The strategy consultancy Oliver & Ohlbaum (O&O) has been commissioned by a group of publisher organisations to attack several industry studies (including one commissioned by CCIA) considering the impact of a text and data mining (TDM) exception. I won’t go line by line here, but instead address the main themes in their criticism.
The common theme throughout O&O’s analysis is using the HM Treasury Green Book as a rhetorical prop, demanding that every study boil the ocean and address all kinds of issues outside their express scope. The studies they are critiquing estimate specific impacts: the impact of a commercial TDM exception on UK AI investment, or AI adoption in the wider UK economy. The Government should take those existing studies into account as part of a more comprehensive impact assessment. Those studies are not, never claimed to be and do not need to be a comprehensive assessment of every potential impact of a commercial TDM exception, however.
Misunderstanding of the purposes of the HM Treasury Green Book
There are endless references to the Green Book and how the studies included apparently do not meet its standards. O&O seems to not understand what the Green Book is intended to achieve.
The Green Book is a guide for organisations “developing, reviewing or approving proposals that use public money” and in “the appraisal of regulations, which is required in regulatory impact assessments (RIAs).”
The Green Book is therefore guidance for organisations looking to develop a complete assessment of a proposed regulation, and recommends taking account of a broad range of potential consequences (e.g. regional impacts). It is not an appropriate standard (or intended to be) for every organisation that seeks to provide evidence contributing to such an overall assessment and the wider public debate over options for government. It only takes a cursory look online to find that as recently as June last year, O&O itself produced a study which does not appear to consider impacts on UK regions, for example.
The Government has asked (through its consultation and other engagement) for evidence about the impact of potential copyright reforms to support AI development in the UK. These studies respond to that need by attempting to address specific relevant questions. It is not “fantasy economics” to produce a study that does not consider every claim by rights holder organisations.
For what it’s worth, CCIA has published an analysis (apparently not considered by O&O) of the appropriate approach to an overall consideration of the economic impacts of a UK TDM exception, which explores many of the impacts O&O claims have been neglected (e.g. licensing).
‘Have your cake and eat it too’ approach to licensing
O&O repeats a rights holder claim that a commercial TDM exception would have an impact on a mature licensing market, which the consultancy claims the studies ignore. However, all of the licensing deals it cites involve at least one U.S. entity. Several involve only U.S. entities. This fits with a broad understanding in the sector that the U.S. is where the most active market for licensing has developed.
In the U.S., firms rely on the fair use doctrine to enable AI training, and training has been found to be fair use in multiple recent court decisions. The point of a commercial TDM exception would be to support UK companies and researchers so they can compete with competitors abroad which, in every other major economy, have similar protections. If a thriving licensing market is incompatible with protection for AI training, why are all the examples cited by rights holders involving organisations training in markets with protection for AI training?
O&O simply repeats the arguments from the organisations commissioning its research, which want to cite positive examples of licensing from the U.S., without reflecting on the policy environment there, which provides the protection for AI training they oppose.
Versus the status quo, a UK commercial TDM exception does not undermine the incentive to strike licensing deals. It simply affects whether activity dependent on such an exception occurs in the UK or not. The same training, using the same data, without a licence, can occur elsewhere.
Even in a scenario where the UK attempts to apply its copyright rules extra-territorially (with serious consequences discussed by Prof Noam Shemtov in a recent paper) the impact on licensing is ambiguous. It is not clear that it would be practical and economical to license for UK-specific AI models on a scale sufficient to make up for the loss of opportunities to licence data in a more dynamic, globally-connected AI market (e.g. data that is not publicly-available).
Incentives for licensing go beyond access to data. In the vast majority of cases, licensing agreements between rights holders and AI developers focus on creating value to the user at the output stage. The existence of a commercial TDM exemption should not impact this incentive, i.e. licensing will still occur with a commercial TDM exemption. This is likely to happen more in a market with a healthier AI ecosystem.
Impact analysis
O&O argues that industry’s studies “do not consider why copyright is decisive compared to other factors in the UK” and ignore other constraints. This is flatly untrue.
The Centre for British Progress study looks at different trajectories in AI investment between the UK and EU with more restrictive regimes and the U.S. with the fair use doctrine. As the study says, it “assumes removal of other capability constraints that might limit how much investment takes place” and it is important to note that this is broadly what the (released at around the same time) AI Opportunities Action Plan sought to do, with then-Secretary of State Peter Kyle pointing to “decisive action to support the AI sector and take down the barriers to growth.” In that context, it seems perfectly appropriate to ask to what extent the UK’s ambitions in the sector might be capped by having a less supportive regulatory regime than every other major economy. The report explores at length why a commercial TDM exception is important to AI development.
