Intellectual property used to sit at the edge of technology policy, important but compartmentalised. Copyright lawyers argued over songs and software; patent specialists contested molecules, chips and telecoms; competition authorities intervened only when exclusion became egregious. Generative and agentic AI has collapsed those boundaries. Models are trained on vast corpora of human expression, deployed through software interfaces, embedded in enterprise systems and increasingly entrusted with delegated action. In that setting, intellectual property is no longer merely a reward mechanism for invention. It is becoming part of the institutional architecture that allocates control over knowledge, behaviour and market access.
That is why this subject matters far beyond legal departments. The disputes now gathering around AI-assisted authorship, machine inventorship, training data, synthetic outputs, standard-essential patents and defensive publication are not technical skirmishes. They are arguments about who may participate in the next layer of economic organisation. If the first internet era was shaped by protocols and platforms, the agentic era may be shaped just as much by claims over datasets, interfaces, model weights, orchestration methods and compliance techniques.
The stakes are especially high for smaller actors. Large firms can absorb licensing uncertainty, litigate across jurisdictions and stockpile patents for cross-licensing. Individuals, one-person enterprises, universities, public-interest labs and smaller states cannot. For them, the structure of AI-related IP will determine whether they can build on common rails or must rent access to private toll roads. The category therefore deserves attention not because every claim of ownership is suspect, but because the balance between incentive and enclosure is being rewritten in real time.
Why AI turns old doctrines into live constitutional questions
IP law has always balanced diffusion against exclusion. Copyright grants limited control over expression, not ideas. Patent law offers time-limited exclusivity in exchange for disclosure of a novel, non-obvious and useful invention. Trade secrets protect confidential know-how so long as secrecy is maintained. Standards bodies, meanwhile, have tried to reconcile interoperability with proprietary rights through disclosure duties and fair, reasonable and non-discriminatory licensing commitments in sectors such as telecommunications.
AI destabilises each of these settlements. A foundation model can absorb patterns from millions of works without reproducing most of them in ordinary use, yet may still generate outputs close enough to raise legal and ethical concern. A researcher can use AI tools to identify molecules, optimise chip layouts or propose engineering solutions, raising the question of where human inventorship begins and ends. An agentic system can execute multi-step tasks according to a governance schema, prompting companies to seek IP protection not only over core models but over control methods, safety architectures, workflow protocols and machine-to-machine coordination.
IP is becoming the operating system of power in the agentic economy.
The result is that doctrines once treated as specialised are now deciding broad questions of industrial structure. Whether training on copyrighted material is licensed, exempted, compensated or prohibited affects who can train advanced systems. Whether AI-assisted inventions are patentable affects where R&D investment flows. Whether governance interfaces become open standards or proprietary stacks affects whether agents from different providers can interoperate safely. These are not peripheral issues. They shape the topology of the market itself.
Authorship after assistance
Copyright law in many jurisdictions still turns on human authorship, even if the precise language and case law differ. That principle has survived the arrival of powerful generative systems. The hard question is not whether a model is an author in its own right; most major systems still reject that proposition. The harder question is how much human creative control is required for AI-assisted output to attract protection, and how that control should be evidenced.
As of 2026, the practical trend is toward recognising copyright in works where a human has exercised meaningful selection, arrangement, revision or direction, while denying protection to purely machine-generated output with no sufficient human authorship. This does not eliminate uncertainty. Prompting can be trivial in one context and deeply iterative in another. Editing may be cosmetic or substantial. Multi-agent creative pipelines further complicate matters: if a person specifies the constraints, chooses among drafts, restructures the result and integrates it into a larger work, there is plainly human agency, but the boundary is not neat.
That ambiguity will matter commercially. Businesses want clean title in marketing materials, code artefacts, designs and media assets. Publishers want assurance that rights can be enforced. Creators want attribution and bargaining power. A likely consequence is the growth of provenance practices: retaining records of prompts, edits, source materials and version histories not as a novelty, but as evidence of authorship and chain of title. The law may move slowly; operational documentation will move faster.
Inventorship and the patent system
Patent law presents a parallel but distinct problem. Inventorship is not a ceremonial label; it determines entitlement, ownership and validity. Courts and patent offices in major jurisdictions have generally resisted listing AI systems as inventors. The underlying logic is practical as much as philosophical: the patent system is built around natural or legal persons who can own rights, assign them and bear duties of disclosure.
IP is becoming the operating system of power in the agentic economy.
Yet the refusal to recognise machine inventorship does not answer the real question. Modern R&D is saturated with software assistance. Scientists use models to generate candidate compounds, engineers to optimise structures, lawyers to draft claims and search prior art. The key issue is therefore whether a human can properly be called the inventor when AI contributed to conception. Patent doctrine has long coped with tools of varying sophistication, but AI increases the degree of apparent autonomy and the speed of ideation.
The sensible approach is neither to romanticise human genius nor to pretend that software is a legal person. It is to preserve a human-centred inventorship standard while demanding candour about AI’s role. Where a human defines the problem, evaluates the proposed solutions, recognises the inventive contribution and reduces it to a claimed invention, inventorship may remain human. Where no human can honestly explain why a specific claimed solution was conceived beyond selecting from model outputs, the case for patentability weakens. A system designed to reward disclosure and skilled contribution cannot rest indefinitely on fiction.
