How accurate is the JULL AI authenticator?
A free AI authenticator is only as valuable as the truth behind its accuracy claim. We built our tool on a proprietary AI vision engine and a deterministic rule engine; here is exactly how well it performs, where it fails, and what we will not pretend to do.
The honest answer
On low-tier counterfeits (AliExpress / DHgate / Turkish replicas), the tool catches roughly 9 out of 10. On mid-tier “factory-grade” fakes — the ones that fool casual buyers but not a hand inspector — it catches around 6 to 7 of 10. On top-tier “super fakes” engineered specifically to defeat photo authentication, it catches at best 2 to 3 of 10.
This is true of every photo-based authentication service, including paid ones. Anyone claiming higher accuracy on super fakes from photos alone is selling certainty they do not have.
How the tool works
Our tool does notask our AI “is this bag real?” That is a question no general-purpose AI can answer reliably. Instead, the tool asks our AI four narrow, objective questions and lets our deterministic rule engine make the judgment.
For each of the four uploaded photos, our AI returns a strict JSON response measuring what is visible: a letter and shape on the blind stamp, a stitch count per centimetre and an angle estimate, a count of feet with micro-stamps, the smoothness and color of the edge paint. These measurements are evaluated against authentic Hermès production patterns by 25+ rules, each weighted by how strong a counterfeit signal it represents.
Our AI is the eye. Our rule engine is the judge. The verdict you see — a score from 0 to 10 with marker breakdown — is produced by code we wrote, not by inference of any single model of authenticity.
What we tested against
We assembled a calibration set from documented reference photos in established authentication archives: LegitCheck's real-versus-fake comparison libraries, Bagaholic guides, Hauteacademics breakdowns, and Madison Avenue Couture's authentication blog. Authentic reference bags came from JULL's own boutique-sourced stock with documented provenance.
We tested across three counterfeit tiers because they are fundamentally different problems:
- Low-tier — AliExpress, DHgate, eBay copies sold openly as “designer inspired”. Wrong leather grain, wrong stitch count, missing feet stamps, sloppy edge paint.
- Mid-tier — Turkish, Vietnamese, and lower-grade Chinese factory replicas sold as “1:1” on Reddit and Telegram. Better stitching, sometimes correct leather, often correct hardware weight, but typically still incorrect feet detail or edge paint.
- Top-tier — high-grade Chinese and Korean “super fakes” sold for £1,500–£5,000 each. Real Hermès leather offcuts, replicated stamping equipment, near-perfect saddle stitching. Engineered specifically to defeat photo authentication.
Calibration results
| Category | Correctly flagged | Detection rate |
|---|---|---|
| Authentic Hermès bags (true positive) | passed pattern check | 92% |
| Low-tier counterfeits | flagged as suspicious | 89% |
| Mid-tier counterfeits | flagged as suspicious | 66% |
| Top-tier “super fakes” | flagged as suspicious | 28% |
Results from our internal calibration run, May 2026. Numbers will shift as Hermès updates production patterns and as counterfeit factories adapt. We re-calibrate quarterly.
What this means in practice
If you are considering a bag listed on Vestiaire, eBay, Instagram, or Telegram at a suspiciously low price, our tool is very likely to flag it. The fakes at those price points are almost exclusively low and mid-tier. A score under 5 from our tool is a strong signal not to proceed.
If you are considering a bag from a serious pre-owned dealer charging market rate, the calculus is different. Top-tier super fakes typically appear at market prices through dealers who themselves were deceived during sourcing. Our tool may report a high score on a top-tier fake. A high score is not a guarantee of authenticity. It means the pattern markers we can read from photos look consistent. Final authentication still requires hand inspection.
This is why we are unambiguous in our disclaimer language. The tool catches obvious and moderate counterfeits very well. It catches super fakes inconsistently. Pretending otherwise would be dishonest.
Where the tool fails
In our calibration we documented every failure case. The recurring issues:
- Blurry photos pass borderline. When photo quality is too low for our AI to confidently measure, the tool returns “insufficient” — but borderline photos can produce false positives. Solution: we require all 4 photos at sharp focus.
- Top-tier super fakes with real Hermès leather. If a fake is built from genuine Hermès leather offcuts, the visual texture is correct. The remaining tells are stitch tension and hardware finish, which photos do not always capture.
- Year code letters with unusual spacing. Hermès has used several stamping fonts. our AI occasionally misreads when the heat-stamp depth varies across the letter.
- Vintage bags with discontinued patterns. Our rule engine encodes 2015-onward patterns most thoroughly. Pre-2015 vintage bags can score low not because they are fake but because we do not yet model older production patterns.
We are documenting these failure cases publicly because that is the only way an authentication tool earns trust. Every authenticator — paid or free — has failure modes. The honest ones tell you what they are.
Compared to paid services
For context, here is how our free tool compares to commercial authentication services on the same dimensions:
| Service | Method | Cost | Claimed accuracy |
|---|---|---|---|
| JULL AI Authenticator | Photo-based AI + rules | Free | 89% on low-tier, 66% on mid-tier, 28% on super fakes |
| Entrupy | Microscopy device + AI | $119/bag | 99.86% (their published number) |
| Bababebi | Human expert + photos | ~$200/bag | Stated as “opinion”, no published rate |
| Real Authentication | Human expert + photos | $20–30/bag | Stated as “opinion”, ~92–95% from industry estimates |
Entrupy is the gold standard for pattern detection but requires their proprietary microscopy hardware — you cannot use them from photos alone. Bababebi is the most respected solo authenticator but is currently closed to new clients. Real Authentication accepts photo submissions and is reasonably priced.
Our recommendation
Use our free tool as a first-pass filter. If it flags a bag, do not proceed. If it gives a high score, you have a useful data point — but use it as one input among several:
- Ask the seller for the original boutique receipt.
- Inspect the bag in person.
- For bags above £15,000, consider paid third-party authentication.
- For maximum certainty, buy from a dealer with traceable boutique sourcing.
JULL exists for the last option. Every bag we sell comes direct from the Hermès boutique chain — the chain of custody never breaks, so the authentication question never arises. Read more about our sourcing.
Try the tool
Free. No signup. Photos analysed by JULL AI in 2–5 seconds. Verdict produced by JULL's rule engine against 25+ authentic Hermès markers.
Open the AI authenticator →