Culture is being turned into raw material at industrial scale and returned to us with the names rubbed off.
The systems doing this are neither illegitimate nor inevitable. They are being built now, under rules that are still being written. The question is not whether generative systems will exist. It is on whose terms they will be allowed to draw on human work.
For too long this has been staged as a fight between creativity and technology. It is not. It is a question of infrastructure.
Whatever settlement you favour, it needs the same plumbing underneath it: a reliable way to know where a work came from, who made it, what permissions attach to it, what use was made of it, and what happens when those permissions are ignored. Without that, consent is theatre, licensing is guesswork, credit evaporates and enforcement belongs only to whoever can afford the lawyers.
The Manchester Principles set out the minimum conditions that infrastructure must satisfy. They are deliberately technology-neutral. They describe obligations and outcomes, not products. The point is not to freeze technology in place. It is to make sure progress has foundations strong enough to carry trust.
Manchester is the right place to say this because this city and the towns around it have seen the argument before: who gets to set the terms when technology reorganises production?
At Rochdale in 1844, the Pioneers built a co-operative on the idea that people could organise the terms of trade for themselves. In Manchester in 1868, working people from across Britain convened the first Trades Union Congress at the Mechanics’ Institute. Peterloo belongs to the same longer history of demanding a voice in systems that govern people without adequately representing them.
The lesson is not nostalgia. It is institutional design: when power becomes concentrated, durable rights require durable machinery.
There is a wider lineage too. Article 27 of the Universal Declaration of Human Rights holds two ideas together: everyone should be able to participate in culture and share in its benefits, and authors are entitled to protection of the moral and material interests arising from their work.
The Manchester Principles take that balance seriously. Access to culture and technological progress are public goods. So is the capacity of the people who make culture to remain visible, solvent and free to decide the terms on which their work is used.
They are a shared statement of the minimum conditions under which human creative work can be used to train, fine-tune, condition or generate synthetic content.
They are not a technical standard. Standards bodies already exist to do that work, and several different standards should be capable of satisfying these Principles.
They are not an endorsement of any company, platform, registry or protocol. A framework that only works if one provider owns the rails has failed before it starts.
They are not a restatement of existing law. Several of these commitments ask more than the law currently requires in some jurisdictions, and that is deliberate. They describe what an adopter undertakes to do, not the minimum it can be compelled to do.
They are not an attempt to displace the exceptions and limitations that already exist in copyright law. Research, private study, quotation, criticism, review, news reporting, teaching, preservation, archiving and accessibility are, where provided for by applicable law, load-bearing parts of a functioning cultural settlement. These Principles are directed principally at the commercial development and deployment of generative systems, not at the ordinary business of scholarship, journalism, libraries and education.
They are not an argument against generative technology. They are an argument that powerful technology should be built on consent, evidence and accountable exchange rather than contested foundations.
The four pillars are presented in the order in which the argument runs, not in order of priority. Each depends on the others.
Permission must be given, not assumed.
Use of protected work to train, fine-tune or condition a generative system requires permission granted in advance by the relevant rights holder.
Silence is not consent. A regime that requires millions of creators to discover invisible uses and object one by one should not be mistaken for a consent regime.
1.1Opt-in as the default. No training, fine-tuning or output conditioning on protected work for commercial purposes without permission obtained beforehand.
1.2Machine-readable permissions. Terms expressed so they can be discovered, read and honoured automatically at scale, without being proprietary to one licensor, platform or model provider.
1.3Routes proportionate to the rights holder. Collective licensing for those who need it, direct negotiation for those who want it, and a workable path for the unaffiliated individual with neither a lawyer nor a collecting society behind them.
1.4Remuneration that can track use. Payment structures capable of reflecting the extent and character of a work’s contribution, rather than treating an entire catalogue as an undifferentiated bucket of training material.
1.5Permissions that travel. A permission granted once should remain legible wherever the work goes, across platforms, formats and jurisdictions.
1.6Lawful exceptions preserved. Nothing in this pillar overrides an exception or limitation properly available under applicable law. Equally, an exception intended for research, preservation, education or another protected purpose should not become a route to commercial-scale ingestion by another name. Where an exception or limitation is relied upon, that basis should be capable of being identified rather than left to assumption.
