Self-Describing Products for Recyclers: The Law Is Written, the Business Case Is Not

The problem

In 2022 the world generated a record 62 million tonnes of electronic waste, and less than a quarter of it (22.3%) was documented as properly collected and recycled¹. The metals embedded in that waste were worth an estimated US$91 billion, yet e-waste recycling meets no more than 1% of demand for rare earth elements¹. The report's authors project that generation will reach 82 million tonnes by 2030 while the documented recycling rate falls to 20% (projected)¹. These are official estimates from the International Telecommunication Union (ITU) and the United Nations Institute for Training and Research (UNITAR), published in March 2024.

Clothing shows the same pattern. Europe generates more than 15 kilograms of textile waste per person each year, only 30–35% is collected, and less than 1% of textile waste is recycled back into new fibre (consultancy estimate, 2022)². The reason is partly informational: most fibre-to-fibre recycling processes have strict input requirements for fibre composition and purity, so waste has to be scanned and sorted to match them². These textile figures are more than three years old.

Behind both numbers sits one practical difficulty. A recycler receiving a mixed load cannot easily tell what each item is made of, which substances it contains or how it comes apart, and finding out item by item costs more than the recovered material is often worth. That economic squeeze is visible now: Plastics Recyclers Europe expects Europe to have lost almost one million tonnes of plastics recycling capacity since 2023 by the end of 2025 (projected, industry-reported), citing low-priced imports of recycled plastic, falling demand for EU-made recyclate and regulatory burden³.

What already works

Products that announce themselves

The closest working example of a product telling a sorting machine what it is comes from packaging. In the HolyGrail 2.0 initiative, brands printed digital watermarks (patterns invisible to the eye that a camera can read, each linking to information about that item) on their packaging. From August to December 2024, a sorting line at a material recovery facility in Swisttal, Germany, logged 5.66 million detections of 5,949 different products, with detection efficiency between 87.9% and 93.8% (operator-reported)⁴. Earlier semi-industrial tests in 2021–2022 reported an average of 99% detection, 95% ejection and 95% purity for the first prototype (operator-reported)⁵. No independent verification of either set of figures was found.

In plain terms: on real household waste, a machine read roughly nine out of ten tagged items. The tagging, not artificial intelligence, did most of the work.

Machines that recognise materials without being told

Where items carry no tag, sorting relies on sensing. The Fibersort machine, developed by a European consortium, uses near-infrared light to classify garments by fibre. A laboratory check of its output by the Belgian textile institute Centexbel in March 2019 found 98.5% cotton in the pure-cotton fraction, but only 82% polyester in the pure-polyester fraction and 61% polyamide in the pure-polyamide fraction, against target thresholds of over 95% and over 80% (project-reported, measured without a control group)⁶. The consortium itself said the technology needed further adjustment for polyamide and viscose⁶. These results are more than three years old.

Machine-vision research on mixed recycling streams shows similar limits. A 2025 preprint (not yet peer reviewed) testing modern detection models on images from an operating recovery facility reached a mean average precision of 51.6, up from a published baseline of 24.2 (measured on a benchmark dataset)⁷. Mean average precision is a 0–100 score combining how many objects a model finds and how accurately it outlines and labels them. The same study notes that cardboard makes up over 66% of the labelled items and metal less than 2%, and describes the stream as cluttered, overlapping and deformed⁷.

What these figures mean for a reader: recognition systems can sort broad categories well, but telling one plastic or fibre blend from another in messy, dirty piles remains unreliable. Recognition alone does not tell a recycler what an item contains.

A legal framework for product data

The European Union has adopted the legal basis for the information side. The Ecodesign for Sustainable Products Regulation, adopted on 13 June 2024, allows the Commission to require a digital product passport: a set of data specific to a product, reachable by scanning a data carrier such as a barcode or two-dimensional code on the product, its packaging or accompanying documents⁸. Recyclers are named among the actors entitled to free and easy access, according to access rights set product group by product group⁸. Where information requirements apply, they must make it possible to track substances of concern, including their location in the product, their concentration and information relevant for disassembly and recycling⁸. The passport must remain available for at least the expected lifetime of the product⁸.

The regulation leaves the content of each passport to later product-specific acts⁸. The Commission's 2025–2030 working plan, adopted in April 2025, prioritises steel and aluminium, textiles (with a focus on apparel), furniture, tyres, mattresses and several energy-related products, plus horizontal requirements on the recyclability of electrical and electronic equipment⁹. The separate EU Batteries Regulation of July 2023 already provides for a battery passport, and states that detailed composition and dismantling information should not be public but should be available to recyclers, repairers and other operators who need it¹⁰.

What does not exist

The research found no sorting operation that reads item-level passports from mixed waste at industrial speed and uses composition data to route each item. No ESPR passport for a consumer product was confirmed to be in force at the time of writing. Products reaching recyclers today, and for years to come, were placed on the market without passports.

