AI transcription for journalists: the recording that stays yours

The NUJ Code of Conduct requires journalists to protect confidential sources. Article 10 of the European Convention on Human Rights protects the journalist-source relationship as a component of press freedom. UK courts have consistently upheld the principle that source confidentiality is not simply a professional courtesy — it is a legal protection that enables journalism that serves the public interest.

A source who speaks to a journalist on condition of anonymity has a reasonable expectation that the recording of that conversation stays within the journalist's control. What they probably have not considered, and what the journalist may not have considered either, is whether their voice is being processed by an AI transcription service operating on servers in the United States.

Cloud AI transcription tools are fast and accurate. Most journalists who use them are not doing so carelessly — they're making a practical decision to reduce the 6–8 hours per hour of audio that manual transcription requires. But source protection is not a practical consideration. It is a professional obligation, and cloud processing of source audio creates a legal exposure that most journalists have not evaluated.

The transcription workflow that most journalists use

A journalist records an interview — on a phone, a Zoom call, or a dedicated audio recorder. The recording is then transcribed, either manually or via an AI service. The transcript is used for quote verification, story construction, and reference.

Manual transcription takes 6–8 hours per hour of audio. This is a well-established benchmark across the industry — a 90-minute interview generates 9–12 hours of transcription work. At that rate, transcription is a significant fraction of the total time invested in a story.

AI transcription services have changed this. Tools like Otter.ai, Rev, and Descript can produce a transcript in minutes. The accuracy is not perfect — accents, crosstalk, and audio quality affect the output — but for most purposes, AI transcription with human review is faster than manual transcription by an order of magnitude.

The trade-off: your audio goes to a cloud server.

What cloud transcription means for source protection

When you upload a recording to a cloud AI transcription service, the audio is transmitted from your device to the company's servers. The transcription is generated on those servers. The audio may be retained according to the company's data retention policy. The resulting transcript may be used to train the company's AI models, depending on the terms of service.

For a journalist recording an interview with a corporate whistleblower, a government source, a witness to a crime, or a survivor of abuse — this transmission is a source protection problem.

The US CLOUD Act. Major AI transcription services — Otter.ai, Rev, Descript, Fireflies — are US-incorporated companies. The US CLOUD Act (Clarifying Lawful Overseas Use of Data Act) permits US federal agencies to compel US companies to disclose data stored on their servers, even when that data belongs to non-US citizens and is stored on servers outside the United States. This is not a theoretical risk. It is the legal framework under which US intelligence and law enforcement agencies can access cloud data, including audio recordings, without the knowledge of the data subject or the journalist.

For a journalist covering national security, organized crime, government corruption, or any subject where source identity could place that source at risk — the CLOUD Act is relevant to every cloud service they use.

The UK Investigatory Powers Act. UK journalists face an additional domestic risk. The Investigatory Powers Act 2016 provides UK intelligence agencies with broad powers to intercept communications and compel technology companies to provide data. The journalistic privilege provisions in the IPA are weaker than the source protection provisions in older legislation, and case law interpreting these protections remains unsettled.

The data breach scenario. Cloud services are breached. When a cloud AI transcription service suffers a breach, the audio and transcripts stored on its servers become accessible to whoever conducts the breach. A journalist's full interview archive — every source, every conversation, every unreported detail — is a high-value target for state actors, organized crime groups, and corporate intelligence operations.

The subpoena scenario. UK and US courts have ordered technology companies to disclose data held on their servers in legal proceedings. If a journalist's source recordings are on a cloud server, they exist as a target for a subpoena. A recording that lives only on an encrypted local device cannot be disclosed via a third-party subpoena, because the third party doesn't have it.

Security by architecture

Journalists who are serious about source protection already understand one principle: the safest data is data that doesn't exist on anyone else's systems.

Signal is end-to-end encrypted because the messages are encrypted before they leave the device — Signal's servers see only encrypted ciphertext. Proton Mail is trusted by journalists for the same reason. The mental model is: if the service cannot read your data, it cannot be compelled to disclose your data.

