logical operations, especially when the problem does not closely resemble examples from training. Puzzles that require tracking multiple changing states, multi-step word problems, spatial reasoning tasks, basic logic gates. The model can talk about these things fluently. It often cannot do them reliably. The fluency is part of what makes the failures surprising. The model sounds confident right up until the moment you check the answer.
Innovascope covers a lot of ground here. The term researchers use is “brittle”: the model performs well in familiar territory and falls apart in ways that feel random and unpredictable outside it, which is the opposite of how a genuine reasoning system should behave.
This matters for anyone using AI tools to do serious work. Summary? Great. Drafting? Useful. Checking your logic on a complex argument? Be careful. Doing arithmetic? Use a calculator. The model giving you a confident answer does not mean the answer is right. It means the model found a pattern it liked.
The AI companies know this. There is active research into giving models better tools for genuine reasoning, including access to code interpreters and calculators that do the procedural work the underlying language model cannot. Some of this is working. Models with tool access perform much better on the tasks where raw pattern-matching fails.
But the core language model itself still cannot reason in the way a person does, and a lot of the impressive demos you see are carefully selected to show the instances where pattern-matching happens to produce the right answer. The failures are quieter and less shareable. The useful mental model is this: AI is very good at recognizing what the right answer usually looks like. It is much worse at actually figuring out what the right answer is. That distinction matters every time you trust it with something important.