The risk of paper thickets
A second concern is volume. AI-assisted drafting and prior-art search can accelerate patent filing, particularly in software-heavy domains. That may improve disclosure, but it may also produce dense thickets of narrow, overlapping claims around implementation details, safety methods and orchestration techniques. Such thickets favour incumbents with litigation budgets and can chill experimentation by smaller entrants who cannot price legal uncertainty. Patent offices will therefore face pressure to examine software and AI claims with greater discipline, especially where abstractions are repackaged as inventions.
The training-data conflict is really about bargaining power
Public debate often treats training-data disputes as a morality play: creators versus machines, or innovation versus permission. The reality is more intricate. Training a model typically involves copying material into technical pipelines for analysis, optimisation and statistical learning. Rights-holders argue that this uses protected works at scale without consent and may erode markets for licensing. Model developers reply that learning from data is transformative, that outputs do not generally substitute for source works one-to-one, and that requiring ex ante licences for everything would freeze competition in favour of the largest incumbents.
Both sides have a point. Creators and publishers are right to resist a world in which their catalogues become free raw material for systems that may compete with them. Developers are right that an unworkably broad permission regime could make only a few firms capable of assembling compliant datasets. The policy challenge is therefore not to pick a moral tribe, but to design a settlement that preserves both a market for original creation and a contestable field for downstream innovation.
The answer will differ by asset class and jurisdiction. Text and data mining exceptions exist in some legal systems, subject to conditions. Licensing markets are emerging, though unevenly. Certain uses may be tolerated where outputs are non-expressive or where access is publicly available under lawful conditions; others may demand consent or remuneration. What should be resisted is the lazy assumption that all training is theft or, conversely, that all ingestion is socially equivalent to reading. Training is an industrial act with distributive consequences. It deserves an industrial policy response, not slogans.
The decisive battles will be fought less over singular inventions than over chokepoints: data, interfaces, standards and compliance methods.
Opacity is becoming a legal risk
One reason the training-data debate has become so combustible is that model development has often outpaced documentation. In many cases, even organisations that built large systems have struggled to provide complete, auditable accounts of what was included, under which terms, and with what downstream filtering. As regulation matures, especially in Europe under the AI Act and related copyright transparency obligations, opacity itself is becoming a liability.
This has implications beyond formal compliance. Investors, insurers, acquirers and public-sector buyers increasingly treat provenance as part of operational due diligence. A system trained on unclear material may still be valuable, but it carries contingent risk. Over time, well-documented data supply chains may matter as much as model performance in regulated sectors. That favours institutions capable of disciplined rights management, but it may also create openings for federated and smaller-scale approaches where training and fine-tuning occur on clearer legal footing.
Standards, patents and the governance layer
The next frontier is less discussed but potentially more consequential: standard-essential patents and adjacent control over agent governance. In telecommunications, standards bodies created common specifications to ensure that devices and networks could interoperate. Patents reading on mandatory parts of those standards became powerful, but their power was moderated by disclosure obligations and licensing norms.
The decisive battles will be fought less over singular inventions than over chokepoints: data, interfaces, standards and compliance methods.
Agentic AI may require analogous layers. Agents will need common ways to identify themselves, authenticate authority, represent permissions, exchange context, call tools, log actions, negotiate safety constraints and prove compliance. If such functions become standardised across sectors, IP claims over indispensable methods could become strategically important. A patent over a crucial coordination, identity, memory or oversight mechanism may not sound dramatic. In practice, if that mechanism becomes unavoidable for secure interoperability, it can become a gate.
This is where the category intersects with sovereignty. Open, royalty-free or low-friction standards can make the agent ecosystem legible and contestable. Proprietary governance rails can do the opposite, embedding dependence deep inside workflows that appear neutral on the surface. None of this means patents on technical contributions are illegitimate. It means the institutional design of standard-setting will matter immensely: who participates, how patents are disclosed, what licensing terms apply, and whether public-interest requirements are built in early rather than litigated later.
Why governance methods will attract filings
Safety and compliance are expensive to build and easier to commercialise when framed as protectable methods. It is therefore rational for firms to seek patents on model monitoring, audit trails, policy enforcement, tool access control, synthetic identity checks or collaborative agent protocols. Some of these claims will be meritorious. Others may amount to abstract business processes dressed in technical language. The danger is not merely higher licensing cost. It is that critical governance functions become fragmented across proprietary estates, making interoperation slower and public oversight harder.
Defensive publication may return to relevance
For much of the software era, defensive publication became a practical response to overbroad patenting. By placing technical details into the public domain as prior art, researchers and firms could block later claims by others. That strategy may regain importance in AI, especially for methods that are socially useful but whose enclosure would impose broad coordination costs.
Universities, public labs, non-profit collaborations and even commercial actors may choose to publish certain governance techniques, safety benchmarks, model evaluation procedures and interoperability methods precisely to prevent them becoming private bottlenecks. This is not anti-property ideology. It is selective stewardship. In areas where cumulative innovation depends on common access, publishing can protect freedom to operate better than a fragile hope that patent offices will reject every expansive application.
