AI Promises Efficiency but the Fashion Industry Needs a Political Reckoning

The operational logic of the Fashion for Good vision rests on a chain of equivalences: AI makes processes more efficient, more efficient processes waste less, less waste means a smaller environmental footprint. Each link in that chain is individually plausible. The chain itself is broken. The mechanism that breaks it has a name, the Jevons Paradox, and it has been documented across every major efficiency-driven industrial transition since the nineteenth century. When the cost of a resource-intensive process falls, the rational response for a growth-oriented industry is to produce more, not to produce the same amount more cleanly.

The dyehouse that achieves first-time-right colour matching does not idle its capacity. It runs more recipes. The demand-forecasting system that reduces deadstock does not constrain buying. It de-risks larger orders. Fashion for Good treats efficiency as a consumption constraint. A market structured around growth treats it as permission to produce more.

The 2026 data does not support that treatment. AI reduces dyeing time by 17.5% and energy consumption by 12.1% in individual facilities, figures cited correctly as evidence of operational progress. What goes unexamined is the trajectory of total garment production, which is moving toward 100 billion annual units, accelerated in part by the same AI trend-forecasting tools the piece frames as waste-reduction instruments. Those tools are marketed to brands as mechanisms for spotting and monetising micro-trends rapidly, compressing the design-to-shelf cycle and enabling more frequent product drops. The same capability that makes production leaner at the facility level makes overproduction cheaper at the brand level. The net effect on absolute consumption is acceleration, not reduction.

The ecological shadow of the AI infrastructure itself compounds this. Fashion for Good describes AI as the connective tissue of the 2036 industry but does not account for what that connective tissue costs to run. Training a single generative model emits carbon equivalent to five cars' lifetimes. Inference requirements for real-time industrial monitoring are projected to consume between 200 and 400 TWh of electricity annually by 2030. Data centres now consume 22% of Ireland's national electricity supply. Google's total carbon footprint has risen 48% since 2019 despite significant per-query efficiency improvements, because the expansion of AI infrastructure has outpaced the gains within it. Water withdrawals for server cooling are projected at between 4.2 and 6.6 billion litres globally by 2030, a figure that does not appear in any lifecycle calculation the piece describes.

This is the contradiction the piece's own framing invites but does not follow through. Facility-level efficiency gains from AI are real and measurable. At the scale of a global industry running on AI infrastructure, those gains are absorbed by production growth, by the resource demands of the infrastructure itself, by the market logic that captures efficiency as competitive advantage rather than environmental dividend. The 2036 vision of a cleaner industry requires not just better tools but a different commercial logic: binding production caps, carbon pricing on virgin materials, mandatory recycled content thresholds. Fashion for Good does not examine whether that logic is coming. The answer, from 2026, is that it is not, at least not at the speed or scale the Jevons mechanism demands.

If that commercial logic were to arrive, through binding international production agreements or carbon-adjusted trade regimes, the efficiency applications Fashion for Good maps would function broadly as the piece describes. The technology is not the obstacle. The question is whether the political conditions that would make it work are on any near-term trajectory, and the regulatory picture examined in the final section suggests they are not.

Fashion for Good's treatment of labour displacement occupies approximately two sentences. The acknowledgement is there: AI-driven automation may displace millions of garment workers across the Global South, and a just transition will be needed. What follows is a pivot back to the technology roadmap. The proportion is not incidental. It reflects an editorial choice about whose perspective the 2036 vision is written from. Brands, investors, and solution providers occupy the foreground; workers and producer-country governments occupy a late-stage caveat. That choice has analytical consequences, because the displacement the piece frames as a future risk is already a present-tense condition, and the distance between the speed of automation and the readiness of any transition policy is growing.

The Bangladesh data is unambiguous. Automation in the country's RMG sector has already produced a 30.58% reduction in the overall workforce, with sweater manufacturing recording a 37.03% decline in labour per production line. These are not projections. They are documented conditions from 2025. Bangladesh's garment exports represent up to 85% of national income, and the sector has historically functioned as the country's primary labour absorber, particularly for low-skilled women workers entering formal employment for the first time. There is no comparable alternative export engine emerging at the scale required to absorb a workforce being displaced at this rate. The just transition Fashion for Good invokes as a necessary counterweight to automation does not have a concrete institutional form, a financing mechanism, or a political constituency in the countries where it is most urgently needed.

