How to read this page
Every fact about Resistant AI below carries the date it was read, and names the page it was read on. Prices, claims and statistics in this market move; the date is what tells you how much to trust a line and what to re-check.
Their accuracy figures are repeated here only as claims. "99.92% accuracy" (AWS Marketplace listing) and "<1% false positives" (their generative-AI page) come with no population, ground truth or method that we could find (read 2026-10-12). A rate with no stated population is a claim, not a measurement, and that applies to every vendor including us. Ours are at /evidence, with the population, the date, the engine version and what they do not cover.
Side by side
| Tamperlens | Resistant AI | Stronger | |
|---|---|---|---|
| Price | Published. Free tier of 50 documents a month; Solo is $29 for 1,000 documents a month, which is $0.029 a document (/pricing). | No rate card on their own site (read 2026-08-06). Their one public price is the AWS Marketplace "Document Forensics" listing: four dimensions (document forgery analysis, customized forgery analysis, document quality analysis, document classification analysis), each $10,000 a month on a 1-month contract, with a private offer encouraged for custom pricing (read 2026-10-12). The listing also offers 12, 24 and 36-month contracts, whose prices are not shown, and its AWS-generated "AI Insights" text mentions per-request billing, so the effective price is unknown and is set in private offers. A flat monthly fee per dimension is not a per-document price, and we cannot derive one from it. | Tamperlens |
| How you buy it | Sign up, get a key, call the API. No conversation required. | Demo-first: "Book a demo" on their product page, or "Request a private offer" through AWS Marketplace. We found no free trial and no self-serve sign-up (read 2026-10-12). | Tamperlens |
| Scale behind the analysis | One engine, and a corpus we measured ourselves: 1,728 documents nobody curated, plus 705 published documents. | "170 million+ documents processed since our founding in 2019": Global Document Fraud Report 2026, published 26 Feb 2026, updated 22 Jul 2026 (read 2026-08-06). Their product page now says "trained on over 200 million documents" (read 2026-10-12). | Resistant AI |
| Breadth of analysis | Structure and metadata: 20 signal families on a PDF, 19 of them on every PDF and the last only with a private issuer baseline, plus image containers and Office files. No classifier, and no data from other customers. | A far broader platform: model-based document forensics (they describe pixel-structure anomalies and detection of images from named generative models), serial fraud correlated across their customer base, and threat intelligence. We have no equivalent of the forensics or the threat intelligence, and our serial alert works only inside one account, never across customers. | Resistant AI |
| Enterprise assurance | None. No SOC 2, no ISO 27001, one server in Finland, no on-premise build (/security). | An enterprise programme sold to banks and regulated lenders. We cannot inspect its contents from outside, which is exactly what a procurement team is buying. | Resistant AI |
| What happens to the file | Held in memory, analysed, discarded with the response. Without an account the document does not survive the request; an account keeps a short derived record (no text, no filename), and the file, encrypted, only if it opts in (/privacy). | We could not establish a retention position from their public pages (read 2026-08-06), so we do not state one here. | — |
| False-positive rate | Published at /evidence, with the population and the date it was measured on. | No rate with a stated population. Their 2026 report does state the basis of its analysis, which is more than most of this field does, and they cite "<1% false positives" for generative-AI detection, but no population or date goes with that figure (read 2026-10-12). | Tamperlens |
| Recall | Measured per forgery operation at /evidence: 100% on an incremental-update edit, a spliced image and an edit with the metadata stripped, 89.4% through a Form XObject, 71.1% after a qpdf rewrite, 65.5% after a Ghostscript re-distill. Built forgeries, not real fraud (engine 1.23.0, August 2026). | No recall by operation that we could find. Their product page claims "3x more fraud detected", without a stated baseline (read 2026-10-12). | Tamperlens |
| Where the engine also runs | Directly, through the API, the n8n node and the MCP server. | Inside Ocrolus Detect since 2026-10-01, and listed with ABBYY, Tungsten Automation and Experian (read 2026-10-12). See the next section. | Resistant AI |
One engine, two contracts: Ocrolus Detect
On 1 October 2026 Ocrolus announced that its Detect product now embeds Resistant AI's Documents forensic engine, "now live across Ocrolus' customer base". Ocrolus's own documentation splits the work in two: the Resistant engine operates at the file and pixel level (metadata anomalies, digital modifications, suspicious PDF structure), and Ocrolus's own algorithmic checks operate at the content level (balances that must reconcile, tax arithmetic, date validity). The two are merged into one Authenticity Status, highest risk wins.
