Truthring vs free browser detectors

This page used to name nobody, which made it safer and less useful. It now names four specific tools and reports what each of them had published on one date, because a category argument that cannot be checked against particular sites is just an assertion with a wider target.

What all four do publish, and why it is not enough

The four sites are more alike than their branding suggests. Every one of them, on 31 August 2026, offered a free check with no sign-up, listed the audio formats it accepts, gave guidance on how long a clip should be, printed one hedging sentence to the effect that no detector is 100% accurate, included a vague line about files being deleted after analysis, and described in general terms the kinds of signal the model looks at.

That set is not nothing. It tells you how to use the tool and it warns you not to over-read the answer. What it cannot tell you is how much weight the answer deserves, because every item in it is a description of the product rather than a measurement of it. “Not 100% accurate” is true of every detector ever built, including ours; it is compatible with being wrong two times in a hundred and with being wrong forty. “Deleted after analysis” is compatible with four seconds and with four months. A list of signals analysed is compatible with any performance whatsoever.

None of the four led with a numeric accuracy claim, which is to their credit — and none of the claims any of them did make carried an attached source, dataset or test date, which is the more consequential fact. The gap here is not honesty. It is measurement, and measurement costs money: a labelled evaluation set covering clean audio, telephone codecs, forwarded voice notes and background noise is expensive to build and expensive to keep current as new generators appear. A free tool has no revenue to spend on it. That is a structural explanation, not an excuse, and it is why price and disclosure track each other loosely rather than tightly.

EyeSift, which does something the others do not

Checked 31 August 2026, EyeSift publishes a methodology page at eyesift.com/methodology/ that carries a refresh date — “refreshed May 31, 2026” — and version identifiers for its detectors, including “text detector v6.3.5”. A dated, versioned method document is precisely the artefact this site argues for, and finding it on a free cross-modal tool rather than on a paid audio specialist is a useful corrective to any assumption that price buys disclosure.

They also write down what the tool cannot do. On the same date the site advised readers to “use this free AI voice detector as a first-pass screen, not as forensic proof”, and disclosed that it “does not decode SynthID watermarks or cryptographically verify C2PA manifests”. That second sentence is unusually specific: it names two provenance mechanisms the tool does not implement, which stops a reader assuming a capability the product does not have. Very few vendors at any price volunteer a list of the things they do not check.

The gaps we found remain. On 31 August 2026 EyeSift published no false positive rate, no per-condition accuracy, no dated generator list and no decision thresholds, and it is counted in every zero on the scorecard above. But it is the clearest instance in this category of a free tool being straight with its users about scope, and a comparison page that could not say so about anyone would not be worth reading.

When free is the right answer

Say this plainly, because comparison pages rarely do: most people who reach a voice detector should not spend money. If your situation is on this list, use a free tool and keep your cash.

• Idle curiosity. A clip sounded odd and you want to know. Nothing follows from the answer, nobody is accused, and an audit trail is overhead you have no use for.
• The audio is public. A political speech, an advert, a podcast, something already circulating. There is no privacy to protect and no person exposed by the upload.
• Zero stakes. No money moves, no employment decision hangs on it, nobody is confronted, nothing is filed. Disclosure only earns its keep when a result will be used against someone.
• You want corroboration. A second detector is a genuinely good habit, and free makes it affordable. Two tools agreeing raises your confidence; two disagreeing tells you the audio is marginal, which is worth knowing on its own.
• Budget is the binding constraint. If there is no money, a free result held loosely is better than no signal at all — provided you never present it to anyone as proof.

Anyone who tells you that a paid tool is always the responsible choice is selling one. The variable that matters is not the price; it is whether a person will be judged on the output.

Using a free result without doing harm

Where we intend to be the better choice

• The result will reach someone with authority over a person. An employer, a bank, an insurer, a court. That needs a written method, both error rates and a record that can be reopened.
• The recording is private. A stated retention period and a written position on training are the difference between a transaction and an act of faith.
• You need the tool to admit uncertainty. We return unclear where the audio will not support a verdict, rather than converting poor input into a confident number.

Questions