The circular economy, in broad terms, seeks to reduce waste and resource extraction by preserving the value of products and materials over time. Yet the authors emphasise that there is no single agreed definition. Frameworks such as slow, close and narrow loops, designing out waste, keeping products and materials in use, and regenerating natural systems all circulate simultaneously, and even core concepts like regeneration remain open to interpretation. This interpretive openness is precisely where the trouble begins. An LLM cannot reproduce the entire field in a single answer. It must select and organise material, bringing some ideas into view and leaving others out. In a discipline whose meanings and priorities are still unsettled, that selection determines which interpretations users encounter most readily, and which fade from sight.

The first of the paper’s three analytical dimensions, the Content dimension, examines how AI constructs and curates circular economy knowledge. The mechanics matter. LLMs learn from vast training corpora spanning websites, books, academic papers, policy documents, corporate reports and media content, adjusting their internal parameters to capture recurring relationships between tokens and their contexts. When answering a prompt, they predict what is likely to follow, one token at a time. Crucially, the answer is generated from statistical regularities in the source material, not retrieved as a settled account or independently assessed for the strength of its claims. A formulation that appears frequently in the corpus may be reproduced more readily than a better-supported but less visible alternative. The authors call the resulting convergence a data-driven consensus: agreement around the patterns most strongly represented in the available material, rather than through debate or critical assessment.

The consequences for circularity are concrete. Recycling, for example, remains one of the most extensively studied and adopted circular economy strategies and dominates implementation narratives, even though many materials can only be recycled a limited number of times, meaning recycling may prolong rather than replace linear production and consumption. Higher-value strategies such as refuse, reuse, maintenance and upgrading, or newer territory like nature regeneration, may receive less attention simply because they are less prevalent in the historical record. The authors illustrate the tendency with an exploratory exercise, reported in the paper’s supplementary information, in which six widely used public-facing LLMs were asked the same open question, can you define the circular economy, across three collection rounds in 2025. The exercise was explicitly illustrative rather than a benchmark, but it shows how models may carry forward the imbalances of their training data and amplify them in the accounts presented to users.

The knowledge base itself compounds the problem. AI scholarship has long flagged uncertain data provenance and opacity around how training material is selected and processed. For the circular economy, this is aggravated by the explosive growth and mixed composition of the literature since the mid-2010s, which blends peer-reviewed research with consultancy reports, policy documents and unvetted grey material. English dominates the field, so perspectives developed in other linguistic contexts are likely underrepresented. Meanwhile, the publish-or-perish pressures of academic life can prioritise quantity over robustness, and even the ISO standard for the circular economy struggles to harmonise competing interpretations. LLMs draw on this heterogeneous material without reliably distinguishing robust, contested or weak knowledge. Concepts may also suffer what the authors describe as knowledge decay, erosion or dilution: the 9R framework presented as the circular economy itself, circularity stretched to encompass almost any corporate initiative, or the term used as a generic label in policy and business. When such formulations become widely reproduced, models carry them into new answers without recognising that their original meaning has been narrowed, stretched or detached from its foundations.

The second dimension, the Transformative dimension, asks what this mediated knowledge does to innovation and entrepreneurship, the twin engines of the circular transition. Here the authors draw a sharp distinction between optimisation and systemic change. AI systems excel at answering specific questions: optimising routing, maintenance and fleet management, or designing more circular single-use packaging. But a meaningful circular shift in the automobile sector, for instance, requires redefining the car as a service rather than a product, alongside changes in consumer behaviour, municipal policy and manufacturer strategy. If AI is used only to make the linear model of private car ownership more efficient, it will not drive systemic reconfiguration. LLMs cannot independently assess socio-technical systems or determine that long-term sustainability goals should take precedence over short-term data patterns, which limits their capacity to propose the bold redesigns a genuine transformation demands.