1.7Past use is in scope. Compliance going forward does not by itself settle what happened before. Where protected work has already been used without permission in circumstances that would require permission under these Principles, adoption carries an obligation to disclose that use and to negotiate in good faith over it, rather than simply to draw a line under it. This commitment is made without prejudice to any party’s position on the legal status of that use.
Origin must be recorded at the point of creation, not reconstructed after the fact.
Provenance is the load-bearing pillar. Licensing without provenance is difficult to enforce, attribution without provenance becomes guesswork, and enforcement without provenance turns into litigation about memory.
Detection after the fact will always matter, but it cannot be the foundation. It is an arms race against systems designed to improve.
Declaration at source is the better starting point: record origin when the work or synthetic output is made, then preserve that record through the chain.
2.1Disclosure of training corpora. Meaningful, auditable disclosure sufficient for a rights holder to determine whether and how their work was used. This obligation is retrospective as well as prospective and extends to systems already trained and in commercial deployment.
2.2Provenance captured at creation. Origin, authorship, contribution and rights information bound to the work when it is made, through tools people already use wherever possible.
2.3Open and vendor-neutral mechanisms. Built on published specifications and open governance. No single company should own the rails on which an entire sector’s provenance depends, and no rights holder should have to surrender their catalogue to a registry controlled by a counterparty.
2.4Durability through transit. Provenance designed to survive ordinary distribution, format conversion, compression and platform ingestion, and to fail visibly rather than silently when it is removed or broken.
2.5Declaration before detection. Synthetic origin asserted by the system responsible for producing it, with independent detection used as a check rather than as the primary source of truth.
2.6Proportionate privacy. Provenance records should carry what is needed to establish origin, rights and contribution, and no more. Transparency about works must not become surveillance of the people who make them. Pseudonymous and collectively held credits must remain possible.
Credit is data, not courtesy.
Attribution has spent two decades being treated as an editorial nicety that gets a little thinner at every handoff until it disappears.
In a synthetic economy it is not decoration. It is part of the mechanism by which contribution is recognised, licensed, paid and remembered.
3.1Persistent identifiers throughout. Creators, contributors, works, recordings and rights holders identified through established persistent identifier schemes, populated as close to source as possible and carried onwards.
3.2Attribution that survives distribution. Credit that persists through the supply chain instead of terminating at the first platform or intermediary that finds it inconvenient.
3.3Machine-readable and human-readable. Legible to systems for licensing and payment, and legible to people for recognition and accountability.
3.4Contribution as well as authorship. Recognition of the session players, engineers, writers, producers, editors and other contributors whose work is inside the finished object and who currently disappear first.
3.5Attribution at inference, not only at training. Where a specific work materially conditions a specific output, that relationship should be capable of being recorded and disclosed. This is stated as a direction of travel rather than a description of current practice. The techniques are immature. This clause should therefore be read together with the commitment to evidence over assertion below: it asks adopters to implement this principle as capability matures and to publish honestly about what is and is not yet possible.
A right without a remedy is decoration.
The first three pillars describe obligations. This one describes what happens when they are ignored, which is why it will be the first thing some parties try to negotiate away. It must not be.
4.1Auditability. Material disclosure obligations subject to independent verification rather than self-certification alone.
4.2Remedy proportionate to the claimant. A route to redress that a solo artist, photographer, writer or small producer can actually use. A right enforceable only by parties able to fund High Court litigation protects catalogues, not creators.
4.3Cross-border recognition. Rules that cannot be escaped simply by moving a training run, server or corporate entity to another jurisdiction, pursued through compatible national rules and international co-operation.
4.4Consequences that bite. Sanctions and remedies material enough that compliance is cheaper than breach, including, where lawful and proportionate, restrictions on commercial deployment of systems built through persistent unlicensed use.
4.5Remedy for past use. Where past use infringed applicable rights, there must be a practical route to redress. A framework that only looks forward risks rewarding those who moved fastest before anyone was watching.
4.6Transparency as the precondition. No enforcement mechanism works without the disclosure and evidence required under Pillar Two. Provenance and remedy must be designed together or both will fail in practice.
