AI Can’t Read

It’s somewhat confusing to see news stories about rogue AIs hacking government systems while chatbots still routinely fail questions a toddler could answer. Tech bros talk about AI like they’re on the verge of building Skynet, while at the same time Taco Bell’s AI drive-through is crashing its system by accepting an order for 18,000 cups of water.

Part of the issue is that while it’s easy to anthropomorphize AI chatbots, they don’t see the world as a human does. In fact, all they see is math. A very specific kind of math. Let’s break it down.

Tokens, not words

One of the most widely shared ways to trip up an AI for a long time was to ask it how many R’s are in the word “strawberry.” Any toddler who knows the letter R can simply look at the word and count three of them, but AI chatbots would regularly answer that there are only two.

Upon failing the challenge, users would ask the AI to read out every letter individually, and it would respond by saying “S-T-R-A-W-B-E-R-R-Y,” clearly saying “R” out loud three times. The users would once again ask the bot how many R’s are in “strawberry.” Again, the AI would say that there were just two.

This bizarre error occurs because large language models (LLMs) don’t actually have a way to work with letters or words directly. Instead, they encode all the information they’re given as tokens, which are units of meaning that are assigned to different numbers. So, for example, the word “cellphones” might be broken up into three tokens, “cell,” “phone,” and “s,” each of which is assigned a number. The AI knows that “cell” and “phone” have meanings on their own, but have a different meaning when they’re stuck together like this. It also knows that when “s” is added on, the meaning will change again, and the other words in the sentence will also change to reflect that.

Let’s say “strawberry” is made of two tokens, “straw” and “berry,” and that they’re assigned the numbers 135 and 933. When you ask the AI to count the R’s in “strawberry,” it doesn’t see the word “strawberry.” It only sees a series of two tokens,  “135, 933,” and has no way to count how many R’s are in the word those tokens represent. It does know that when tokens like these show up after tokens like the ones in your question, the proper response is usually the token corresponding to the word “two,” so that’s what it’s going to type.

A mathematical engine that can’t do math

LLMs don’t actually see or understand the words that you give them. They just turn your input into long sequences of numbers and perform complex mathematical transformations on them. Ironically, this also makes them terrible at simple math problems.

If you ask a base LLM to multiply 722 x 185, it will give you an answer that looks plausible, but is completely wrong. Having seen plenty of examples in its training data, it might correctly guess that multiplying any numbers with a 2 and a 5 in the ones place will end in a zero. It might also correctly guess that numbers of this size usually have products that are six digits long. It might make hundreds or thousands of statistical inferences about what the answer should look like, but it won’t actually perform the simple multiplication operation. Because that’s not what LLMs do. Tens of billions of parameters in its statistical model, yet an LLM can’t do arithmetic as well as a pocket calculator.

Skirting shortcomings through tooling

To deal with such cases, most AI apps have ways around them through specific tooling. Tooling refers to building additional software applications around a language model to shape the way users interact with it.

For example, voice recognition tools will convert spoken words into text that can then be fed into the model. The chat interface of an AI chatbot is also a program that makes the output of an LLM look less like fancy autocomplete and more like a back-and-forth conversation. There are retrieval-augmented generation (RAG) tools that copy search results into the LLM, and there are coding tools that allow models to create and evaluate code. There are even basic calculators to help out with the arithmetic.

So if you ask a chatbot to perform a simple math question, it knows to send that part of your request to an actual calculator program instead of letting the LLM try to answer the question itself. Likewise, the strawberry error is so widely publicized that many models are specifically tooled to answer that specific question correctly.

Of course, the strawberry error is just one example of a bigger problem. In a 2025 study by OpenAI researchers Adam Tauman Kalai, Edwin Zhang, and Ofir Nachum as well as Santosh S. Vempala from Georgia Tech, they tested several AI models on a similar question: How many D’s are in the word “DEEPSEEK”? All the models they tested got the question wrong, with some guessing as high as six or seven.

The problem of LLMs being unable to see letters and numbers not only isn’t fixed, it literally can’t be fixed because that’s what an LLM is.

How to get the most out of an LLM

Large language models may not be able to read or reason, but a human can. The key to getting the most out of an AI tool isn’t to just use it for everything. The goal should be to use AI in ways that supplement users’ own cognitive abilities, rather than hinder them.

AI can be a powerful tool for summarizing information, for interfacing with complex software, and for automating repetitive tasks with a low cost of failure. If a chatbot gives you a list of legal sources to cite and a few of them are bogus, it’s not a big deal. You simply verify the accurate sources and skip the others. If you let a chatbot write a legal brief for you and then submit it to a judge, you’re going to be in serious trouble.

Treat LLMs like the specialized tools that they are. Understanding how they interact with data will help you better determine which tasks are best suited for an LLM and which ones are more easily accomplished by a human.

Additional resources

• How To Search by Image
• What Is a VPN?
• Understanding Age Verification