We told listeners a podcast host was an AI and watched what they did about it. They minded a little, mostly in the first fifteen seconds, and mostly when we stopped and asked them directly.
By Will Nash
26 August 2026
We told a group of listeners that a podcast host was an AI, and then we watched what they did about it.
The short version is that they minded a little, mostly in the first fifteen seconds, and mostly when we stopped and asked them directly. The longer version is more interesting, and includes a few results that went against us.
The test
Everyone heard short podcast clips. Each clip carried one of three labels: a normal host name, a line reading "Hosted by an AI interviewer", or a solo version of the same episode with the guest speaking alone.
The important part is that the AI-labelled clips and the normally-labelled clips were the same audio file. Only the words on the screen changed. Any difference between them is caused by the label and nothing else.
We set the pass mark before we started. The AI label had to keep at least 80% of the audience a normal host keeps. Thresholds and the analysis plan were frozen before anyone read a result.
What we found
Left alone with a clip, people treated the AI-labelled version almost exactly like the normal one. The difference was around one listener in a hundred.
Then we interrupted at ninety seconds and asked whether they wanted to keep listening or play something else. Asked directly, the gap widened to roughly seven in a hundred. People also hesitated longer before abandoning the AI-labelled clip than the normal one.
Almost all of the label's cost lands in the first fifteen seconds. After that, the two groups behave about the same, and by the end of the window they have nearly converged. The label triggers a quick judgement in some people, and it is largely spent by the time anyone has heard enough of the episode to judge it.
There is a gap between what people reported and what they did. Asked how natural the host sounded, listeners rated the AI-labelled audio noticeably lower than the identical file with a human name on it. Their listening behaviour moved far less than their ratings did.
The result that passed
We also tested format. Episodes recut so the guest speaks alone, with the interviewer's questions removed, held up against the two-person original. Among the people who turned down every conventional interview clip they were given, just under a third still stayed with the solo version.
That is the only claim in the study that cleared the bar we set, and it did so in a setup built to work against it.
What we could not prove
The AI label question did not resolve. Our best estimate cleared the 80% mark, but the sample was too small to be sure, and the range runs below the bar. The honest verdict is "probably fine, cannot promise".
The penalty also leans heavily on one episode. Remove it from the analysis and the effect disappears entirely. With four episodes we cannot tell whether that episode is the exception or the honest signal, so we would not treat the penalty as a fixed cost.
The solo result has limits too. What we tested was a careful one-off recut rather than a production process we have proven at scale, and the stronger reading of the result softens in some cuts of the data.
What we take from it
The design was deliberately tilted towards finding a penalty. People listened in a foreground browser tab with nothing else to do, and everyone saw both label types in one sitting, which makes a label far more noticeable than any real subscriber would find it. Under those conditions, a result that survives means something and a result that fails is harder to read.
For anyone weighing up how much an audience really minds, the practical finding is that the reported opinion and the recorded behaviour came apart more than we expected. If your own audience research asks people how they feel about AI, it is worth asking what it would show if it measured what they did instead.
The full study, including the methodology, the numbers and everything that went against us, is here:
Read the Aloudable Listening Study 2026 (PDF)
If you are thinking about a show of your own and want to talk it through, get in touch.
