Truthring vs Reality Defender
Of the ten vendors checked on 31 August 2026, Reality Defender publishes the deepest research and the richest result metadata of any of them. Named papers on arXiv, a confirmed ASVspoof5 submission, per-model names and scores on every result, a release version, a unique request identifier, and a status vocabulary that includes UNABLE_TO_EVALUATE — a detector saying out loud that some inputs do not support a verdict.
What we did not find that day: a false positive rate, a methodology document, any per-condition accuracy, or a generator list. Both halves are true at once, and this page holds them together rather than picking the flattering one.
They cover audio, video, image and text. We cover audio and nothing else. That is the shape of the difference, and it is a genuine trade rather than a defect on either side.
How this page is sourced. Every statement about Reality Defender below was checked against their own live pages and developer documentation on 31 August 2026 and carries the source. Where we write not found, it means checked on that date and not present at the URL named — it is not a claim that they publish it nowhere. One item on this page could not be reached at all and is marked as a fetch failure rather than an absence. Corrections: corrections@aivoicedetctor.com
The research, which is the strongest in the field
Most vendors here publish a percentage and a customer logo wall. Reality Defender publishes papers, and the difference matters, because a paper can be read, argued with and reproduced by people who do not work there.
Alethia: A Foundational Encoder for Voice Deepfakes is on arXiv as 2605.00251. Alongside it, FoeGlass and ICLAD, and a confirmed submission to ASVspoof5 dated 10 September 2024. The paper authors are named: Yi Zhu, Brahmi Dwivedi, Jayaram Raghuram and Surya Koppisetti. Leadership is named too — Ben Colman as chief executive, with Milos Fulton, Alex Lisle, Brian Levin and Bennett Morrison.
Submitting to a public challenge means being scored on somebody else’s data, under rules you did not write, with the result published whether it flatters you or not. Truthring has done none of this: we have published a methodology in advance of having any results, which is a different and lesser thing. Founded 2021 in New York, with a Series A expanded to $33m announced 22 October 2024; investors named include IBM Ventures, Booz Allen Ventures, Accenture, DCVC and Y Combinator.
The result metadata, which we intend to copy
This is the part of their product we admire most, and there is no reason to be coy about it.
A result from their API does not arrive as a bare score. It carries models[].name with concrete slugs such as rd-cedar-img and rd-elm-img, a score per model rather than one blended figure, a top-level releaseVersion, and a requestId identifying the submission. If a verdict is challenged six months later, that establishes which models ran, at what version, on which upload. Almost nothing else in this field gives you that.
And the status vocabulary includes UNABLE_TO_EVALUATE. A detector that can decline is a detector whose positive answers are worth more, because it is not being forced to produce a number for audio that cannot support one. Truthring returns an explicit unclear for the same reason, and finding the same intellectual honesty in a competitor’s status enum is genuinely reassuring rather than inconvenient.
They also cover more ground than anyone else checked: audio, video, image and text, the last with a stated 200-word minimum and support for PDF and Word. Breadth here is not a compromise. One vendor across a mixed queue means one integration, one security review and one report format — and cross-modal disagreement is itself a finding a specialist cannot see.
Two published descriptions that do not match
Their marketing describes real-time detection. The API documented for developers, as we read it on 31 August 2026, is upload-and-poll: submit a file, then poll for the result.
We report both as published and deliberately do not reconcile them. A real-time path may well exist that the public documentation does not describe. Guessing which is the case is exactly the inference this page exists to avoid. If latency decides your purchase, put the question to them directly.
One number that is not what it looks like
The only percentage we found on their technology page is 73%, and it is a human baseline, not a product claim. The page presents it as the rate at which humans can detect audio deepfakes, attributed to a University of Florida study. It is a statistic about people.
We flag it because a comparison page quoting it as a product accuracy figure would be misrepresenting them, and the reverse error is just as easy to make.
What each of us publishes
| Published item | Truthring | Reality Defender |
|---|---|---|
| Modalities covered | Audio only, deliberately | Audio, video, image and text — the broadest of any vendor checked. Text requires a 200-word minimum; accepts PDF and Word realitydefender.com · checked 31 August 2026 |
| Peer-reviewed and challenge research | None. Pre-launch | The deepest of any vendor checked. Alethia (arXiv 2605.00251), FoeGlass, ICLAD, and a confirmed ASVspoof5 submission dated 10 Sep 2024. Authors named: Zhu, Dwivedi, Raghuram, Koppisetti arxiv.org/abs/2605.00251; realitydefender.com · checked 31 August 2026 |
| Result metadata | Committed — engine version and a Ringmark reference on every verdict | The richest of any vendor checked. Per-model names and scores, a top-level releaseVersion, and a unique requestId per upload Reality Defender API documentation · checked 31 August 2026 |
| An explicit “cannot judge this” outcome | Yes — an unclear verdict, by design | Yes. Status vocabulary includes UNABLE_TO_EVALUATE Reality Defender API documentation · checked 31 August 2026 |
| False positive rate | Committed, not yet measured. Framework published at accuracy, both directions, by condition | Not found on the pages checked. The only percentage on the technology page is a human baseline (73%, University of Florida study), not a product claim realitydefender.com · checked 31 August 2026 |
| Accuracy by recording condition Codec, noise, clip length | Committed — table published, cells unfilled | Not found on the pages checked realitydefender.com · checked 31 August 2026 |
| Written methodology document | Published in advance of results, versioned | Not found on the pages checked realitydefender.com · checked 31 August 2026 |
| Stated failure conditions | Published as product documentation | Present, but only as an all-caps warranty disclaimer in the terms rather than as product documentation realitydefender.com terms · checked 31 August 2026 |
| Generator coverage list | Committed — named and dated | Not found on the pages checked realitydefender.com · checked 31 August 2026 |
| Real-time or after the fact | After the fact only, on an uploaded file | Both descriptions published: marketing says real-time; the documented API is upload-and-poll. Reported as found, not reconciled realitydefender.com; API documentation · checked 31 August 2026 |
| Named people | One named publicly — team | Named. Ben Colman (CEO), Milos Fulton, Alex Lisle, Brian Levin, Bennett Morrison, plus the paper authors above realitydefender.com · checked 31 August 2026 |
| Third-party certification | None held. Pre-launch — compliance | [VERIFY: fetch failed 2026-08-31, check manually before publishing] trust.realitydefender.com loaded but returned no readable certification list to our fetcher. This is a fetch failure, not an absence. Do not read this row as a statement that they hold or publish nothing trust.realitydefender.com · checked 31 August 2026 |
| Retention period in time units | Not set. Design published, periods unset — data retention | A no-retention and no-training commitment appears in the terms; no period in time units found realitydefender.com terms · checked 31 August 2026 |
| Published price | Yes at launch — pricing published in advance | /pricing returned HTTP 404 realitydefender.com/pricing · checked 31 August 2026 |
Right-hand column checked 31 August 2026 against Reality Defender’s own live pages and developer documentation. Not found means checked on that date and not present at the URL named. Re-checked quarterly.
