AI Writes for Us – and Almost No One Notices

You receive a message from a colleague, a friend, or a stranger online. The text is polite, to the point, and doesn't seem odd. You reply – without even thinking about who or what wrote it. According to a new study from the University of Michigan, this is exactly what happens in most cases: people hardly ever notice when they're communicating not with a human, but with AI-generated text.

People Can't Tell AI-Generated Text From Human Writing

We Think We Can Tell – But We Can't

Intuitively, it seems easy to distinguish a «live» text from a machine-written one. Many people are confident they would sense something unnatural – too smooth, too perfect, too impersonal. But experiments show the opposite.

Researchers tested whether people could determine the origin of a text in real-life, everyday communication scenarios: email exchanges, short messages, and typical casual remarks. The result was quite telling – participants very rarely suspected they were reading a text written by artificial intelligence. Simply put, AI blends into a normal conversation almost seamlessly.

And it's not because people are inattentive or uncritical. The problem most likely lies elsewhere: we simply don't have a firm idea of what an AI text «looks like.» We're looking for a robotic quality that is no longer there.

Modern AI Models Write Like Real Humans

Modern Models Write Like Humans – and That's the Core of the Problem

Just a few years ago, generated text really gave itself away: odd phrasing, mechanical repetitions, a feeling that it was «translated from an alien language.» Today, things are different. Modern language models are trained on vast datasets of human speech – and they reproduce its style, intonation, and even minor imperfections quite convincingly.

This is especially noticeable in casual correspondence. When a message is short and neutral – «okay, sounds good», «I'll check tomorrow», «thanks, got it» – the reader simply has no clues. The text contains nothing suspicious, and the brain automatically attributes it to a human.

This isn't a bug or a coincidence – it's a direct consequence of how modern text generation systems work. They are optimized for naturalness. And they have succeeded.

Why AI's Unnoticed Presence in Communication Matters

Why This Matters Beyond the Lab

One might argue: so what? If the text is relevant and doesn't offend anyone, what difference does it make who wrote it? That's a reasonable question, and there's no simple answer to it.

On the one hand, using AI in correspondence is already common practice. People ask models to help draft an email, soften a confrontational tone, or simply save time. In this sense, AI is becoming a kind of smart editor – a tool, not a substitute for a person.

On the other hand, if the person you're talking to can't tell a live human from an automated response, it changes the very nature of communication. Trust in written correspondence has traditionally been built on the assumption that behind the words stands a specific person with intentions, a mood, and accountability. If this assumption can no longer be considered reliable, it's no longer a technical issue, but a profoundly human one.

The researchers aren't claiming that AI in correspondence is unequivocally bad. But they are highlighting that the widespread invisibility of such texts creates an environment where the line between «written by a human» and «written by an algorithm» is gradually blurring – and most people simply don't realize it.

How to Address AI-Generated Text Detection

What to Do About It – Is Still Unclear

One approach being actively discussed in the research community is the so-called «labeling» of AI-generated content: technical markers or watermark signals embedded in the text, invisible to the reader but detectable by special tools. The idea is viable but comes with significant caveats – such markers can be circumvented, and a universal standard doesn't exist yet.

Another approach is to increase user «AI literacy»: teaching people to question the origin of a text, pay attention to context, and not take communication for granted. It sounds reasonable, but in practice, it works less effectively than one might hope – especially in situations where a person is tired, busy, or simply not expecting a trick.

For now, the study offers more of a diagnosis than a prescription: we live in a communication environment where AI has already become an almost invisible participant in the conversation. And most of us don't suspect a thing – not because we're naive, but because we don't yet have the tools to notice it.

A Final Bit of Irony

The study's authors accompanied their publication with an ironic remark: the press release was written without the use of artificial intelligence. And even if it hadn't been, you likely wouldn't have suspected a thing.

This is, perhaps, the best illustration of the study itself.