Two earlier articles in this series have already touched circular fashion from adjacent angles. The dead stock article covered the AI-matched liquidation channels and sorting platforms — Renewlane, Circ — that route unsold inventory toward recycling rather than landfill. The computer vision quality control article examined how machine vision assesses returned garments for repair versus recycling routing at brands like Patagonia. Both pieces, by necessity, treated circularity as a downstream sorting and routing problem.

This article goes underneath that layer, to the more fundamental technical and regulatory problem that determines whether any of that sorting and routing infrastructure can actually work at scale: fibre identification. You cannot recycle a garment fibre-to-fibre — meaning genuinely turning old fabric into new fibre, rather than simply downcycling it into insulation or wiping cloths — unless you know with precision what it is made of. And for the overwhelming majority of the world’s textile waste today, nobody actually knows.

A significant EU regulatory development now underway is forcing this problem into the open in a way that will reshape circular fashion infrastructure across the industry, well beyond companies that sell directly into the EU. This article examines the fibre identification bottleneck, the AI technology now solving it, and the Digital Product Passport regulation that is about to make solving it mandatory.

Less than 1% of textile material globally is currently recycled fibre-to-fibre — turned back into new yarn rather than downcycled into a lower-value application like industrial wiping cloths or insulation, or simply landfilled and incinerated. The reasons for this abysmal figure are often discussed in terms of economics and chemistry: fibre-to-fibre recycling processes are expensive, and blended fabrics — a cotton-polyester mix, for instance — are genuinely difficult to separate into pure component fibres that can be respun.

But underneath both of those well-known problems sits a more basic one that receives far less attention: a sorting facility processing a mountain of donated and discarded clothing frequently does not know, with any reliable precision, what most of those garments are actually made of. Care labels are missing, illegible, or — a persistent and underappreciated problem — simply wrong, either through manufacturing error or because the label was never updated when a fabric blend changed during production. A facility cannot make a sound fibre-to-fibre recycling decision based on information that is this unreliable, which is precisely why the overwhelming majority of collected textiles default to the lowest-value disposal path: it is the only path that does not require knowing exactly what the material is.

This is the genuine bottleneck that AI-powered fibre identification technology is built to solve — not the chemistry of separating blended fibres, which remains a real and only partially solved materials science challenge, but the much more basic problem of knowing, reliably and at the speed a sorting facility actually operates, what is in front of the sensor.

The dominant AI-powered approach to fibre identification uses near-infrared (NIR) spectroscopy, often combined with hyperspectral imaging — a technique that captures how a material absorbs and reflects light across a much wider range of wavelengths than the human eye or a standard camera can detect. Different fibre types — cotton, polyester, wool, nylon, and their countless blends — each produce a distinct optical “fingerprint” in this near-infrared range, determined by their underlying molecular structure, regardless of dye colour, surface print, or physical condition.

The genuine AI contribution sits in interpreting these optical fingerprints, because raw spectral data alone does not resolve cleanly into a confident fibre identification — particularly for blended fabrics, where the spectral signal is a combination of multiple fibre fingerprints overlapping in ways that are not always easy to separate mathematically. Convolutional neural network models, trained on large libraries of reference spectra from known fabric samples, learn to distinguish these overlapping signatures with a level of pattern recognition that traditional statistical methods could not reliably achieve, particularly as the number of distinct possible blends and fibre combinations grows.

This data dependency creates a genuine industry coordination problem worth naming directly: these models are only as good as the reference spectral libraries they are trained against, and building those libraries at the scale and diversity required — covering not just pure fibres but the vast range of real-world blends, conditions, and degradation states found in actual post-consumer textile waste — has historically been fragmented across individual companies and research institutions rather than shared as common infrastructure. NIST, the U.S. National Institute of Standards and Technology, released an open-access reference database in 2025 specifically to address this gap — a collection of near-infrared spectral fingerprints across 64 distinct fabric types, including pure fibres, common blends, and genuine real-world fabric samples sourced from thrift stores rather than only laboratory-pure references, intended as shared training infrastructure that any sorting technology developer can use to improve their models rather than each company rebuilding the same reference data independently.

ZORITEX has built an AI-powered sorting system using hyperspectral near-infrared technology capable of identifying more than 15 distinct fibre types and their blends at sorting-line speed — the company reports the approach can increase sorting throughput roughly tenfold compared to manual sorting while reducing per-unit sorting costs by 50 to 75%, a combination that begins to make fibre-to-fibre recycling economically viable at a scale that manual identification never could.

Research groups including VTT Technical Research Centre of Finland have published open datasets of hyperspectral textile imaging specifically to support broader research and commercial development in this space, reflecting a pattern across the field where the underlying sensor hardware is increasingly commoditised while the genuinely differentiating asset is the quality and breadth of the training data and classification model built on top of it. Once a fibre is identified with confidence, separation technologies like Aalto University’s Ioncell process — using ionic liquid solvents rather than harsh chemicals to extract cellulose fibres from blended fabric — can achieve recycling rates above 95% for the cellulose component, demonstrating that the downstream chemistry, while still developing, is considerably further along than the upstream identification problem that AI sorting is now addressing.