The question worth asking them
Not “how accurate is the platform”. Ask: what is the false positive rate for the audio path specifically, on audio like mine, and where is it written down?
A platform-level figure blends modalities that fail for unrelated reasons, and an average across them describes no decision anyone actually makes. Given the depth of research they publish, there is a fair chance the audio numbers exist and simply are not on a public page — enterprise vendors routinely release validation material under confidentiality during procurement. Absence of an answer on a marketing page is not itself evidence of anything.
Where Reality Defender is the better choice
- Your input is mixed and you do not get to choose. Trust and safety queues, newsrooms, identity onboarding, insurance claims. Text coverage in particular is rare.
- Cross-modal disagreement is the finding. A face that passes and a voice that does not points at a re-voiced original. Handed the soundtrack alone, we cannot see it.
- You want research you can read. Papers with named authors and a public challenge submission are checkable by people with no stake in the outcome. Very little else here is.
- You need result metadata that survives a dispute. Model names, per-model scores, a release version and a request identifier are what an opposing party will ask for.
- Scale of investment. A funded platform can staff research across four modalities in a way a single-product company cannot.
Where we may be the better choice
- The evidence is a voice and nothing else. A voicemail, a voice note, a call recording — no frames to corroborate it.
- You need condition-level detail. A detector’s error rate on a doubly compressed voice note is a different number from its average, and the difference is what you will be asked about.
- A written error rate matters more than breadth. That is what we commit to publish in both directions before anything ships. As at 31 August 2026 we could not find one for their audio path.
Every item above is a design commitment made before launch, not a measured result. Nothing on this site claims otherwise.
Questions
Does Reality Defender publish a false positive rate?
We did not find one. Checked 31 August 2026 across realitydefender.com and their developer documentation, no false positive rate, no per-condition accuracy and no vendor accuracy claim of any kind was present on the pages we could reach. That is a statement about what we found at those URLs on that date, not a claim that no such figure exists anywhere.
What is the 73% figure on their technology page?
It is a human baseline, not a product claim. The page presents it as the accuracy rate at which humans can detect audio deepfakes, attributed to a University of Florida study. Reading it as a statement about the product would be wrong, and any comparison page that quotes it as one is misleading you.
Is Reality Defender real-time or upload-and-poll?
Both descriptions are published and we are not going to reconcile them for you. Their marketing describes real-time detection; the API documented for developers, as we read it on 31 August 2026, works by uploading a file and polling for a result. If latency matters to your use case, that is a question to put to them directly rather than to infer from either page.
What does Reality Defender publish that Truthring does not?
Peer-reviewed research, for a start. Alethia, a foundational encoder for voice deepfakes on arXiv, plus FoeGlass, ICLAD and a confirmed ASVspoof5 submission, with authors named. Also the richest result metadata of any vendor we checked: per-model names and scores, a release version, and a unique request identifier on every submission. Truthring has published no papers and has not launched. That gap is real and we are not going to talk around it.
Why does their UNABLE_TO_EVALUATE status matter?
Because it is a detector admitting, in its own API, that some inputs do not support a verdict. That is the same discipline behind Truthring returning unclear rather than forcing a call on a clip that is too short or too degraded. Most detectors return a number no matter what they are given, and a number produced under those conditions is worse than no answer, because somebody will act on it.
Which one is more accurate on audio?
Nobody can answer that from outside, us least of all. Truthring is pre-launch and has no measured figures. Even between two shipping products it would take the same held-out clips at the same thresholds, scored by a party with nothing riding on the result.
Sources
- realitydefender.com — modality coverage, the 73% human baseline, leadership, research index, funding. Checked 31 August 2026.
- arXiv 2605.00251 — Alethia: A Foundational Encoder for Voice Deepfakes; named authors. Checked 31 August 2026.
- Reality Defender developer documentation — per-model names and scores, releaseVersion, requestId, UNABLE_TO_EVALUATE, upload-and-poll flow. Checked 31 August 2026.
- Reality Defender terms of service — failure conditions as a warranty disclaimer; no-retention and no-training commitment without a period. Checked 31 August 2026.
- trust.realitydefender.com — could not be read on 31 August 2026. Marked for manual check; nothing about certifications is asserted from it either way.
Claims about Reality Defender checked 31 August 2026 against publicly available material. Re-checked quarterly. Corrections are published with a date rather than made silently.