Defensive publication, however, has limits. It does not fund ongoing maintenance in the way some licensing regimes can. It does not resolve copyright disputes over training material. Nor does it stop enclosure through trade secrets, platform terms or de facto market dominance. Still, as part of a broader institutional toolkit, it offers smaller actors a means of shaping the knowledge commons rather than merely reacting to it.
Open standards versus proprietary moats
The rhetoric of openness can mislead. Not every open standard produces competition; not every proprietary system is harmful. Some closed architectures can deliver accountability, security and coherent user experience. Some ostensibly open ecosystems are captured by a few dominant implementers. The policy question is therefore more precise: where does society benefit from mandatory interoperability and low-friction access, and where do stronger exclusive rights still encourage useful investment?
In AI, the answer is likely to vary by layer. Core research breakthroughs may justify patenting in some fields, particularly where development costs are high and disclosure is meaningful. By contrast, basic governance and communication interfaces for agents may warrant a more open approach, because their social value rises with broad compatibility. Identity, permissions, auditability and protocol-level safety are closest to infrastructure. Allowing these rails to become proprietary moats would recreate, in more automated form, the dependency patterns that policymakers have spent years trying to unwind in digital markets.
A healthy settlement would therefore distinguish between inventions that deserve returns and interfaces that should remain contestable. That distinction is difficult in practice, because companies naturally seek to turn interfaces into assets. Yet making everything ownable is not neutral. It rewards control over chokepoints more than contribution to the wider ecosystem.
Trade secrets, model weights and the quiet enclosure of know-how
Not all control will come through copyright or patents. Trade secret law may prove just as important. Model weights, data-cleaning pipelines, reinforcement processes, system prompts, evaluation suites and post-training recipes can all be shielded through confidentiality rather than disclosed through patenting. For many firms this is rational: trade secrets avoid publication and may last indefinitely if secrecy holds.
A healthy AI commons requires both enforceable rights and deliberate limits on how far those rights can reach.
But a heavy turn toward secrecy changes the innovation bargain. Patents at least offer public disclosure. Trade secrets can preserve advantage while limiting independent verification, reproducibility and safety scrutiny. In the context of agentic systems used in critical settings, that raises awkward questions. How much opacity should society tolerate in exchange for commercial incentive? Regulators are unlikely to abolish trade secrecy, nor should they. Yet in safety-critical or infrastructure-like domains, access mechanisms for auditors, courts and competent authorities may become a necessary counterweight.
The geopolitics of AI IP
IP settlements will not emerge in a vacuum. The United States, European Union, United Kingdom, China, Japan and other jurisdictions are approaching AI through different mixtures of industrial policy, copyright tradition, competition law and administrative practice. That divergence matters. A world in which one jurisdiction tolerates broad text and data mining, another emphasises licensing, and a third backs national champions through procurement and standards participation will produce fragmented compliance and strategic forum shopping.
For smaller countries and firms, fragmentation is costly. They often lack the leverage to shape global terms yet must live with them. This is one reason international bodies and standards organisations matter more than they appear to in domestic debate. WIPO, OECD guidance, technical standards groups and national patent offices are slowly building the procedural environment in which future disputes will be interpreted. The most durable influence may come not from grand legislation alone, but from mundane examination practice, licensing norms and standard-setting procedures.
What a balanced settlement should aim for
The most serious mistake would be to imagine that the choice is between maximal ownership and complete freedom. Both are unstable. Overreach can lock innovation behind permission layers and turn governance into rent extraction. Under-protection can weaken incentives for expensive creation and transfer bargaining power from creators to compute-rich intermediaries. The aim should be a narrower, more disciplined settlement.
- Human accountability should remain central. Authorship and inventorship doctrines should preserve meaningful human responsibility rather than performative attribution.
- Transparency should be treated as infrastructure. Documentation of training inputs, licences, model lineage and output provenance will increasingly underpin legal certainty.
- Patent quality matters more than patent volume. Offices should scrutinise AI and software claims carefully, especially around abstract governance methods.
- Interoperability layers deserve protection from enclosure. Where agent ecosystems depend on common protocols, standards processes should favour fair access and resist hidden tolls.
- Defensive publication should be used strategically. Public-interest actors can preserve freedom to operate by placing critical methods into prior art.
- Creators need viable bargaining structures. Training-data policy should not simply assume that either uncompensated extraction or universal bilateral licensing is workable.
None of this will make the field neat. AI is not one technology and IP is not one instrument. But the direction of travel is clear enough. Ownership disputes around models are maturing into governance disputes around ecosystems. The deepest question is no longer who pressed generate, nor even who first conceived a machine-assisted idea. It is whether the legal architecture of AI will distribute agency broadly or concentrate it at the points where everyone else must connect.
A healthy AI commons requires both enforceable rights and deliberate limits on how far those rights can reach. That is why intellectual property and patents belong at the centre of any serious analysis of the machine economy. They are not merely about protecting invention. They are about deciding the terms on which intelligence itself may be organised, exchanged and governed.