The possibility

One way this could work: a garment, appliance or piece of furniture carries a durable data carrier linked to its passport. When it arrives at a sorting plant, a reader on the line looks up the item and retrieves its material composition, substances of concern and disassembly notes. The line then routes it to the right recycling process, or to a dismantling station with instructions, instead of to a mixed bale or an incinerator. Where the carrier is missing, damaged or unreadable, sensors and recognition models estimate what the item is, and the plant treats that estimate with less confidence.

The role of AI in this picture is narrower than the headline suggests. Reading a code and looking up a record is ordinary software and data plumbing. Setting access rights, keeping records accurate for a product's whole life and paying for the system are matters of law, standards and money. AI would carry three specific jobs. First, recognition: identifying items whose tags cannot be read, which would be most of the stream for a long time. Second, matching: linking a partially read or crushed item to the most likely passport record. Third, decision support: combining passport data, sensor readings and current market prices for recovered material to decide, item by item, whether dismantling is worth the labour.

If passports become widespread, recognition would shift from guessing composition to confirming identity, which is an easier task. If they do not, AI would remain the main tool, with the limits described above.

What would have to be true

1. Passports would have to cover the products that matter. The working plan's priorities include textiles and furniture, but electronics appear mainly through horizontal recyclability requirements⁹. Composition data would need to be required, not optional, for the product groups recyclers handle in volume.
2. Recyclers would need data they can use. The European waste management association FEAD warns that complexity limited the usefulness of the existing chemical-substances database (SCIP) for recyclers, and asks for access to concentrations of regulated chemicals and materials¹¹.
3. The carrier would have to survive the bin. FEAD asks that passports be accessible even for damaged, crushed or compacted products, and notes that radio tags suffer metal interference in batteries and durability problems in textiles¹¹.
4. Data would have to be accurate for decades. The regulation requires accurate, up-to-date data available for the expected lifetime of the product⁸. For durable goods, that means records maintained long after sale, possibly after the manufacturer has gone.
5. Recognition models would have to handle real waste. Current benchmark performance in cluttered streams is around half of the maximum score⁷. Models would need to be trained on the specific streams each plant handles.
6. Someone would have to pay, and someone would have to buy. Better sorting adds value only if there is a buyer for the recovered material at a price above the cost of recovery. The closures reported by Plastics Recyclers Europe suggest that condition does not hold today for parts of the plastics market³.
7. Access rules would have to protect commercial secrets without starving recyclers. The battery rules restrict composition detail to operators with a need¹⁰; similar balances would be needed for every product group.

What could stop it

What the evidence shows

Europe's plastics recycling industry has been shrinking: almost one million tonnes of capacity is expected to have been lost since 2023 by the end of 2025, with low-priced imports and weak demand for EU recyclate named as causes (projected, industry-reported)³. Waste operators report that an earlier EU product-data tool, the SCIP database, was hampered by complexity and limited usability for recyclers¹¹. Automated fibre sorting fell short of its own purity thresholds for several fibre types in 2019 testing⁶, and in 2022 automated fibre sorting accurate enough for recycling was described as not yet developed². The strongest field result for item-level tagging, the HolyGrail trial, is reported by the initiative itself⁴.

Reasoning

The challenge question is whether this possibility would be less likely, less useful or more harmful than it first appears. Several points suggest caution.

Information may not be the binding constraint. If recovered material is cheaper to import than to produce in Europe, knowing exactly what an item contains does not make recovering it profitable. Better data would raise purity and reduce sorting cost, but would not by itself create demand.

The legacy stream is long. A sofa or washing machine sold before passports exist would still reach recyclers years later. For that period, which for durable goods could be long, sorting would depend on recognition, not passports.

The cost falls on some and the benefit on others. Manufacturers would bear most of the cost of creating and maintaining passports; recyclers would capture most of the benefit. Unless rules or fees connect the two, incentives would work against data quality.

Simpler measures may do more. Designing products from fewer materials, banning hard-to-separate combinations or paying recyclers through producer-responsibility fees could each improve recovery without any new data system. A passport that records a hard-to-recycle blend does not make it easier to recycle.

There are risks to people. If disassembly instructions or substance data were wrong, dismantling staff could be exposed to hazards they were told were absent. Liability for errors in passport data is unresolved; FEAD explicitly asks that recyclers not be made responsible for substances the passport fails to declare¹¹.

Signs to watch

• The first ESPR product-specific act setting passport content, for example for textiles or another working plan priority, and whether it requires material composition⁹.
• Whether the Commission's passport registry, due by 19 July 2026 under the regulation, is operating and what it holds⁸.
• The rules on access rights for recyclers in each product act, compared with the battery model¹⁰.
• Results from the HolyGrail 2030 early-adopter demonstrations in Belgium and Germany, especially any independently verified purity and cost figures⁴.
• The Commission's horizontal recyclability requirements for electrical and electronic equipment⁹.
• Updated capacity figures from Plastics Recyclers Europe for 2026³.
• Peer-reviewed results from recognition systems tested in operating plants, not only benchmark datasets⁷.