The Machine Transcript Is a Witness Too: Auditing AI in the Human-Rights Archive

There is a particular kind of silence in an archive: the pause before a survivor begins to speak, the breath a lecturer takes before the passage that matters. Anyone who has sat in a hall in Oxford, or listened back to a recording months later, knows that the gaps carry meaning. Machine transcription is very good at filling them — with punctuation, with guesses, with plausible words that were never said.

Human-rights organisations have embraced automated transcription for understandable reasons. A small campaigns team with a shoebox of audio and a threatened budget can now produce a searchable record in an afternoon. For an archive like ours, which holds decades of Amnesty lectures, panels and testimony, the temptation is obvious. Yet a transcript is not a photocopy of speech. It is a reading of it, produced by a system with its own assumptions, and it becomes the version that researchers, journalists and lawyers will cite for years.

That is a human-rights question, not merely a technical one. Under the UK GDPR and the Data Protection Act 2018, voice recordings and the transcripts made from them are personal data; under Articles 8 and 10 of the European Convention on Human Rights, as given effect by the Human Rights Act, privacy and expression are engaged on both sides of the microphone. The following seven tests are offered as a practical audit for anyone in Britain who is putting a machine to work on the spoken record.

Test One: Has Consent Named the Machine?

A speaker who agrees to be recorded for an event has not automatically agreed to be processed by a third-party AI service, possibly hosted abroad, possibly used to train a model. Consent obtained in 2019 cannot reasonably cover tools that did not exist. Revisit permissions for archived material before it is fed into a new pipeline, and make the wording specific: which organisation, which purpose, which retention period, which categories of onward use. Where consent is not the lawful basis — for example with legacy recordings held under legitimate interests — record that reasoning in writing and be prepared to defend it.

Test Two: Run an Accent Audit, Not a Demo

Vendors demonstrate their systems on clean studio audio. Human-rights archives are not clean. They contain Birmingham and Belfast voices, Welsh-language passages, Scots, British Sign Language interpretation, whispered questions from the floor, and speakers for whom English is a fourth language. Before adoption, test the tool against a sample that reflects your actual holdings — at least ten recordings, deliberately chosen to be awkward. Measure word error rates by accent and language, and ask what happens to names. The mishearing of a place, a family name or an allegation is not a cosmetic glitch; it is a factual error with consequences.

Test Three: Context Is Not Decoration

Speech-to-text flattens the things that make testimony legible: irony, hesitation, interruption, the translator speaking over the witness, the chair cutting off a hostile question. Punctuation inserted by an algorithm can invert meaning. Decide in advance which events require a human-verified transcript and which can survive as machine output labelled as such. Where the record matters — legal, historical or reputational — insist on a second pass by someone who was in the room or who has the cultural and linguistic competence to hear what the machine could not.

Test Four: Read the Vendor Contract Like a Witness Statement

The critical clauses are rarely the ones in the sales deck. Ask where processing occurs, whether audio is retained, whether the provider trains models on your material, who can access it, and how deletion is evidenced. UK public bodies also need to consider the Cabinet Office's cloud security principles and their own DPIA obligations. For charities and event organisers, the questions are simpler but no less urgent: could you produce a lawful answer if a speaker asked, in writing, what happened to their voice?

Test Five: Retention Without a Kill Switch

Automation multiplies copies. A recording becomes a transcript, a summary, a search index, a backup, a vendor log. Set a retention schedule before ingestion, not after, and make deletion a tested procedure rather than a policy sentence. Sensitive testimony — from people at risk, from whistleblowers, from those who spoke on condition of anonymity — should be segmented and, where possible, processed locally or transcribed by hand.

Test Six: Access and Redaction Are Rights, Not Extras

A searchable archive is a public good, but full-text search also makes it easier to find what should not be found. Build redaction into the workflow: names of private individuals, safeguarding details, health information, anything given in confidence. At the same time, respect the access side of the ledger. Transcripts published on a website should meet accessibility standards under the Public Sector Bodies Accessibility Regulations, and should be offered in formats that work for disabled readers rather than as an unlabelled wall of text.

Test Seven: Keep a Named Human at the End of the Chain

Every automated record needs an accountable owner: a person whose name appears in the metadata, who signs off publication, and who can be contacted when an error is found. Publish a correction route. Log which transcripts are machine-generated and which are verified. When the technology changes — and it will, repeatedly — the audit should be repeated, not assumed.

None of this is an argument against the tools. It is an argument about who holds the pen. Amnesty's own history is a reminder that the written record is a form of protection: it can shield a witness or expose them, dignify a testimony or distort it. If we let a model write the first draft of that record, we owe it to the people who spoke to make sure a human remains the editor, the conscience and, where necessary, the correction.

The lecture hall falls quiet again. What we choose to write down afterwards is still our responsibility.