Why Everyone Hates Autocorrect

Autocorrect has a perception problem.

If you ask people what they think about it, most of them will tell you it’s terrible. They’ll immediately have an example ready. The time it changed a word into something embarrassing. The time it completely misunderstood what they were trying to say. The time it turned a normal text message into something that made no sense. People remember those moments because they’re funny. Or frustrating. Or both. What people don’t remember are the thousands of times autocorrect worked.

Think about how often you type on your phone. You’re using your thumbs on a piece of glass with no physical keys. You’re typing while walking, watching television, standing in line at the grocery store, riding in cars, and half paying attention to three other things at the same time. By any reasonable standard, this should be a terrible way to enter text. And yet it mostly works. Not because we’re all incredible thumb typists. It works because autocorrect is constantly cleaning up our mistakes. It’s fixing missed letters, correcting spelling errors, interpreting fat-fingered taps, and making educated guesses about what we intended to say. It’s doing all of this so effectively that most of the time we don’t even notice it’s happening.

That’s the problem. When autocorrect succeeds, it becomes invisible. If I send a text message and autocorrect quietly fixes three mistakes along the way, I don’t stop and think, “Wow, great job, autocorrect.” The message gets sent and I move on with my day. The success disappears into the background. There’s no memorable moment because nothing interesting happened.

Failures work differently. If autocorrect changes one word incorrectly and suddenly I’m texting about ducks instead of something else entirely, that’s memorable. Now I have a story. Now I have evidence that autocorrect is stupid. Never mind the fact that it may have successfully corrected dozens of other words in the same conversation. The one visible failure outweighs all the invisible successes.

The more I think about it, the more I realize this isn’t really an autocorrect problem. It’s an AI problem. Every day, millions of people use technologies powered by machine learning and artificial intelligence. Spam filters catch unwanted email. Recommendation engines surface content we actually want to watch. Navigation systems route us around traffic. Voice assistants transcribe our speech. Search engines help us find information. Increasingly, AI systems help us write, code, summarize, translate, and answer questions.

Most of the time these systems work surprisingly well. And most of the time, nobody notices. What people notice are the failures.

A chatbot confidently invents a legal citation. An image generator gives somebody six fingers. A recommendation algorithm suggests something ridiculous. A language model gets an answer wrong. Those failures get shared, screenshotted, reposted, and discussed. They become evidence that the entire technology is broken.

To be clear, the failures matter. AI systems absolutely make mistakes, sometimes serious ones. But our perception of those mistakes is distorted because we only see the failures that rise above the noise. We rarely stop to count the thousands of successful interactions that happened beforehand.

Autocorrect taught us this lesson years ago. Imagine if every successful correction generated a notification.

“Autocorrect fixed a typo.”

“Autocorrect corrected a spelling error.”

“Autocorrect inferred your intended word.”

Your phone would become unusable within minutes. The only reason autocorrect feels unreliable is because its successes are silent and its failures are loud. AI has the same problem. Success is silent. Failure announces itself.

That doesn’t mean AI is perfect. It doesn’t mean the criticism is unfair. It simply means that human beings are not particularly good at evaluating technologies whose successes disappear into the background. We remember the embarrassing text message. We forget the thousands of messages that arrived exactly as intended. The ultimate irony is that the best technologies often receive the least recognition. When a technology works well enough, it stops feeling like technology at all. It becomes infrastructure. It becomes expected. It becomes invisible.

Autocorrect may be one of the best examples of that phenomenon. Millions of people rely on it every day, most of them convinced it’s terrible, even as it quietly helps them communicate. The same fate may be waiting for AI.