AI transcription ethics: Where the limits lie in journalism

The sentence nobody said

Hallucination, where text appears that was never in the audio, is a common problem in AI transcription. For a journalist it is more dangerous than a misheard word, because the result is a clean sentence no one ever said.

The best-known study is Careless Whisper. In 2023 its authors ran 13,140 English audio segments through the Whisper API, found phrases absent from the audio in roughly 1 percent of transcriptions, and judged 38 percent of those harmful. Longer pauses raised the risk. The same study shows that after a December 2023 update most of the problem segments no longer produced invented text.

That figure needs careful reading. It comes from English recordings and one service at one date, and it is not a rate for other languages, today’s models or LiveType. What is certain is that the problem is real and cannot always be spotted by reading the text.

Then there are the ordinary errors: similar-sounding words, misspelled names, mixed-up numbers, and a lost negative. “We did” and “we didn’t” are one small sound apart, so check those in the audio and not on the page.

AI transcription accuracy gives examples of what gets written in silence most often.

The line between transcript and summary

Everything here is about writing speech down. Asking AI to summarise a conversation or to pull quotes from it is a different job and a much riskier one: A language model can produce a sentence that sounds said and was not.

A real case came in 2026, when Ars Technica retracted a story containing quotes generated by an AI tool and attributed to someone who never said them. The person quoted noticed. Press Gazette keeps a running list of similar cases.

The rule that follows is simple: Every word inside quotation marks comes from audio the journalist heard. A summary, or a chatbot’s “they said this”, has no place there.

Where your source’s voice and words go

The ethics of transcription go beyond accuracy, because protecting a source means knowing where a recording goes. The Freedom of the Press Foundation notes that online services must receive audio in a form they can read, and suggests tools that run on your device for sensitive work. LiveType is also software that processes speakers’ voices on your own device and provides privacy.

Before choosing a tool, ask: Who receives the audio? Is it kept? Is it used for training? Where does the text live?

For LiveType, the answers are:

• With This device selected, audio does not leave your device. You can check this yourself.
• With the cloud option, it does: To the provider with your own key, or through our server without being stored when you use credit. Do not use it for a sensitive conversation.
• The text lives on the server. Keeping the audio on your device does not protect the text. Teams and link holders you share with can see it, and the system administrator can open a transcript.

So for a confidential source, “the audio stays on my device” is not enough. Do not put identifying names and details in speaker names or the title, and export and delete the text when you are done. For a conversation that could put someone’s life at risk, using no internet-connected tool at all is a legitimate decision. The recording on your phone and the file on your computer need the same care.

For the personal data side in Turkey, see KVKK and voice recordings.

Should you tell the reader?

There is no easy answer, and we will not hide the reasoning. Turkey’s Press Council principles state the public’s right to know when and how AI was used, and the Journalists’ Association asks that the program be named. In US surveys reported by The Journalist’s Resource, most respondents want disclosure. None of these names transcription separately.

This article’s view: Transcription should not affect the substance of a story. It only makes writing easier. It does not write, pick quotes or decide, and once checked against the audio, a quoted word is simply what the speaker said. So there is no need to tell readers separately that AI helped transcribe the conversation.

That view has three limits:

• If your outlet’s policy asks for disclosure, follow it. An outlet that reads these codes as covering transcription has a fair case.
• If AI summarised, wrote the story or chose quotes, this view does not apply. That is a direct contribution to the content.
• If the transcript is what you publish, say that it was produced automatically and checked. Readers cannot see how far it was checked.

The error is the journalist’s

“The software got it wrong” is no defence. If a published quote is wrong, the journalist corrects it and answers for it. They chose the tool, did the checking and published. When it happens, correct it openly and talk about how it happened inside the newsroom.

Tools that make checking easier

A good transcription tool does not remove the need for checking, but it makes the job easier. The less time checking takes, the lower the chance of error. That is why LiveType offers:

• Telling speakers apart: You can follow who said what. It can be wrong and you do the naming, so check attribution separately.
• Times: Paragraphs carry their times, and in files clicking one moves the player to that moment, so you need not hunt for the sound. Taking timecodes helps on long recordings.
• Correcting as it runs: With editing while transcribing you fix a wrong name when it appears, and later text does not undo it.
• Guards against invention: The engine does not send trailing silence to the model or re-run unchanged speech. These lower the risk and do not remove it.

They have their limits. LiveType removes hesitation sounds, so its text is not strictly verbatim. It does not record live streams, so checking a quote from one needs the broadcaster’s replay or your own recording. Without the original recording, you cannot catch the error. Keep it.

Checklist before publishing

The texts cited here were read on 2 November 2026, so check their current versions at the source. For hands-on steps, see interview transcription and writing a news story from a transcript.