Data, Content, Context, Slop

There are many stakeholders in the information economy. For decades, creators, workers, market intermediaries, media companies, and digital platforms have tussled over how information, articulated in different forms such as “works”, “content”, or “knowledge”, can be monetised and who should benefit. More recently, AI firms have entered this contest over how to define and justify the forms of profit that can be drawn off information as it flows from creators to consumers via intermediaries, platforms, and AI models.

These are tussles over the rules that organise the information economy and take shape as sectoral battles. Entities that would typically be competitors instead coordinate, deploying massive resources in efforts to ensure legal categories are interpreted in ways that provide entitlements, power, and structural advantages for that sector’s monetisation strategies.

This dynamic is acutely visible in the copyright disputes against AI companies for training their models, without licenses or permission, on materials that are ordinarily controlled and exploited by large media companies. The resulting litigation asks that courts evaluate how copyright doctrine and its internal categories like “reproduction”, “originality”, “creativity”, and “fairness” ought to adapt to manage AI as a form of computation and set of commercial practices. The settlements reached on the meaning of these categories become the legal background against which commercial operators then engage in pragmatic deal-making that results in the commodification, assetisation, monetisation, or other valorisation of information.

The editors of this Symposium asked me to contribute a piece addressing how law ought to govern what counts as AI-generated knowledge, with a focus on whether existing legal categories continue to be helpful, or whether more attention ought to be paid to the AI tools used in the generation of knowledge. The comments above about the conceptual battles structuring the AI economy are intended to emphasise that the legal regimes governing information – what legal categories capture it and how it ought to be exploited – are resolving in a manner far more pragmatic than principled. The solutions to questions such as the legal status of AI-generated information are rarely (or are only ever partially) resolved in the legal doctrine, and more clearly manifest in the ways that commercial actors work to shape legal environments around their preferred forms of AI-inflected profit-seeking.

“Data” or “Content”

These dynamics look different at different points in the pipelines, supply chains, and forms of industrial organisation that make up what we call AI. I offer three examples below. The first is early in the AI supply chain, at the stage of model training, where massive amounts of text, art, video, and other media forms are ingested to create AI models capable of parsing, classifying, creating, and acting. Many of the copyright disputes have been levelled at this point, but this jurisprudence is struggling to conceptualise the ways in which the legal-economic status of information is adopting a critical duality in its encounter with AI. On one hand, this information is “content” or “knowledge” whose value is defined by what it expresses to a human. On the other, this information is also “data”, whose value is defined by what it reveals, statistically and in aggregate, about the relationship between concepts.

The companies in these disputes are trying to realise one or the other of these value forms – expressive value and data value – in different ways. Traditional media companies, for instance, want to maintain the legal status of this information as content, to be exploited and monetised through regimes like copyright, that they have successfully influenced and corralled over decades to maximise profitability. AI firms, alternatively, have built businesses organised around exploiting this information as data – where all sorts of materials are ingested into AI models to then power the AI systems that are sold to users on different terms.

Copyright doctrine simply doesn’t have the conceptual capacity to mediate between these different business models or value forms. Both sectors’ claims are correct – yes, “content” is being reproduced for model training, but the result is not a reproduction or collage of the inputs. Rather, the outcome is a statistical model of how media systems work that is capable of transforming the conceptual units of media into reproducible styles. Law alone cannot solve the conundrum, especially with the stakes as high as they already are. The role of doctrine is, instead, defining the background conditions of sectoral negotiations in ways that create price pressure one way or the other. There is no principled legal way through this reality. The legal status of information is being resolved almost exclusively in private deal-making (and legal reforms that water down legal doctrine to create more pragmatic contexts for deals, as we are currently seeing in Australia).

“Context”

The second point in the AI supply chain where existing legal categories struggle to conceptualise the nature of information is in the operation of AI software products used for information work. Academia, specifically academic publishing, is a good example.

On one hand, academic publishers similarly straddle the data-content divide. They are businesses traditionally built around exploiting copyright. The parasitic dynamics of that business are well understood, but, extremely simply, they usually involve academics providing exclusive copyright licences to their research outputs to publishers, and publishers monetising that licence by selling subscriptions to their journals back to universities. Where authors would provide non-exclusive licences or other open access arrangements, the publishers demand cash for the privilege. AI software is a new revenue pathway for publishers that leverages their existing control over the content academics produce to simultaneously exploit its data value to power their new generative AI search, query and workflow tools.