The displacement is not uniform, and its structure matters. A fully automated sewing cell costs between $100,000 and $350,000, viable primarily for large-scale operations producing high-volume basics or high-margin goods. This means automation hits the factories employing the most workers first: the high-volume, low-margin facilities that function as the primary point of entry into formal employment across South and South-East Asia. The workers displaced are disproportionately women, older workers, and those with the least transferable skills—precisely the population for whom alternative employment is hardest to secure and for whom transition support would need to be most intensive.

The ILO adopted its first-ever conclusions on AI in manufacturing work in 2026, a signal of institutional recognition that the problem has arrived. No binding framework for wage protection, industrial transition policy, or worker co-determination is yet operational at anything approaching the scale the displacement requires. Meanwhile, the agentic supply chain Fashion for Good describes also depends on a layer of human labour it does not name: the data labellers, content moderators, and AI trainers whose work keeps autonomous systems functional, and whose conditions—low wages, documented mental health harms, geographic concentration in the same regions where garment production is being automated—replicate the vulnerabilities of the workforce the technology is displacing.

The analytical error is not optimism about technology. It is the assumption that the social consequences of deployment move on the same timeline as the deployment itself: that a just transition can be designed and implemented in the years between now and 2036 while automation proceeds at the pace the efficiency arguments demand. The 2026 evidence suggests the opposite. The displacement is faster than the policy, the automation is cheaper than the transition, and the regions absorbing the social costs have the least leverage over the commercial decisions driving them. The horizon problem is not only about the technology moving too fast to forecast. It is about the consequences arriving faster than the institutions built to manage them.

Were producer-country governments integrated into the governance of AI deployment in their industries, through binding social protection clauses in trade agreements or through co-determination rights over automation timelines, the displacement dynamic would look materially different. That integration is not currently on the table. Its absence is not a technical failure. It is a choice about whose interests the transition is designed to serve.

Fashion for Good raises the accountability question directly: when an autonomous supply chain makes a harmful sourcing decision, who is responsible? It then declines to answer, deferring resolution to an entirely new governance category that the industry must develop before 2036. The deferral is honest but significant in what it concedes. Agentic systems are not waiting for that category to exist. Early versions are already being tested in procurement platforms, managing spending limits and supplier substitutions without human sign-off. The decisions are being made. The architecture to hold them accountable is not there. The legal architecture that would need to underpin it is a fundamental mismatch between how liability has always been assigned and how autonomous multi-step reasoning actually works.

The mismatch runs through every relevant body of existing law. Tort law requires proof of duty of care and foreseeability, both of which assume a human actor whose decisions can be reconstructed and evaluated. Contract law requires legal personhood, which autonomous systems do not possess. Agency law depends on explicit or implicit delegation from a human principal, which dissolves when a system is executing multi-step reasoning across chains of tools and permissions that no individual authorised in full.

As of April 2026, financial and trade regulators in the US and EU have formally found that existing law mostly fails to address transactions initiated by autonomous agents. The EU AI Act focuses on compliance requirements and mandatory registration for high-risk AI systems. It does not resolve what happens when a registered, compliant system causes harm. Compliance and accountability are not the same thing, and the Act does not bridge them.

The black-box problem is what makes the jurisdictional failure irreparable in the near term. Even where liability could in principle be assigned, to the brand that deployed the system, the software provider that built it, or the supplier that integrated it, the chain of reasoning that produced a harmful decision is often not reconstructable. Deep learning systems do not generate auditable decision logs in the way that human procurement processes do. A sourcing decision that results in a labour rights violation, an environmental breach, or a contractual failure may be traceable to a model weight, not a choice.

Recent scholarship proposes Law-Following AI systems designed to refuse illegal actions, but the legal boundaries such a system would need to respect are themselves unresolved across the cross-border supply chains Fashion for Good describes, where a single logistics decision may simultaneously engage Bangladeshi labour law, EU environmental regulation, US trade compliance, and Vietnamese export controls.

Were a recognised accountability standard to emerge for agentic systems, one that required decisions to be reconstructable and liability to be assignable before deployment proceeds, the black-box problem would become partially tractable. Jurisdictional agreement would still be required, but the evidentiary basis for redress would exist. No such standard is currently in force. Its absence is not a gap in the regulatory calendar. It is the condition under which deployment is already occurring.

The returns from agentic deployment accrue to brands immediately, while the legal and social costs of harmful decisions fall on workers, suppliers, and states with no mechanism for redress. Fashion's supply chains are among the most geographically dispersed, the most legally complex, and the most dependent on the lowest-margin actors absorbing risk, which makes unresolved accountability particularly acute and concentrated among those least able to absorb it. Fashion for Good frames the governance question as something the industry must answer well before 2036. The industry is already testing agentic systems in conditions of legal ambiguity, and the harms accumulating in that gap compound with each deployment cycle.

