Can readers tell when a book was translated by a machine?

Every author considering a translation asks the same thing: if I use a computer to translate my book, will readers feel it? Someone tested this properly, and the answer is that they mostly cannot tell — yet they still prefer the human version, without being able to say why.

What the test looked like

Researchers took 15 novels that professional translators had translated into English, and made a second version of each, translated by computer. Then they gave real readers — people who read books for pleasure, not experts — passages from each book, and asked two things: guess which passage was machine-translated, and say which one you prefer.

First result: they could not tell

Their guesses were about as good as a coin flip. A machine-translated passage read to them like any other book. That tells us something important about where things stand: the era of computer translation sounding clunky and strange is over. If you are still picturing translation that reads like the instruction manual for a washing machine, reality has moved on.

Second result: they preferred the human one anyway

Here it gets interesting. Those same readers who could not identify the machine version chose the human version more often when asked which one they liked better. And the more slowly and closely they read, the clearer that preference became.

In plain terms: the difference is real, but it is not loud. It is not a wrong sentence that leaps off the page. It is a cumulative sense that something is a little less alive.

What this means for your book

Two practical conclusions.

First, do not spend your review asking "does this feel like a machine wrote it". Experienced readers could not answer that, and neither will you — certainly not in a language you do not read. It is a check that returns no information.

Second, do check other things, and those turn out to be easy to find. Whether a character has the same name in chapter one and chapter ten. Whether they speak in the same style throughout. Whether a recurring term was translated the same way every time. These are exactly what readers experience as "something is off", and they are also the only things you can check without reading the language.

The limitation worth knowing about

Readers were given the opening chapters of these books, not complete books. The researchers say so themselves. That matters to you, because an opening chapter does not yet contain the problems that build up over a full-length book — a name that shifts halfway through, a style that drifts. The longer the book, the more room such things have to accumulate.

A fuller comparison of the options is in a human translator versus AI, and the practical checks are in how to check a translation in a language you cannot read.

And what about books that are not novels

The test was run on literary fiction, where style is part of the product. For non-fiction, a how-to book or a professional title, the equation shifts: the reader mainly wants to understand, and slightly less musical prose bothers them far less. Precision in terminology, on the other hand, matters more.

So what counts as good enough depends on the kind of book you wrote. It is worth asking what your reader comes for — an experience or information — and reviewing accordingly.

For anyone who wants the test itself: the full study.