A used Hermès Birkin can sell for more than a new luxury car. So when one arrives at a resale warehouse, the question is not just "what is it worth" but "is it real." Getting that answer wrong, even once, can destroy a marketplace's reputation overnight.
Luxury resale is now a large and growing business. Platforms like The RealReal, Vestiaire Collective, and Fashionphile take in used handbags, watches, and jewelry, verify them, and resell them.
The bottleneck is trust. A superfake (a counterfeit good enough to fool trained human authenticators) can pass through resale channels and end up on a customer's arm. When that happens, the platform is liable, both legally and reputationally.
Traditionally, authentication relied on human experts. A specialist inspects stitching, hardware, date stamps, and the smell of the leather. This works, but it is slow, hard to scale, and dependent on scarce talent. AI enters here as a way to make expert judgment faster and more consistent.
Computer vision is a branch of AI where models learn to interpret images. For authentication, the model is trained on thousands of photos of genuine and fake items, learning the visual signatures that separate them.
The key insight: authenticity lives in the microscopic details, not the overall look.
Entrupy built a handheld microscope-camera device paired with an AI model. An operator scans specific zones of a bag at high magnification. The system compares those microscopic images against a large database of authentic and counterfeit examples and returns a verdict, typically in under a minute, along with a certificate.
The value is not just speed. It is consistency. The same input produces the same output, unlike a tired human on their hundredth bag of the day.
The RealReal combines in-house human authenticators (gemologists, horologists, brand specialists) with technology tools that flag suspicious items and speed up routine checks. AI here acts as a first filter and a second opinion, not a full replacement for expert eyes on high-value pieces.
This hybrid model is common across the industry: machines handle volume and pattern recognition, humans handle judgment calls and edge cases.
Under the hood, authentication is a binary classification task: given an image, output "authentic" or "counterfeit," usually with a confidence score.
Input: high-magnification image of stitching region
Model: convolutional neural network (CNN)
Output: {"authentic": 0.94, "counterfeit": 0.06}
Decision rule:
if authentic_score > 0.90 -> pass
if authentic_score < 0.60 -> reject
else -> escalate to human expertThat middle band, the "escalate" zone, is the whole point. A well-designed system does not force a yes or no on ambiguous cases. It routes them to a person. The economics work because the model confidently clears the easy majority, freeing experts to focus on the hard few.
A convolutional neural network (CNN) is a model architecture that is especially good at images. It scans for local patterns (an edge, a stitch angle, a grain texture) and builds up to a whole-item judgment.
🎬 [VIDEO: "How AI Authenticates Luxury Goods" — youtube.com — a walkthrough of computer-vision authentication in the resale market]
AI authentication is powerful but not magic. Understanding the failure modes matters more than believing the marketing.
Counterfeiters improve. As authentic reference data leaks or gets reverse-engineered, fakes get closer to real. The best superfakes now use genuine leather, correct hardware suppliers, and near-perfect stitching. A model trained on last year's fakes may not recognize this year's.
This is a moving target. The model must be retrained continuously as new counterfeit techniques appear. A system that is not updated degrades silently.
The model only knows what it has seen.
Genuine goods vary. Handmade items, factory changes over decades, honest wear, and repairs can all make an authentic bag look atypical. A model that is too strict generates false rejections, flagging real items as fake. This frustrates sellers and creates disputes.
The opposite error, a false acceptance (passing a fake), is worse for the brand's trust. Tuning the system is a constant balance between these two failure types.
Microscopic vision captures surface detail. It does not smell the leather, feel the weight, or open the lining to check internal construction. Some authenticity cues remain physical and human. This is why high-value items still get human review.
As with any AI, there is a theoretical risk of adversarial attacks, deliberately crafting a fake to exploit blind spots in the model. In practice this requires knowing how the model works, but it is a reason platforms keep their exact methods confidential.
Vérification des acquis
1. Why does computer vision authentication focus on microscopic details like stitch pitch and leather grain rather than the overall appearance of a bag?
2. The excerpt frames the core bottleneck of luxury resale as 'trust.' What does this reveal about why AI authentication matters to these platforms?
3. According to the lesson's reasoning, what limitation of traditional human authentication makes AI appealing as a complement?
4. Select ALL correct answers about why leather grain is a useful authentication signal.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers describing how a computer vision model learns to authenticate items.
Sélectionnez toutes les réponses correctes.
Why do platforms invest here? Three concrete reasons.
1. Scale. Resale volume is growing faster than the supply of trained human experts. AI lets a platform handle more items without a proportional increase in specialist headcount.
2. Trust as a product. For a resale marketplace, verified authenticity is the core value propositionvalue propositionA clear statement of the benefits your product delivers, the problems it solves and why customers should choose you over alternatives.Voir la définition complète →. A shopper pays a premium to buy from The RealReal instead of an anonymous seller precisely because of the authentication guarantee. AI makes that guarantee cheaper to deliver and easier to document with a certificate.
3. Data compounding. Every item scanned adds to the training database. The more a platform authenticates, the better its models get, which is a genuine competitive moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.Voir la définition complète →. A newcomer cannot easily replicate years of accumulated authentic and counterfeit imagery.
Counterfeiting is a very large global problem. The OECD has published extensive research estimating that trade in counterfeit and pirated goods runs into the hundreds of billions of dollars annually, with fashion and leather goods among the most affected categories. Precise figures are estimates and vary by methodology, but the scale explains why brands and platforms invest heavily in detection.
The direction of travel is clear: not full automation, but augmentation.
The realistic 2026 workflow looks like this:
1. Item arrives and gets photographed under controlled conditions.
2. AI runs a first-pass classification and flags risk zones.
3. Confident authentic items clear automatically.
4. Ambiguous or high-value items escalate to human experts, who see the AI's findings as a starting point.
5. Every outcome feeds back into the training data.
The human expert does not disappear. Their role shifts from checking every item to resolving the hard cases the machine flags, which is a better use of scarce expertise.
For luxury brands themselves, there is a strategic question: should authentication technology stay in the resale market, or should brands build it into products from the start? Some are experimenting with embedded digital identifiers (NFC chips or QR-linked tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.Voir la définition complète →) so a bag can prove its own provenance. That approach sidesteps the vision problem entirely, but it only helps for items made with it, leaving decades of existing luxury goods to the computer-vision approach.