Truthring vs aivoicedetector.com

We make a competing product, so this is not a neutral document. It compares what each of us has put in public, at a stated URL on a stated date. Nothing rests on our word — every finding can be re-run in a browser in about ten minutes.

Start with what they do better

Their privacy policy, marked v1.0 and dated 25 May 2026, states that submitted audio is “deleted within 24 hours by default”, with a one-way 256-bit hash kept afterwards. That is a commitment expressed in a unit of time, under a version and a date. Our own retention page is honest that our windows are not fixed yet — the correct thing to write and a worse thing to offer. On this row they are ahead of us.

Second, on 31 August 2026 their /detect page named twenty specific generator systems. Naming the systems a detector was built against is more informative than counting them, and most of this field counts. The list carries no date and the copy says “24+” where twenty are named — but naming twenty beats naming none.

The document the argument rests on

Most of the table is ordinary. Plenty of detectors omit a false positive rate; plenty name nobody. One row is different in kind, because it is not an omission but a load-bearing claim.

Their case for being usable in a dispute reads: “Each verdict carries a methodology version stamp, so the same test can be run against the same snapshot later, which is what makes it defensible.” The homepage compresses that to “Defensible methodology in deposition.” The whole structure depends on a snapshot existing somewhere a third party can read. A version number pointing at nothing cannot be re-run by opposing counsel, a regulator, or you.

So we looked, on 31 August 2026. Ten plausible paths — the ones a reader would try and a search engine would index — each returned HTTP 404. Their sitemap.xml lists 53 URLs and includes none of them. The footer link reading “Read the methodology” lands on their about page; a second link labelled with the word resolves to a blog post. That is four routes to the document, and on that date none arrived.

We are stating what those requests returned on one day, not what exists on their servers. The document may sit somewhere we did not look, or behind a login, and if so we will correct this page and date the correction. But a snapshot a reader cannot locate does not do the job the claim assigns it, and a deposition is exactly the setting where unlocatable and absent come to the same thing.

The check is worth learning regardless of vendor. Find the sentence claiming a published method, find the link, follow it. That sequence separates rigour from a claim about rigour, and it takes less time than reading the page the claim appears on.

What is missing from “99% on clean audio”

Their headline, verbatim on 31 August 2026: “99% on clean audio across 24+ generator models. We always report the confidence number; we never claim certainty we do not have.” The second sentence is a good instinct. The difficulty is not the figure but everything not attached to it.

A detection number measures one model against one set of clips at one threshold on one date. Strip those away and 99% is unfalsifiable: no dataset, no sample size, no test date, and no reachable method to consult, so it cannot be reproduced, challenged or aged. It is also one-directional. A detection rate rises simply by flagging more aggressively, and that is paid for in genuine speech wrongly called synthetic — the quantity that decides whether a detector can be pointed at an employee or a witness. On 31 August 2026 the phrase “false positive” appeared nowhere in the eight pages of theirs we read.

We are not asserting the figure is untrue; we have no way to know, and saying otherwise would be the same species of unsupported claim. We are saying a number with no dataset, no sample size, no date and no reachable method gives a reader nothing to work with — and that “clean audio” is the easiest condition there is, least like the compressed forwarded voice note in most people's actual problem.

Where they may be the better choice

• You want a deletion window you can point to today. Their 24-hour default is published, dated and specific. Ours is not yet, and if that decides it, it decides in their favour.
• You want to know which generators a tool was built against. Twenty named systems beats a count, and it is on their site now.
• They are shipping and we are not. A live product, a free tier of 50 checks a month, paid plans from $49 to $599 a month, and years of submissions. Truthring is pre-launch, and a working tool today beats a stated intention today.
• This page may already be stale. If they publish a false positive rate and a methodology URL that resolves, most of the argument above disappears — and the category is better for it. Re-run the checks yourself.

Where we intend to be the better choice

• When the verdict has to survive being questioned. What holds up is a readable method, both error rates by condition, a version stamp and a reference that reopens the analysis — not a percentage.
• When the audio is realistic rather than clean. A verdict on a twice-forwarded voice note is a different claim from one on a studio file, and a single figure across both misleads by construction.
• When you need the tool to admit it cannot tell. We return unclear where the audio will not support a verdict, and unknown generator where we cannot name the system.

Questions