Can Listeners Tell AI Narration? What the Published Research Measured

The published research we could read is thin: a 2026 journal paper reports very small differences between AI and human narration in listener ratings, a 2023 paper exists whose findings we could not confirm from a readable abstract, and a vendor-sponsored survey reports that 61% of listeners who heard the AI version thought the voice was human, against 65% of those who heard the human version. None of this is a result from NarratedIndie, because we have not run a listening panel.

The three sources at a glance

Each source answers a different question, so the first step is to separate what was measured from what was read.

What the 2026 paper in Computers in Human Behavior Reports measured

The newest study we found, "Who cares about artificial intelligence? Human and artificial voices in audiobooks" by Ahrens-Schwabe, Kosch, Stocker and Boomgaarden, was published in 2026 in Computers in Human Behavior Reports (volume 22, article 101068). The abstract page gives the month as May. According to the University of Vienna abstract page, the study measured evaluations by self-identified audiobook listeners of literary excerpts read by AI voices or professional actors. It reports that what listeners believed about the narrator type did not change their ratings, and that the differences between the two were very small.

The caveats: we read only the abstract page, the sample size was not on it, and we do not know which AI voices were used. A voice that was current when the study ran may not be the one you would pick today. "Very small" also needs the paper's actual figures before anyone leans on it, and we do not have them.

What we could confirm about the 2023 paper in New Media and Society

"Synthetic versus human voices in audiobooks: The human emotional intimacy effect" by Rodero and Lucas appeared in New Media & Society 25(7), pages 1746 to 1764 (DOI 10.1177/14614448211024142). The repository record confirms the paper exists. We could not read an abstract, so we state no findings from this paper. If you rely on it, read the paper's abstract yourself.

The Edison Research survey for Spoken

A survey by Edison Research, commissioned by Spoken, an AI audiobook company, found AI multi-cast narration rated higher than human narration. One trade report (RAIN News, July 14, 2026) summarizes it as 61% of respondents being unable to tell AI from human, but the press release gives 61% of the AI group (and 65% of the human group) as thinking the voice was human. We read that trade article, Digital Trends' report and the press release as carried by Podnews, but not Edison's own full report.

As reported, it surveyed more than 1,000 US adult audiobook listeners using an excerpt from a science-fiction thriller; the coverage says the survey focused on character-driven fiction. Digital Trends gives 1,005 respondents, surveyed in May 2026. It was sponsored by a company that sells AI audiobooks, and it is not peer reviewed. Its method comes to us only secondhand. Weigh it accordingly.

The release's own numbers add detail the headline loses:

Four things change how those numbers read:

• The 61% is not a side-by-side test. Listeners were split into two blinded groups and each heard one version. So 61% of those who heard AI, and 65% of those who heard a human, thought the voice was human. The two groups can be compared, but nobody was asked to pick the AI out of a pair.
• It compares unlike things. The release pits an unedited, one-click AI production with a different voice for each character against one human narrator. That says little about single-voice AI narration.
• The human narrator won on exposition. The release says that for passages without multiple characters, the human narrator was rated higher.
• Purchase intent is a stated answer. The release calls 46% and 49% statistically comparable. It gives no margin of error, and an answer in a survey is not a sale.

Why the studies do not settle your question

These limits apply to the sources we could read, and they matter to a buying decision:

• Different questions. One measures ratings, one whether listeners can tell. "Cannot tell" and "prefers" are not the same.
• Short excerpts. Studies play excerpts, and an author sells a book of many hours. Consistency over a full book is not what the two we could read address.
• Voices in a study are a snapshot. A study fixes the voices available when it ran; we did not check how these voices compare with current ones.
• Genre and listener. Fiction excerpts were used where we know, and nonfiction, children's or technical books are not covered by what we read.
• Stated reactions, not sales. We found no study of what actually happens to sales when a title carries an AI label.

Those labels are often added anyway: Spotify adds a sentence to the description of digital voice titles, and see the labelling guide for other stores. Some retailers state positions of their own too: Libro.fm said in a February 18, 2026 post that it will never knowingly prioritize, publicize or market AI-narrated audiobooks.

What this means if you are choosing a narrator

The evidence supports a narrow conclusion: in at least some published work, listeners did not rate AI narration much lower, and the strongest "cannot tell" claim comes from a company that sells AI audiobooks. A second academic paper's findings are unconfirmed. It does not tell you whether your book, your genre and your voice choice will hold up.

Judge your own book. Run the free same-chapter comparison on a passage with your dialogue and hard words, and share it with a few readers without telling them which is which. Two habits borrowed from the Edison design, plus two of ours:

1. Give each reader one version, then ask about it, as Edison's two-group design did, rather than asking them to spot the AI.
2. Use one passage of dialogue and one of plain exposition, because the Spoken release reports different results for each.
3. Our suggestion: ask whether they would keep listening before you reveal anything, then ask again after you say what it was.
4. Our suggestion: listen to a long stretch yourself. Excerpts cannot show what happens over hours.

A handful of readers can expose a voice that grates or a name that is wrong. It cannot prove that listeners in general will accept the book, so treat it as a check for problems, not a result.

For which stores accept AI narration, see the retailer overview, for tool picks best AI audiobook generators, and for the wider workflow the AI audiobooks hub.

NarratedIndie has listed a panel as a possible future project, with consent, a scoring sheet and raw results, but none has been run and no date is set.