The number that should govern your process is the false positive rate
Not the accuracy figure. You are running a process whose failure mode is an individual person losing their job over something they did not do, and the only statistic that speaks to that is how often the model calls a genuine recording synthetic.
This page touches on legal duties and is not legal advice. Take your own advice in each jurisdiction where you employ people.
Why this is landing on your desk
Employees record things. A grievance now routinely arrives with a voice note attached: a manager saying something in a one-to-one, a colleague on a call, a voicemail left after hours. Until recently the argument was about context and about whether the recording was permitted. Now there is a third argument available to anyone who does not like what a recording contains, and it costs nothing to make: that is not me.
You are therefore exposed at both ends of the same case. A fabricated recording used to support a malicious complaint. And a genuine recording dismissed as fabricated, leaving a complainant with real evidence they cannot get taken seriously. Both outcomes land on the investigating officer, and neither is solved by buying a tool.
What this cannot do for you — and why it matters more here than anywhere else
It cannot establish that anyone did anything. It analyses audio. It does not know who is speaking, whether the words were said in the way the complainant describes, or what happened around them.
Its false positive rate is not zero, and it never will be. A genuine recording made on a poor connection, over a speakerphone, in a noisy room, or forwarded through two messaging apps can attract a synthetic verdict. We publish that rate rather than hide it. [VERIFY: state the false positive rate and the audio condition it was measured on]
The person on the wrong side of that rate cannot prove a negative. There is no evidence an innocent employee can produce that establishes a recording of them is real. If your process allows a verdict to carry the decision, you have built a process with no defence available to the accused.
A likely human verdict does not clear a recording either. It means no known signature was found. It is the weaker verdict and should not be reported to a panel as authentication.
It cannot make the decision. A named human decision-maker considers the analysis alongside everything else and reaches a conclusion they can explain. That is not a formality; in several regimes, decisions with significant effects on a person taken solely by automated means are restricted, and the restriction exists precisely because of the paragraph above.
What it can honestly add
Used within those limits, an analysis does two useful things. It gives an investigating officer a documented, dated, technical view on a question they are otherwise answering by ear — and people are close to chance at identifying synthetic speech by ear. And it creates a record that the question was asked properly, which matters when the process is reviewed later.
It is also a reason to look harder in a specific direction. A synthetic verdict on a recording submitted with a complaint is not a finding against the complainant; it is a prompt to establish where the file came from, on what device, and whether the original still exists. A likely human verdict on a recording the respondent denies is not proof, but it does mean the denial now needs its own explanation.
Your duties when you use it
| Duty | What it means in the case file |
|---|---|
| Transparency | Tell both parties the analysis was run, before the outcome. Put the report in the bundle they see. Covert analysis can invalidate a sound process. |
| Lawful basis | Analysing a voice recording is processing personal data in many regimes. Record the basis you rely on and check it covers a recording of someone who is not the complainant. [VERIFY: verify under your applicable regime] |
| Human decision-maker | Name the person who decided. Record what else they relied on. The verdict must be visibly one input among several. |
| Right to challenge | Give the employee the reference code, the engine version and the error rate, and record their response in the file. |
| Proportionality | Do not analyse recordings beyond the ones actually in issue. A grievance is not a mandate to screen an employee’s call history. |
| Retention | The recording, the report and the reference code follow your investigation retention policy, not a longer one because the file is small. |
Wording that keeps a process defensible
How the finding is written into the investigation report matters as much as the finding. Two examples of the difference:
Do not write
“The recording was found to be an AI fake, so the complaint is not upheld.”
States a conclusion the evidence does not reach, and makes the tool the decision-maker.
Write instead
“An automated analysis (Truthring, engine [VERIFY], ref [VERIFY]) returned a synthetic verdict at [VERIFY] confidence. The stated false positive rate for audio of this quality is [VERIFY]. Taken with the absence of the original file and the account given at interview, I concluded on the balance of probabilities that…”
Attributes the finding, states its uncertainty, and leaves a person accountable for the conclusion.
Questions from HR
Can we use this as the sole ground for dismissal?
No. This is not a limitation we are being modest about — it is the one use we would refuse. See the position stated on the compliance page: permitted with transparency and a human decision-maker, never as the sole ground for an adverse outcome.
What if the recording is too poor to analyse?
You will get unclear, and that is the correct answer rather than a failure. Record it in the file as inconclusive and decide the case on the rest of the evidence. Do not run the clip repeatedly hoping for a different verdict.
Should we tell employees we may do this?
Naming it in your investigation and disciplinary policy is cleaner than introducing it mid-case, and it gives you something to point to when transparency is questioned. [VERIFY: check with your employment counsel before amending policy]
Can a manager run a check themselves?
Better that it sits with the investigating officer, logged, than with a line manager acting on a suspicion. Ad hoc checking by managers is the pattern most likely to produce an unrecorded false positive that later shapes a decision nobody can account for.
Before you write it into policy
Read the limitations page in full, and have the false positive figure in front of the panel whenever a verdict is discussed.
Reviewed