On this page
Digital watermarks: do they work at a real sorting facility?
Marks and reference graphics are shown for identification, academic research and commentary. The image does not grant permission to apply a scheme mark to packaging; consult the issuing body's current eligibility and artwork rules.
Introduction
The other marks in this section address a human reader. Digital watermarking addresses a machine. Imperceptible patterns are embedded in print or moulded texture. Industrial cameras recover identification data at conveyor speed. The aim is to identify each package at the sorter without relying on the consumer or on a visible label.
The technology and the programme
A digital watermark tiles the package's surface with a repeating, sub-visible code, carried either in artwork through minute luminance modulations or in the substrate through surface topology, so that any sufficiently large fragment presents the identifier to a camera. At the sorting line, detection units image the waste stream and resolve watermarked items to their identities, which permits sorting decisions on attributes no spectroscopic sensor can register, among them food-grade as against non-food-grade instances of the same polymer, brand-level streams, and format distinctions. The implementation at issue in the published trials is Digimarc's, and the sector vehicle is HolyGrail 2.0, the initiative driven by AIM, the European Brands Association, together with the Alliance to End Plastic Waste; it succeeds a first HolyGrail pioneer project and constitutes the coordinated attempt to move watermark sorting from laboratory conditions to industrial validation (Digimarc, 2025; AIM, n.d.).
The programme's staging bears on the interpretation of its figures. Semi-industrial trials, conducted on controlled test streams and purpose-run lines, established technical feasibility with detection figures in the high nineties; those figures circulated widely and established the expectations against which the subsequent test would be read. That subsequent test concerned unstaged municipal waste, which is contaminated, deformed and overlapping, and which passes at commercial throughput.
The industrial validation
The reference evidence was published in April 2025, when Digimarc announced industrial-scale validation from HolyGrail 2.0 trials at the Hündgen Entsorgung materials recovery facility in Swisttal, Germany, a working MRF processing real post-consumer rigid household packaging, with the trial running from August to December 2024 (Digimarc, 2025; Packaging Europe, 2025).
The published results record 5.66 million detections across 5,949 unique SKUs, averaging approximately 56,000 items detected per day, with "detection accuracy … consistently above 90%" under real-world conditions of surface contamination, deformation and overlap (Digimarc, 2025). Independent coverage reports the distribution underlying that headline figure: detection efficiency ranged from 87.9 to 93.8 per cent across the trial, with cloud-based analysis indicating consistent efficiency above 90 per cent (Packaging Europe, 2025). The programme's announced next phase addresses a more difficult substrate, namely flexible polypropylene food packaging, trialled in Belgium (Digimarc, 2025).
The staging distinction requires explicit statement, since it is frequently elided in secondary coverage: the industrial figure is "above 90%", whereas the semi-industrial trials produced figures of approximately 99 per cent. The difference between the two is the measured effect of operating conditions, and the industrial figure is the one applicable to policy and investment assessment. On that figure, the validation supports a specific claim: watermark detection functions at commercial scale, on real waste, at accuracies exceeding 90 per cent, for rigid household packaging, at one facility, across five months.
Scope and limitations of the validation
The trial's boundaries define the extent of what has been demonstrated, and each boundary corresponds to a live research question.
Detection is not sortation. The validated metric is detection accuracy, that is, the proportion of items identified as they pass. A sorting improvement requires the downstream chain: ejection precision, purity gains in output bales, and ultimately yield economics. Detection above 90 per cent is accordingly the enabling condition rather than the outcome, and facility-scale evidence linking watermark sorting to measured bale-purity and revenue effects has not been published. The gap between detection and outcome corresponds to the gap this collection's evidence theme documents for labeling interventions generally.
Rigids are not flexibles. The validated stream is rigid packaging, in which fragments are large and surfaces regular. Flexibles, being crumpled, folded and filmic, present a harder case and constitute the announced next frontier for that reason; no industrial flexible validation existed at this article's verification date.