LexisNexis, for example, now calls itself a tech company rather than a publisher. This makes sense considering its background in information retrieval. But this orientation also signals the centrality of that company’s AI software to its business model. Important here, however, is the way that this type of AI ecosystem produces another techno-legal information category: “context”. A specific further hybridisation of “data” and “content” into a new form that is operationalised through product harnesses and model context protocols. The exploitation of context is effectively the value that software firms offer when building on top of models provided by AI companies.

Here, the business is less about having AI models train on content so that it can regurgitate that material as an answer to a query. Rather, the goal is to use that information available to a query from an AI model or agent. This becomes a new way to monetise access to information – not as academic content with expressive value consumed by humans, or as data with data value absorbed by models, but as context whose value manifests through its structural availability and machine readability within an AI ecosystem – a kind of traffic value. The business model for this arrangement is called subscribe-to-context, and it depends largely on enforcement mechanisms that do not exist yet. Specifically, inchoate techno-legal mechanisms of query-based licensing that define the availability of material to AI agents. But at the moment, this depends more on technical information controls than legal ones, which aren’t really able to grasp the doctrinal consequences of information being read and reasoned over by an AI agent for the sake of automating the production of new information.

“Slop”

The third point in the AI economy that legal categories stumble is at the point of information creation. It is worth looking at this issue as an extension of the developments outlined above. The legal arrangements that make the academic publishing business models feasible, at their heart, still rely on an academic or author creating an information artefact that is legally excludable and exchangeable as a property right.

As this copyright-protected work encounters datafication and contextification in service of automation, a contradiction emerges. Academic publishers obtain a licence to the copyright created by an academic to build new tools that automate elements of an author’s workflow. Taken to the end point, this automation begins to challenge the human’s authorial status and contribution. Because software systems cannot be authors that hold copyright, there may no longer be a copyright in works that publishers can exclusively exploit.

Were academic publishers not currently advertising jobs for humans to edit fully AI-generated content, that would be a less ridiculous proposition. But they are, and the reason is not necessarily to ensure the quality or rigour of the materials produced. Those human editors serve to resolve the copyright contradiction by giving automated content an authorial copyright “wash”, sufficient to create and sustain the property right that anchors the business. This is humans mopping up “slop” to ensure its commercial value. Rather than any principled or doctrinal evolution of the meaning of originality, creativity, or attribution, the status of AI-generated information is mediated by brute force commercial arrangements that site a human within an automated workflow in order for them to accumulate the legal privileges of authorship.

That dynamic is equally clear in the outcome of the dispute between Hollywood studios and the US Writers Guild in 2023. The resulting agreement stipulated that AI cannot be considered a writer capable of creating literary material. But it does clarify that studios may give fully AI-generated material for human writers to rework into literary material. This maintains the human’s pay and status as a writer (rather than those of a rewriter), while protecting the studios’ strategic imperatives around ensuring they still generate and control intellectual property while using AI to automate the production process. The terms of the agreement do not speak to the special creative powers of humans, only the capacity of humans to attract copyright protection. This is how the strange and unstable nature of human-machine relations is being managed in law – by deals, not by doctrine. The Writers’ Guild Agreement even specifies that when a writer uses generative AI, incorporates generative AI, or rewrites generative AI outputs, the outcome is literary material produced by the human and not by the AI. The legal status has effectively no relation to actual material practice.

Unquestionably, the way information is being generated has changed. But managing attribution is less about the nature of creativity in the AI context, and more about the manipulation of accountability and privilege in ways congenial to commerce. What counts as legitimate knowledge production is outsourced to labour deals, research integrity bodies, or sectoral convention (like court rules) that specify acceptable configurations in particular contexts. The result is that attribution indexes industrial and technological practices, not the contributions of humans or machines.

The result is that the legal categories defining the ownability of intangibles, such as creativity and originality, are legal terms of art modulated by doctrinal disputes, but ultimately determined by more practical considerations like market power.