Regulation is the most concrete of the prerequisites the Fashion for Good vision requires, and it is already in motion. The EU's legislative programme for fashion and textiles is the most serious attempt by any major jurisdiction to build the compliance floor the 2036 vision presupposes. As of July 2026, large enterprises operating in the EU are prohibited from destroying unsold clothing, a hard deadline with real enforcement consequences. The Ecodesign for Sustainable Products Regulation is in force. The ECGT Directive, effective September 2026, converts unsubstantiated environmental claims from reputational risk to legal liability, with fines reaching 4% of annual turnover for fraudulent data.

Extended Producer Responsibility for textiles is expected around April 2028, making producers financially responsible for end-of-life collection for the first time. By 2027, mandatory Digital Product Passport compliance for textiles will require granular fibre-level breakdowns of material composition, origin, and repair instructions from every brand selling into the European market. This is not marginal regulatory activity. It is a structural reordering of what market access requires.

The DPP deserves particular attention because it is the instrument closest to the traceability and circularity logic Fashion for Good describes. Every product will carry a unique identifier linking to a verifiable record of material composition, origin, and end-of-life routing, precisely the data infrastructure that AI-powered circular systems depend on to sort, redirect, and recycle at scale. The regulatory timeline and the technological roadmap are genuinely aligned here: the DPP creates the data layer that makes AI-enabled circularity operationally possible, and the EU's enforcement mechanisms create the incentive that will drive adoption beyond early movers. The Fashion for Good vision is regulatory-dependent in this domain, and the regulation is arriving.

What the regulation is not designed to do is equally important. The DPP enforces traceability; it does not constrain production volume. EPR makes producers pay for end-of-life costs; it does not make recycled inputs cheaper than virgin materials. Current data confirms that recycled inputs do not yet cost less than virgin materials, contrary to the 2036 projection. The unsold goods ban penalises destruction; it does not guarantee that unsold goods are recycled rather than exported to secondary markets in West Africa or South-East Asia, where they arrive as waste rather than resource.

Each of these measures shifts incentives at the margin. None of them, individually or in combination, delivers the financially self-sustaining circular loop the piece describes. The transition to circular economics in 2026 is driven primarily by the financial penalties of non-compliance and the risk of losing market access, not by the commercial logic of circularity becoming genuinely more profitable than linearity. That inversion has not happened. The regulatory push will accelerate it. It will not produce it on its own.

The geographic limits of the regulation compound its scope. The EU framework governs market access for brands selling into Europe. It does not reach the producing countries where displaced garment workers are absorbing the costs of automation, where accountability failures in agentic procurement are most consequential, and where the Jevons dynamics are most directly felt. A Bangladeshi factory displaced by automation, operating under agentic procurement systems with no legal accountability, producing garments that will carry a Digital Product Passport when they reach Rotterdam—that factory is subject to the compliance demands of the EU regime without access to its protections. The asymmetry is the condition of a framework built to protect consumers and enforce brand accountability in wealthy markets, not to distribute the costs and benefits of industrial transition equitably across the supply chain.

Were the EU framework extended through binding trade conditionality, linking market access to labour protection and automation governance in producing countries, not only to product traceability, the geographic asymmetry would narrow materially. That extension has no legislative timeline. Fashion for Good ends by calling on the industry to act now so that 2036 becomes reality. What the regulatory evidence from 2026 actually shows is more constrained: the foundations are being laid, and they are real. But the distance between those foundations and the self-sustaining, equitable, AI-native industry the piece imagines is a political gap—in production volume governance, in worker transition infrastructure, in cross-border accountability architecture—that ten years of deployment will not close on its own. Whether that gap narrows depends on decisions being made now by institutions, brands, and governments whose current trajectory does not yet point in that direction.

The Questions Worth Keeping

The Fashion for Good piece matters precisely because it takes the technology seriously enough to map it in detail. That rigour creates a shared reference point, and shared reference points are where substantive industry debate begins. The structural questions this article has examined run alongside that mapping: what the rebound dynamic does to efficiency gains, what displacement looks like without a transition infrastructure, what autonomous decisions cost when accountability is absent, what regulation reaches and what it cannot. The 2036 horizon may be the wrong distance. The questions it has generated are the right ones.