One facility is not a system. The five months at Hündgen establish feasibility rather than fleet economics. The business case for retrofitting detection units across a national MRF estate turns on the penetration of watermarked packaging, which turns on brand adoption, which turns in turn on the sorting premium that watermarks command. That reciprocal dependency is characteristic of two-sided infrastructure, and it explains why the programme's brand-coalition structure, in the form of AIM's membership, bears on adoption as directly as the technology does.
Identification is not information governance. A watermark resolves to whatever its database records. The integrity questions this theme raises of every claim — which party attests the attributes, and which party audits the register — transfer intact to the invisible layer, with the difference that no consumer, journalist or competitor can read the mark in order to check it. Machine-facing labeling therefore concentrates trust in the resolver.
Governance questions arising from machine-only marking
Machine-only marking raises governance questions that the visible mark families did not face, and those questions bear on the technology's regulatory position as directly as its detection rates do.
The first concerns asymmetry of auditability. Every visible mark examined in this theme is checkable by any party holding the package, and the enforcement ecology of claims law, NGO investigation and competitor complaint operates on that basis. A watermark's assertions are checkable only by parties possessing detection equipment and database access: the marking's content, its currency and even its presence are inaccessible to the publics on whose trust the system depends. The programme's coalition structure, comprising brand-association leadership and published trial results, supplies interim transparency; a mature system would require what visible marks obtained from law, namely defined access rights to what the package silently states, for regulators at minimum.
The second concerns the register as the locus of meaning. A watermark resolves identity, while every attribute — food-grade status, material composition and the other sorting-relevant facts — resides in the database that the identity keys into. Attribute integrity therefore concentrates at a point that certification's machinery of audit, revocation and licensee codes does not presently reach: no scheme currently certifies the truth of a watermark's linked record, and the sorting decisions the system exists to improve are only as sound as unaudited entries permit. The certification architectures this theme documents for visible claims supply an evident template, whose extension to the invisible layer had not been constructed at this article's verification date.
The third concerns standardisation posture. The EU's carrier provisions consistently specify open, standardised digital marking, the recital language governing EPR identification being explicit on the point, and a proprietary detection ecosystem does not sit readily within that specification. Any path from the present position into mandated architecture runs through standardisation, which a commercial incumbent has limited incentive to pursue. The configuration is the familiar one in infrastructure, in which a first mover's advantage consists in precisely what public adoption would require it to relinquish.
None of the three questions constitutes an argument against the technology; each is a question that arises once detection is demonstrated to work.
Position relative to other carriers
The technology's position may be fixed by comparison with its neighbours in this group.
By comparison with the QR code, the watermark trades visibility for coverage. A QR code occupies dedicated artwork space, must survive intact, and addresses consumers and scanners alike, whereas the watermark is distributed across the pack, tolerant of fragmentation, and addressed to machines only. The carriers are accordingly complementary rather than competing — Digital Link identity for the addressable package, watermarks for the conveyor — and dual deployment is the pattern emerging in the trials' brand cohort.
By comparison with spectroscopy, the near-infrared sorting that constitutes the MRF's incumbent sensor, the watermark's contribution is granularity: NIR reads polymer, whereas the watermark reads product, which permits food-grade separation and the closed-loop streams that polymer-level sorting cannot construct. The incumbent's contribution is universality: NIR reads everything, whereas watermarks read only what has been marked. The architecture indicated is therefore layered, with spectroscopy handling the bulk and watermarks the premium streams, from which it follows that watermark economics depend on the value of the streams that watermarks alone can separate rather than on displacement of the installed base.
By comparison with the EU's regulatory carriers, the watermark is absent. The packaging regulation's data-carrier obligations specify QR codes and open standardised marking, and watermarks, being proprietary and machine-only, fall outside the mandated architecture. The technology is accordingly a sorting-side one whose adoption case rests on facility economics rather than on compliance obligation.
Upstream effects of sorting-side identification
If watermark sorting scales, its effects extend backward through the labeling system, redistributing burdens mapped elsewhere in this collection.