Entrepreneurship presents a similar tension. Circular entrepreneurs often rely on inspiration, contextual insight and unconventional thinking, and grassroots circular entrepreneurs are frequently driven by non-economic motives such as passion for the environment and concern for social issues, prioritising long-term environmental and social benefits over immediate commercial returns. LLMs can simulate aspects of entrepreneurial thinking and assist with generating business ideas, crafting marketing strategies or analysing trends, but they lack the courage, emotions and willingness to defy prevailing odds that characterise many founders. More pointedly, because circular entrepreneurs often pursue ideas that challenge established ways of thinking, while LLMs tend to reproduce familiar and well-represented approaches, AI assistance may favour conventional solutions precisely where more radical innovation is needed. Innovation itself also depends on collective sense-making, trust-building and learning alliances through which actors negotiate standards and share risks; these are human, socially attuned processes that models cannot replicate.

The third dimension, the Moral dimension, confronts the ethics of relying on such systems. A central concern is what the authors term operational indifference to truth: LLMs optimise statistical likelihood rather than evaluate reality, prioritising plausibility over verification and occasionally producing hallucinations, plausible-sounding but fabricated content. In the circular economy context this can mean references to non-existent research, exaggerated claims about circular initiatives, or unsupported performance figures. Compounding the risk is AI assertiveness, the tendency to present responses with high confidence regardless of underlying uncertainty, and compliance bias, whereby models affirm or accommodate incorrect or misleading user statements because they are designed to generate cooperative, engaging responses. Repeated interactions with accommodating systems may create an echo chamber that reinforces conformist thinking, reduces critical scrutiny of circular practices and narrows the range of options considered legitimate. There is also the question of who bears the costs when decisions are poorly informed: biased outputs may direct resources towards established actors and familiar technologies while neglecting locally appropriate alternatives in emerging economies or local communities.

Underlying all of this is the absence of moral agency. LLMs may reproduce the language of ethical deliberation, but they do not act for moral reasons or understand the consequences of what they recommend, and they cannot assume responsibility for decisions informed by their outputs. The authors argue that responsible judgement, understood in the virtue-ethics tradition as integrity, fairness and consistency in weighing competing environmental, social and economic claims, must remain with human actors. Without it, circularity risks becoming cosmetic: products promoted as recycled or recyclable suggesting that consumption may continue largely unchanged, diverting attention from the need to reduce material throughput while leaving the structures of the linear economy untouched. Uncritical reliance on LLMs could even foster intellectual passivity, weakening the very reflection needed to scrutinise their assumptions.

The paper closes with governance priorities rather than despair. Model developers and organisations deploying LLMs should make the sources and assumptions behind their responses explicit, opening them to expert scrutiny by interdisciplinary teams of data scientists, environmental experts, social scientists and policymakers. Training techniques such as reinforcement learning from human feedback could encourage models to signal uncertainty and acknowledge contested claims, supported by standardised benchmarks for epistemic humility, the principle that AI outputs should not claim more certainty than the evidence supports. The human-in-the-loop principle offers a practical safeguard for ethically sensitive choices, and education should equip users to treat LLMs as a critical collaborator rather than an authoritative oracle. Used critically and responsibly, the authors conclude, LLMs can widen access to circular economy knowledge and support more informed deliberation. Used without sufficient scrutiny, they may reinforce prevailing interpretations and lend legitimacy to circularity in name only. The framework also opens a research agenda: comparing LLM responses across models, prompts, languages and user contexts, and testing model recommendations against expert judgement where environmental and social goals conflict with economic priorities. The stakes, the authors suggest, could hardly be higher, because whether AI helps challenge linearity or quietly preserves it will depend on the judgement of the humans who use it.

Subject of Research: How large language models mediate and shape knowledge of the circular economy

Article Title: The circular economy, explained by AI

Article References: The circular economy, explained by AI. (n.d.). https://doi.org/10.1007/s44498-026-00192-z

Image Credits: AI Generated

DOI: 10.1007/s44498-026-00192-z

Keywords: circular economy, large language models, artificial intelligence, epistemic intermediaries, sustainability, recycling, innovation, entrepreneurship, AI ethics, hallucination, governance, Journal of Industrial Ecology