Get Quartz in your inbox

Free daily briefing on global business news.

10 simple things AI is still bad at, from counting letters to reading clocks

Things AI is still bad at include counting letters, reading clocks and citing real sources — here are 10 everyday tasks where chatbots and image tools can still slip

An AI chatbot can draft a cover letter in seconds, translate a contract into Spanish and write working code for a website. The same kind of system has also miscounted the letter r in the word strawberry. That gap between complex tasks handled well and simple ones handled badly has a technical explanation.

It lies in how these tools are built. A large language model, the technology behind chatbots, is trained to predict the next likely piece of text based on patterns in enormous amounts of writing. It does not look facts up, count characters or run calculations unless it is connected to a separate tool that does. Image generators work in a similar way. They assemble pixels that resemble patterns in their training images, without an explicit model of what a hand, a clock or a word actually is.

That design produces a recognizable set of AI limitations. Some show up in language, such as invented citations and confident wrong answers. Others show up in pictures, such as garbled lettering on a sign. Still others appear when a task requires keeping track of something over time, like the position of pieces on a chessboard or details from a conversation last week.

What follows is a selection of 10 things AI is still bad at, drawn from well-documented behaviors. It is not a ranking, and it is not a complete list of AI mistakes. Each entry explains what goes wrong and why the underlying technology makes that error likely.

These weaknesses are not fixed in place. AI developers have added calculators, code execution, web search and memory features to their products. Those additions have reduced some of the problems described here. Newer models have also improved on tasks that earlier versions handled poorly. The causes still matter, though, because the underlying models still work by predicting patterns.

For anyone using AI at work, these limits point to a practical habit. Treat a chatbot's output as a draft, not a verified answer. Check names, numbers, dates and references against an original source. Use a dedicated tool when precision matters, whether that is a spreadsheet, a calculator or a chess engine. Knowing where large language model errors tend to cluster makes it easier to use these systems for what they do well.

Counting letters in a word, like the r's in strawberry

Large language models have miscounted the letters in simple words, including the r's in "strawberry." They do not read text one letter at a time. They read it in chunks called tokens, which hide the spelling a person sees at a glance.

The word "strawberry" contains three r's. One sits near the start, right after the "st." The other two appear side by side in "berry." Anyone who can spell can count them in seconds by pointing at each letter.

A language model starts from a different place. Before any text reaches the model, a component called a tokenizer splits it into pieces. A token might be a whole common word, a fragment of a longer word, a punctuation mark or a single character.

Each token is then converted into a number. The model processes those numbers and the patterns between them, not the shapes of individual letters. From the model's side, a word is a short sequence of IDs. The letters inside each ID are not directly visible to it.

Many tokenizers are built around frequency. Character combinations that appear often in training text become single tokens. Rarer strings get broken into smaller parts. Byte-pair encoding, a widely used method, works this way by merging frequent pairs of characters step by step. The approach is efficient for predicting text, but it does not line up with spelling.

That mismatch makes a whole family of spelling-level tasks harder than they look. Counting how many times a letter appears depends on seeing characters one by one. Writing a word backward, finding its fifth letter and solving an anagram depend on the same thing. A model has to infer that information indirectly, from patterns it absorbed during training about how words are spelled.

Some prompts reduce the problem. Asking a model to spell the word out with a space or comma between each letter tends to split it into single-character tokens before it counts. Chatbots that can run code can also hand the job to a short program, which counts characters exactly.

Multi-digit arithmetic without a calculator tool

Chatbots built on large language models generate answers by predicting the next chunk of text, not by running a calculation. As a result, long multiplication, division and other multi-digit math can go wrong when the model has no calculator or code tool to lean on.

A pocket calculator follows fixed rules. It carries every digit, lines up every column and returns the same result every time. A language model works differently. It has learned from huge amounts of text what the answer to a math problem tends to look like. It then produces the most likely-looking sequence of characters.

For small, common sums, that approach usually lands on the right answer. Basic facts such as times tables appear so often in written text that the pattern and the correct result line up. The trouble starts with large or unusual numbers. A problem involving two long figures may never have appeared anywhere in the material the model learned from. In that case, there is no reliable pattern to reproduce.

Tokenization adds another layer of difficulty. Models break text into tokens, which are chunks that can hold several characters at once. A long number may be split into uneven pieces. Those pieces do not always match the ones, tens and hundreds columns that written arithmetic depends on. Carrying a digit from one column to the next becomes harder when the model does not see the columns cleanly.

The errors can be hard to spot. A chatbot may lay out neat step-by-step working, with each line formatted the way a textbook would show it, and still slip a wrong digit into the middle. The explanation reads fluently. The mistake can pass unnoticed unless someone checks the final figure.

Some AI products route math to a separate tool. Certain chatbots can........

© Quartz