5.1Voice, likeness and performance. Protection must extend beyond rights in works to the voice, image, performance and identity of a person. Many of those most exposed to synthetic reproduction hold no rights in the material used to train the system: the session singer whose voice is cloned, the performer whose likeness is generated, the writer whose manner is imitated. The obligations in these Principles apply to personality and performance interests as they do to rights in works, through whatever legal route a given jurisdiction provides. Style itself is not claimed as property. What is in scope is identifiable stylistic imitation where protected by applicable law, or where deployed so as materially to misrepresent identity, endorsement or performance.
5.2Proportionality. Every obligation here must work for a freelance musician, an independent journalist, a small publisher and a bedroom producer, not just for a major rights holder with a compliance department. A framework that only functions at scale simply entrenches those who already have it.
5.3Openness. Governance of any shared mechanism arising from these Principles should be independent of any single commercial interest, including the interests of those who helped write them.
5.4Evidence over assertion. Claims about what technology can and cannot do should be tested, published and revisited. The Principles should be reviewed against what has actually been built, not what vendors say is possible.
5.5Interoperability with existing work. These Principles are intended to sit above and alongside existing standards, identifier schemes, licensing systems and rights frameworks, not to bulldoze them and start again.
Rochdale is invoked here as architecture, not atmosphere.
Three ideas from the international co-operative tradition are particularly relevant: democratic control, autonomy and independence, and co-operation between independently governed organisations. They correspond here to open governance, vendor neutrality and interoperability.
The Principles are offered in that spirit: a common floor that many bodies can stand on together, held in common rather than owned.
These Principles do not pretend the argument began here. They are deliberately in conversation with work already underway across copyright, creator rights, metadata and content provenance. Their job is to connect those efforts into a minimum operating settlement rather than invent another silo.
Creative Rights in AI Coalition. Its UK principles centre copyright protection, creator control, transparency and a functioning licensing market. The Manchester Principles share those objectives, then ask what infrastructure makes them executable across sectors and borders.
CISAC and the global creator community. CISAC’s AI principles emphasise authorisation, remuneration and transparency. The Manchester Principles share that foundation and add provenance as an evidence layer connecting those principles to practical attribution and enforcement.
Human Artistry Campaign. Its principles begin from the value of human creativity and the responsible use of AI. The Manchester Principles are compatible with that position but are more deliberately infrastructural: less a statement about what we value than a test of what systems must actually do.
C2PA, DDEX and other standards work. C2PA provides an open technical framework for content provenance and authenticity. DDEX develops interoperable data standards for the digital music value chain. These are examples of the sort of standards through which parts of the Principles may be implemented. Neither is treated here as the single answer.
The historical claim is equally modest. Rochdale, Peterloo and the first TUC meeting in Manchester are not invoked as decoration, and the Manchester Principles do not claim institutional descent from them. They are precedents for the same civic instinct: people affected by a system should have a hand in setting its terms, and those terms need institutions capable of outlasting the immediate dispute.
Any organisation, creator, institution or public body, in any territory, may adopt the Manchester Principles. There is no fee and no licence to negotiate.
Adoption commits the adopter to three things: implementing the Principles within their own operations; publishing within twelve months a short statement of how they intend to do so and where they currently fall short; and asking the same minimum conditions of relevant commercial counterparties.
An adopter may state publicly that it has adopted the Manchester Principles and may cite individual clauses. No adopter may describe itself as certified, accredited or endorsed under the Principles. There is no certification scheme.
Adoption may be partial. An adopter that cannot yet meet a clause should say so, say why and, where possible, say when it expects to do so. Honest partial adoption is more useful than unexamined full adoption, and Pillar Four applies to the accuracy of the statement either way.
Adoption attaches to a stated version. An adopter is not automatically bound by a later amendment without adopting it.
The Principles are published under a Creative Commons Attribution 4.0 International licence. They may be reproduced, translated, quoted, adapted and built upon by anyone, including commercially, subject to the terms of that licence. Attribution to the Manchester Principles should travel with the document in accordance with Pillar Three.
The Principles need a custodian, an amendment process and a review cycle. This version does not name them. Those are constitutional questions, and they belong to the people who agree with the Principles rather than to the person who drafted them.