The first effect concerns design-for-recycling economics. The DfR guidelines' classifications, examined earlier in this theme, price features according to their behaviour in sorting and reprocessing, and sorting-relevant classifications assume the present sensor stack. Watermarked identification alters what sorting can distinguish: full-sleeve decorated containers that defeat optical sorting, and carbon-black pigments invisible to near-infrared, which are among the guideline red-lists' most persistent entries, present to a watermark reader as packages carrying identities. A scaled watermark layer would therefore not merely add a sorting capability but re-price design features across the assessment systems, relaxing constraints the guidelines currently enforce. That interaction between two of this theme's technical families is not at present modelled by the governance of either.
The second effect concerns claims substantiation. Recyclability claims are disciplined, throughout this collection's subject matter, by access and processing evidence aggregated at category level. Item-identified sorting generates evidence of a different granularity, namely which SKUs, at which facilities, were detected, ejected and baled, which is the facility-scale outcome data whose absence this collection's evidence theme repeatedly records. A brand whose packaging is watermarked at scale could in principle substantiate recyclability by reference to its own products' measured sorting performance rather than by category proxies, which would constitute a substantiation improvement under tightening claims law and would at the same time be available only to firms able to fund the marking, thereby operating as a further mechanism by which compliance capability concentrates.
The third effect is definitional. "Recyclable" has always meant, in operational terms, recognisable by the system, since a package the MRF cannot identify becomes residue whatever its polymer composition. Identification layers redefine recognisability, and with it the boundary of the recyclable, so that a concept this theme treats as infrastructure-relative is at its core sensor-relative. The watermark's implication is accordingly not confined to improved sorting of what is already recyclable but extends to the scope the term can cover, which raises the further question of which party determines what an expanded system is obliged to recognise.
Summary
The validation record consists of a specific claim — industrial-scale detection above 90 per cent on real rigid waste — demonstrated at a named facility, with the semi-industrial and industrial figures documented separately and the further questions of flexibles, sortation outcomes and fleet economics expressly open. The technology's significance for this theme lies in what it decouples. Packaging's end-of-life performance has depended on the visible label's comprehension chain of design, print, reading, understanding and action. Watermarking routes identification around that chain, so that the package communicates its identity to the machine independently of what any human understood. Should the remaining bridges hold — from detection to sortation, from rigids to flexibles, and from one facility to many — the consumer label's sorting function would migrate to infrastructure, leaving the visible label to instruct behaviour rather than to encode identity. The validation record does not establish that outcome; it establishes that the outcome is technically attainable in the conditions tested.
Two constraints follow for the citation of that evidence. The claim requires scaling to the demonstration, in that the validated result concerns detection of rigid packaging at one industrial facility, and each extension — to flexibles, to ejection-and-bale outcomes, and to fleet deployment economics — is a separate empirical question the record does not answer, whatever roadmaps assert. The claim also requires dating, since machine-identification is among the faster-moving subjects in this theme, and the interval between a technology's validated state and its cited state widens as a source ages; citation practice of the kind applied to statutes applies equally to pilot results.
References
AIM (European Brands Association) (n.d.) Digital watermarks — HolyGrail 2.0. Available at: Open source (Accessed: 18 August 2026).
Digimarc (2025) Digimarc Recycle earns industrial-scale validation in HolyGrail 2.0 trials, 14 April. Available at: Open source (Accessed: 18 August 2026).
Packaging Europe (2025) Industrial-scale HolyGrail 2.0 trial sorts 56,000 rigid packaging items daily. Available at: Open source (Accessed: 18 August 2026).
Note on sources and verification
The industrial validation figures — 5.66 million detections, 5,949 SKUs, ~56,000 items daily, accuracy consistently above 90 per cent, the Hündgen facility, and the flexible-PP next phase — are from Digimarc's press release, with the trial window (19 August–19 December 2024) and the 87.9–93.8 per cent efficiency range from the independent Packaging Europe report. The characterisation of earlier semi-industrial results at approximately 99 per cent follows earlier research conducted for this collection on the programme and is presented as context for the staging distinction rather than as a load-bearing figure. Technical description of watermark embedding and detection is background synthesised from the programme's public materials. The observation that watermarks sit outside the EU's mandated data-carrier architecture is an interpretation of the carriage provisions as documented in the compliance analyses cited elsewhere in this group.
Last verified: 18 August 2026.