The clip arrived at 4pm and the desk wants it by six
A detector will not tell you whether to publish. It will narrow the question, put a number on one part of it, and give you something specific to write in the piece. Everything else is still your job.
Your actual exposure
You are exposed in two directions and they pull against each other. Publish a fabricated recording and you have handed your masthead to whoever made it. Sit on a genuine one because you could not close the verification loop in time and someone else runs it, or nobody does.
The material that reaches a newsroom is also the worst material for analysis. It is rarely the original file. It has been forwarded through a messaging app, re-encoded, sometimes screen-recorded off a social platform with a video player’s compression on top. By the time it lands in your inbox the signal a detector reads has been partly scrubbed by the distribution chain, not by anyone trying to deceive you.
That is worth knowing before you run it, because an unclear verdict on a forwarded clip is a statement about the copy you were given, not about the recording. If you can get the source to send the original file rather than a re-share, do that first. It changes the answer more than any other single step.
What to write in the piece
If a detection result appears in your story, it should appear with enough detail that a reader — or a rival newsroom, or the subject’s lawyer — can interrogate it. That means four things in the copy or the methodology note:
- The tool, named. “Analysis by a detection tool” is not a source. Name it.
- The engine version. Models are retrained. A result from v2.4 is not the same claim as a result from a later build, and the version is what makes the finding reproducible.
- The error rate for that audio condition — not the headline accuracy figure. Telephone audio, short clips and re-encoded files carry different rates. Ask for the one that applies to your clip. [VERIFY: link the accuracy page section that gives per-condition rates]
- The reference code. Every Truthring report carries one. Publishing it lets anyone pull the same report and check that you have described it accurately.
Write the verdict as what it is. “A detection analysis found patterns consistent with synthetic generation, at [VERIFY] confidence” is defensible. “The recording is an AI fake” is a different sentence and the evidence does not reach it.
The harder problem: a source who denies a genuine recording
Most of the anxiety about synthetic speech is about fakes being believed. The editorial problem that will actually land on your desk is the reverse. You have a recording you are confident is real. The subject says it is a deepfake and demands you prove otherwise.
A detector cannot get you out of this. A likely human verdict means no generation signature was found in the audio. It is the weaker of the two positive results, and a synthetic clip that has been cleaned up after generation, or made with a system newer than the last training run, can produce it. Handing that verdict to your lawyer as proof of authenticity is asking it to carry weight it does not have.
What rebuts a denial is provenance. Who made the recording and on what device. Whether you hold the original file rather than a copy. What the file’s own metadata and codec history say about where it has been. Whether anything in the content can be independently corroborated — a detail only the speaker knew, a fact confirmed by a second person in the room, a timestamp that matches something else. The detection result is one line in that case, and not the strongest one.
Plan for the denial before you publish, not after. If your answer to “how do you know it is real” is only a detector score, you are not ready.
What this cannot do for you
It cannot establish that a recording is authentic. It cannot tell you who is speaking, who made the clip, when it was made, or why. It does not read the words, so it has no view on whether the content is true.
It cannot tell you whether a clip was edited before it reached you, only sometimes that it was. It will often return unclear on exactly the compressed, forwarded, short material a newsroom receives — and on those clips likely human should be read as inconclusive rather than reassuring.
It is not a substitute for a source. A detection result cannot be the reason you publish, and it cannot be the reason you spike. The editorial judgement is made by a person who can be named, questioned and held to it.
Where it fits in the desk process
Chase the original
Ask the source for the file as it left the device, not a re-share. Record how it reached you. Preserve your copy untouched before anyone converts anything.
Run the analysis early
A verdict is worth having before you invest a week in the story, not the hour before the legal read. Note the engine version and reference code in the file.
Verify around the audio
Source, timestamp, second witness, documents. Treat the verdict as one strand. If the story stands only on the score, it does not stand.
Publishing a claim about someone based on a recording carries defamation exposure that varies considerably between jurisdictions, and a detection verdict does not reduce it. [VERIFY: verify the position with your own legal team in the markets you publish in] Nothing here is legal advice.
Questions from newsrooms
Can I cite a Truthring result in the piece?
Yes, with the version, the applicable error rate and the reference code. A verified report includes a file hash, a timestamp, the engine version, the limitations that applied to that clip, and a permanent URL a reader can check. Cite the analysis, not a conclusion the analysis does not support.
What if we get a different verdict from a second tool?
Report the disagreement. Two detectors trained on different data will diverge, especially on degraded audio, and a split result is genuine information about how uncertain the question is. Suppressing the one that disagrees with your story is the thing you will be asked about later.
Does submitting a clip put our source at risk?
Audio will be deleted after analysis and a one-way hash retained [VERIFY: retention not yet set] so a report can be tied to the exact file. Check the privacy page and security page against your own source-protection policy before uploading anything sensitive, and consider whether the clip itself identifies the source.
The subject wants our analysis before publication. Do we share it?
That is an editorial call, but the reference code makes it easy: they can pull the same report you did. Offering it tends to strengthen a right-of-reply process rather than weaken it, because it moves the argument onto the audio rather than onto whether you checked.
Run one before the deadline, not after
A free tier is planned at three checks a day and will not need an account. If the clip is going into a story, take the verified report so the reference code exists in public.
Reviewed