This sequencing matters for how the industry should think about where investment is most urgently needed: the recycling chemistry to turn an identified, separated fibre into usable new material is advancing steadily and, in several specific cases, already works well at meaningful scale. The genuine bottleneck constraining the entire system is upstream of that — accurately and cheaply identifying what is in the waste stream in the first place, at the speed and volume a real sorting facility requires.

The European Union’s Ecodesign for Sustainable Products Regulation (ESPR), in force since July 2024, introduces a requirement that will reshape this entire landscape: the Digital Product Passport, a standardised digital record — accessed via a QR code, RFID tag, or similar data carrier attached to the physical garment — containing verified information about a product’s materials, manufacturing, durability, repairability, and end-of-life recycling instructions. Textiles have been selected as one of the EU’s first priority categories for this requirement, specifically because of their high potential for improved material efficiency and extended product lifetimes.

The detailed delegated act setting the specific ecodesign requirements for textiles is expected in 2027, with a transition period meaning practical implementation is anticipated from 2028 — but the regulation applies to any apparel and footwear company selling into the EU market regardless of where that company is based, making it a genuinely global compliance question rather than a Europe-only consideration for any brand with meaningful EU sales. The European Commission’s Joint Research Centre has already conducted detailed life-cycle assessment work on representative garment categories — knitted, denim, and other woven products — establishing that raw material production accounts for 60 to 63% of a typical garment’s total environmental impact, the largest single contributor by far, which is part of why accurate material composition data sits at the centre of what the passport is designed to capture.

The connection to AI-powered fibre identification is direct and consequential: a Digital Product Passport that accurately and verifiably records a garment’s exact fibre composition at the point of manufacture — rather than relying on a care label that may be missing, illegible, or wrong by the time the garment reaches end-of-life — solves the identification problem at its source, before the garment ever reaches a sorting facility. This is a genuinely different and more powerful approach than AI sorting systems trying to identify fibre composition from scratch using only optical sensing at the waste stage, because the passport, if implemented and maintained accurately, removes the need to re-derive information that was already known and verified when the garment was made.

A Digital Product Passport is, by itself, simply a structured data record — it does not generate the information it contains. The genuine AI opportunity sits in how that data gets created, verified, and used across the garment’s life. At the manufacturing stage, AI-powered supply chain traceability systems, building on the supply chain mapping capability examined in this series’ supply chain article, are increasingly the mechanism by which a brand can populate a passport with verified rather than self-reported material composition data — cross-referencing supplier invoices, material certificates, and in some emerging deployments, the same NIR fibre verification technology described above, applied at the factory stage to confirm that a finished garment’s actual fibre composition matches what its passport claims.

At the end-of-life stage, the passport’s data becomes the input that AI sorting systems can use directly, rather than needing to derive fibre composition from scratch through optical sensing alone. A sorting facility scanning a passport-equipped garment’s data carrier receives verified, manufacturer-confirmed composition data instantly — and AI’s role shifts from inference (identifying an unknown fibre from its optical signature) to verification and exception-handling (confirming the passport data is consistent with what the garment actually appears to be, and flagging the cases — degraded labels, counterfeit passports, garments that have been altered or repaired with different materials since manufacture — where the passport and the physical reality have diverged).

This is a meaningfully more efficient division of labour than relying on optical AI identification alone for every single garment passing through a sorting line — verification of an existing data claim is a computationally and economically easier problem than identification from nothing, in roughly the same way that the computer vision quality control systems examined elsewhere in this series found defect verification against a known specification considerably more tractable than open-ended anomaly detection. The optical fibre identification technology described earlier in this article does not become obsolete once passports are widespread — it remains essential for the enormous existing stock of pre-passport garments already in circulation, and as the fallback and verification layer for passport-equipped garments where the data carrier has failed or the passport record is incomplete.

This series’ dead stock article examined Burberry’s 2018 disclosure of destroying £28.6 million in unsold goods, and the public backlash that eventually led several luxury houses to abandon the practice. The ESPR framework now driving the Digital Product Passport is also the legal vehicle introducing a more direct measure on the same issue: a formal prohibition on the destruction of unsold textile and footwear products, with specific and limited derogations — safety concerns, product damage — overseen by national authorities, alongside a separate requirement that businesses disclose, in a standardised format, the actual volumes of unsold goods they discard.

This connects circular fashion regulation directly back to the dead stock economics examined earlier in this series: a brand that can no longer legally default to destruction for unsold inventory has a considerably stronger incentive to invest in the AI-matched liquidation channels, redistribution systems, and now-mandatory disclosure infrastructure this series has covered — not as a sustainability nice-to-have, but as the only remaining legally compliant path for inventory that does not sell. The regulation is, in effect, forcing the commercial logic of circular fashion AI into a position where it is no longer optional for any brand serious about EU market access.