For a buyer that means the same forensic engine can reach you through two contracts: directly from Resistant AI, or inside Ocrolus for bank statements, pay stubs and W-2s. If you already use Ocrolus Detect, you may already be using it. The Ocrolus side of the picture, including the content layer that this engine does not supply, is at /alternatives/ocrolus.
Where Resistant AI is the better answer
- You need a platform, not a signal. Identity, transaction and document risk under one contract, with a customer-success function attached, is a thing we do not sell and are not trying to.
- Your procurement process requires certification. If the review will not clear a vendor without SOC 2 or ISO 27001, we fail that review today and saying otherwise would waste your time.
- You want the benefit of what other customers have already seen. Correlating a forged template across many organisations is structurally unavailable to an API that never pools customers: our serial alert sees only your own account's records. That is a real capability and it is a real cost of our design.
- You need pixel-level and generated-image analysis. They describe detectors for pixel-structure anomalies and for images from named generative models (read 2026-10-12). Tamperlens has neither: it reads the structures an encoder wrote, decodes no pixels, and ships no "is this AI?" model. Whether their detectors hold up is for you to test; we cannot say.
Where Tamperlens is the better answer
- You want to try it before talking to anyone. Fifty documents a month, free, with an API key.
- You need to know the price. Ours is on a page. A cost you cannot see until the third call is a cost you cannot budget.
-
The document must not be retained. By default the file never
outlives the response, and a key set to
nonekeeps nothing at all, not even the short derived record an account keeps by default. There is no store of your customers' bank statements to reason about later unless you turn one on. - You want the same answer twice. The engine is deterministic: the same bytes produce the same signals, and a signal's explanation names the object it came from, so a decision can be audited rather than trusted.
Two live numbers that disagree, and why this page carries dates
Both quotations read 2026-08-06. The drawing shows two published totals and their dates. It does not show which is right, and we do not claim either is an error. The most likely explanation is a redefinition, not a mistake.
The newer, more carefully sourced document cites the smaller number. The most likely explanation is a redefinition rather than an error: the report scopes itself to a narrower population than the marketing page counts. We raise it for one reason only. It is the clearest illustration of why a competitor fact without a read date is worth very little, and why the ones on this page have them.
For the same reason we quote no figure from their home page: two loads of it returned different statistics, so there is no stable number there to quote.
What we cannot do
This is the limit of structural analysis, ours included. How often it bites on real fraud is not measured by us, and we found no published figure from others. Catching the bottom row needs a reference for the issuer's genuine output, or verification at the source.
- Recall is measured only on forgeries we built. On /evidence: 1,006 forged files made from 160 genuine documents by eight known edits, applied by a script. Detection falls as the forger does more to the container, from 100% for an incremental-update edit to 65.5% after a Ghostscript re-distill, and a separate null result found 0 of 20 freshly generated documents above the low band. Not measured: a true-positive rate against real fraud, or against a forger who has read that page.
- A document forged from scratch and never edited may carry no structural trace at all. Structure evidence is evidence of editing, not of truth.
- A cleanly regenerated PDF launders most byte-level signals. The one family that survives whole-file laundering is redaction exposure.
- No identity proofing, and no verification of a person.
What none of us publishes
As far as we could find, no vendor in the table above publishes a false-positive rate on legitimate documents together with its population and date: how often the system flags a genuine file. It is the number that decides how much manual review a detector costs you, and it is the one number this field does not print.
Ours is at /evidence, with the population it was measured on, the date, and the engine version that produced it. The other comparisons are at /alternatives.
Read a report before you believe a comparison
Open the “Edited after creation” sample: a precomputed report on a document whose printed total was lowered by a second revision appended after creation. No account, no upload, no quota, and nothing on this page has to be taken on trust to look at it.
Tamperlens reports risk signals, not authenticity verdicts. Signals can have benign causes; combine them with your own decision logic.