On-device AI transcription applies this principle to recordings. When a recording is processed entirely on the journalist's device — when the AI model runs locally and the audio never leaves the phone — there is no server to subpoena. There is no data breach that exposes the audio. There is no CLOUD Act or IPA order that can compel a third party to disclose audio that was never in a third party's possession.

This is the architecture that Kuulo uses. The AI models for transcription, summarization, and speaker diarization run on the Neural Engine of the journalist's iPhone or Mac. The audio stays on the device. The transcript stays on the device. The summary stays on the device. Nothing leaves.

The practical workflow

Recording the interview. Kuulo records from the iPhone's microphone directly. This means any in-person interview, phone call with a recording app, or environment where audio can be captured by a phone microphone.

Transcription. Kuulo transcribes in real time, on-device, with no internet connection required. A 90-minute interview generates a full transcript by the time the interview ends. Speaker diarization attributes the journalist's questions to one speaker and the source's answers to another — the most time-consuming part of manual transcription is already done.

Summary and quotes. After the interview, generate a structured summary: key claims made, direct quotes highlighted, factual assertions for verification, and the arc of the conversation. This is the basis for quote verification and story construction, generated in under two minutes from a 90-minute interview.

Archive. The recording and transcript live on the journalist's device. Encrypted on an iPhone under Apple's Secure Enclave, accessible only with the journalist's biometric authentication. Not on a server anywhere.

The import path for existing recordings

Many journalists record interviews on dedicated voice recorders, on Zoom with the recording saved locally, or on a phone's native voice memo app. Kuulo supports audio file import — an existing recording can be imported and processed on-device, applying the same local processing to audio that was originally captured on another device.

This is significant for journalists who have a backlog of interviews recorded on tools without AI transcription, or who use a dedicated recorder for audio quality reasons but want AI processing without cloud exposure.

Broadcast research: archive, inquiry transcription, long-form

Documentary journalists and broadcast researchers face a specific transcription challenge: large volumes of recorded material — archive footage, inquiry hearings, planning applications, council meetings, court proceedings — that must be transcribed for story development.

A three-hour public inquiry session generates a transcript that is useful for finding key moments, cross-referencing with other sources, and building a timeline. Manual transcription of three hours of audio is a full working day. Kuulo processes the same audio in minutes on-device.

For sensitive investigations — where the inquiry relates to state actors, national security matters, or sources who participated in the public inquiry but whose specific contributions the journalist does not want processed by a cloud service — on-device transcription is the only appropriate tool.

The speed argument

Journalists who use cloud AI transcription do so because the speed is genuinely valuable. Reducing 9 hours of transcription work to 10 minutes of processing time has a real impact on how quickly a story can be developed and published.

On-device AI transcription on Apple Silicon achieves processing speeds comparable to cloud services for standard-quality audio. A 90-minute interview is transcribed in real time — the transcript is available by the time the conversation ends. There is no meaningful speed trade-off between cloud and on-device on current iPhone hardware.

The speed case for cloud AI transcription — the main practical argument in its favour — does not hold against on-device alternatives. The privacy case for on-device — the legal, professional, and source-protection argument — is substantial. When the tools are equally fast and one of them doesn't create a server-based record of your sources' voices, the choice is not a trade-off. It is the obvious conclusion.

Practical note: recording consent

In England and Wales, it is legal to record a conversation you are party to without disclosing that you are recording, unless you intend to share the recording with a third party in a way that would breach the other person's reasonable expectation of privacy. For journalism purposes, recording an interview for your own transcription and quote verification purposes is generally permissible; sharing the recording externally raises different considerations.

Many journalists inform interviewees that they are being recorded as a matter of professional practice, regardless of the legal requirement. With Kuulo, the complete disclosure is: "I'm recording this for my notes. The recording stays on my phone — nothing goes anywhere else." This is accurate, and it is substantially more reassuring than the alternative, which involves an AI company's servers and data retention policy.

The NUJ guidance on recording and source protection is available at nuj.org.uk. Journalists working in high-risk environments — covering state actors, organised crime, or national security — should additionally consult the Committee to Protect Journalists' digital safety resources for a fuller security model.

The interview is on your device. It stays on your device. That is the level of source protection the profession requires.