The other two studies estimate impacts versus existing activity. In the case of Public First, current levels of TDM activity. In the case of Europe Economics, current investment. That existing activity will reflect a broad range of factors, including energy costs and the like. The studies then consider the impact of a commercial TDM exception based on direct engagement with the sector and surveys engaging directly with those developing and/or using these services. For example, in the survey used by the Europe Economics study commissioned by CCIA, 99% responded that AI development is reliant on text and data mining using publicly-available data, with a full 58% responding “very reliant.” Thinking about the work respondents and their business do in the UK, if other jurisdictions implement protections for text and data mining and the UK does not, 66% reported that projects would have to take place in other countries, with 25% reporting that “many” projects would have to do so.
On that basis, the studies develop reasonable scenarios for the impact of choices around copyright specifically on the activity they are studying. Europe Economics’ study finds that AI-related TDM investment would be 17-38% lower. Public First finds a loss in GDP contribution, across the total economy, of 43%. In neither case, therefore, are these studies assuming that no AI investment is possible without a commercial TDM exception, or that there is no other constraint on growth.
Alleged contradiction
O&O regards it as a contradiction that the studies acknowledge existing investment in AI training and other TDM-related activities in the UK, while arguing that the lack of an exception undermines such investment.
There is no contradiction here. These studies are engaging with a complex sector in which there is a diverse range of activity more-or-less affected by whether a commercial TDM exception is in place. They are also attempting to understand the behaviour of companies that are making decisions (whether to invest in the UK, whether to invest in using AI tools more, etc.) based on expectations for the future regulatory environment. At the moment it is unclear whether the UK intends to permanently depart from the approach to these issues seen in every other major economy. If we resolve that uncertainty, and the UK has a less supportive regime than the U.S., Japan and the EU, it will affect the choices of companies deciding whether or not their future is here or in other jurisdictions.
O&O specifically points to the Europe Economics study on this point but misunderstands its argument. The Europe Economics study takes the conservative approach (reflecting the investor decision-making process described above) that, while a commercial TDM exception will affect the growth of commercial TDM-related activity in the UK, it does not assess the impact on past investment and the loss that has occurred as other countries have taken action and the UK has progressively fallen behind. This is not a contradiction.
Empty claims about UK comparative advantage
O&O asserts that “the use of licensed and high-quality data in AI development supports the transparency and auditability of AI models, which can support trust and adoption across the economy; this is not assessed.”
This is misleading and does not reflect the technical reality of model training. Mandatory licensing considerably limits the availability of data and thereby considerably decreases the quality and reliability of AI models. It also drives smaller players out of the market. It is also unclear why licensing would add any meaningful transparency since most deals are confidential.
Regardless, they do not provide any argument for why this should be considered. With or without a commercial TDM exception, there is nothing to stop companies adopting such models if the benefits are there. As they note repeatedly, at present, in the United States, which has protection at least equivalent to a proposed UK TDM exception, licensing does take place.
“Secondly, the next phase of AI development is in high quality data for RAG and for the development of sector-specific and expert AI tools. It is here that the UK has potential competitive advantage, but this is not assessed by any of the studies.” The Public First study in particular does consider sector-specific development extensively. However, O&O again does not provide any meaningful argument for why these developments should be considered. The reality is often that this kind of specialist development is hard to separate from the more general large language models that would benefit most from a commercial TDM exception, one is an input to the other, so there will be an impact on UK competitiveness in attracting exactly this kind of activity.
International examples
O&O confusingly cites the Green Book in arguing that assessments should look to international comparators, which all of the studies discussed do to varying extents.
They then cite some rights holder complaints about the Japanese copyright exception. If any of these studies purported to provide a comprehensive analysis of the arguments for and against that copyright exception, that might be relevant. However, O&O provides no evidence that these rights holder complaints are credible or material to a consideration of the impacts in the UK that the studies are seeking to assess.
For what it’s worth, the Japanese example has been considered by Prof Andres Guadamuz, who points out that there is no apparent impact on Japanese creative industries.
Japan has also adopted a notably AI-friendly copyright law, which permits the use of copyrighted works for the purpose of information analysis without the need for a licence, provided it does not “unreasonably prejudice the interests of the copyright owner.” The Japanese exception for TDM has positioned them as an attractive location for AI research and development, and again, there is no substantive evidence that this has led to a decline in its vibrant creative sector. While the reach of the existing exception is debatable, it is clear that its existence helps to debunk any argument that the adoption of TDM exceptions affects negatively the creative sector.
Conclusions
There are obviously challenges in estimating the impact of policy changes in the context of a fast-moving technological ecosystem like AI. All of these studies have made a good faith attempt to provide evidence on important impacts relevant to policymakers considering the case for a UK TDM exception. O&O is arguing those efforts are illegitimate because they do not cover all the bases expected of a comprehensive UK policy appraisal, while uncritically citing every unevidenced claim from rights holders they could find.
It is disappointing that, instead of commissioning credible analysis on questions that are clearly important to them like licensing, publishers have instead found a consultancy willing to write a hit job.