The summit is therefore asked to settle four things: who holds the authoritative text and any public register of adopters; how amendments are proposed, published for comment and agreed; how often the Principles are reviewed against what has actually been built, tested and enforced; and how translations and derivative versions are related back to the authoritative text.
Two conditions are proposed for whoever takes custody. The custodian holds the text but does not own the ideas expressed in it and cannot restrict uses permitted by its Creative Commons licence. And no commercial party with a material interest in mechanisms implementing the Principles, including the author’s own company, should hold custody of them.
Manchester’s contribution to the industrial age was not simply machinery. It was the argument about who machinery was for, who benefited from it, and what rights followed when technology changed the terms of human work. We are having that argument again.
The Manchester Principles begin with a simple proposition: human beings should remain visible in the systems built from human culture.
Permission should be knowable. Contribution should be attributable. Origin should be provable. Rights should be enforceable.
None of that requires stopping technology. It requires building it properly.
This text is published for discussion and amendment. Comments, objections and dissent are welcome. If these Principles are worth anything, they should be capable of surviving argument.
These definitions are working definitions for the purposes of this document. They are deliberately broad and are expected to be tested and tightened through discussion and subsequent amendment.
Protected work. Any work, recording, performance, image, text, dataset or other subject matter in which copyright, related rights, database rights, performers’ rights or equivalent applicable rights subsist under the law of a relevant jurisdiction.
Rights holder. The person or entity entitled to authorise the relevant use, including authors, performers, producers, publishers, labels and their assignees and licensees, and collective management organisations acting under mandate.
Generative system. Any computational system that produces text, audio, image, video, code or other content whose behaviour is derived in whole or in part from patterns learned from existing material. The term is not limited to any particular technical architecture.
Training, fine-tuning and conditioning. Respectively: the initial derivation of model parameters from a corpus; subsequent adjustment of those parameters using further material; and the supply or use of material at or near the point of generation to shape a specific output, including retrieval, prompting with reference material, embeddings and style transfer. These activities are treated as relevant uses for the purposes of these Principles without determining their legal status in any particular jurisdiction.
Material conditioning. The use of a specific work in a way that makes a discernible and non-trivial contribution to a specific output. The threshold is deliberately left to be developed through evidence, practice and review rather than fixed here.
Provenance. Verifiable information about the origin, authorship, contribution, rights status and subsequent handling of a work or output, recorded at or near the point of creation and preserved through distribution.
Synthetic content. Content generated in whole or substantial part by a generative system, whether or not it is presented as such.
Persistent identifier. A durable, resolvable identifier for a person, work, recording, performance or organisation, issued under an established scheme and intended to remain stable over time.
Personality interests. A person’s voice, likeness, image, name and performance, however protected in a given jurisdiction, including through performers’ rights, image or publicity rights, passing off, unfair competition, data protection or other applicable law. The term extends to identifiable stylistic imitation only where protected by applicable law, or where the imitation is deployed so as materially to misrepresent identity, endorsement or performance. It does not assert property in style as such.
Adopter. An organisation or individual that has adopted all or part of a stated version of the Principles under the Adoption section above.
International Co-operative Alliance. Co-operative Identity, Values and Principles, and historical material concerning the Rochdale Pioneers (1844). https://ica.coop/en/cooperatives/cooperative-identity
Trades Union Congress. Historical material concerning the first Trades Union Congress at the Manchester Mechanics’ Institute (1868). https://www.tuc.org.uk/about-tuc/our-history
United Nations. Universal Declaration of Human Rights, Article 27. https://www.un.org/en/about-us/universal-declaration-of-human-rights
Creative Rights in AI Coalition. Principles. https://www.creativerightsinai.co.uk/principles
CISAC. Principles and policy concerning artificial intelligence and creators’ rights. https://www.cisac.org/AI
Human Artistry Campaign. Core Principles for Artificial Intelligence Applications in Support of Human Creativity and Accomplishment. https://www.humanartistrycampaign.com/
Coalition for Content Provenance and Authenticity. C2PA Technical Specification. https://spec.c2pa.org/
DDEX. Standards for data exchange across the digital music value chain. https://ddex.net/standards/
Creative Commons. Attribution 4.0 International (CC BY 4.0). https://creativecommons.org/licenses/by/4.0/