An ifo Institute model says tighter access to AI training data could sharply reduce Germany’s potential gains by 2034, reopening the dispute over innovation, creators’ rights and who captures the value.

The financial figure is large enough to sound like an invoice for a decision Europe has not yet taken. On Tuesday, October 13, the Munich-based ifo Institute said tighter restrictions on material used to train artificial-intelligence models could put €155 billion in potential annual German economic value added at risk by 2034. The claim injects a number into one of Brussels’ most difficult policy arguments: how to protect authors and publishers without making Europe’s digital industry dependent on systems built elsewhere.

The important qualification is that the €155 billion is not money already lost, nor a budgetary cost, nor an official prediction of German output. It is the estimated difference between two modelled futures. The distinction matters because policy choices will shape both whether AI improves productivity and whether the people producing the information on which systems train are paid fairly. A viable outcome has to consider both sides of that ledger, not treat one as an obstacle to the other.

A large number with a precise meaning

The ifo Institute release published on October 13 puts the potential annual value added from generative AI at approximately €440 billion for Germany by 2034. If access to training data became substantially more difficult, the same modelling exercise reduces that opportunity by €155 billion, leaving a theoretical €285 billion. Those figures describe a conditional potential at a future point, not a finding that German companies will earn precisely those sums. Nothing in the publication establishes that a particular draft law has already eliminated a corresponding share of national income.

That distinction is crucial for readers assessing economic warnings from research institutions. A headline about billions being ‘at risk’ compresses a series of assumptions about technological progress, adoption, business organisation and legal compliance into one visible result. Change the assumptions and the result can change. The most defensible interpretation is that training-data access may be a material variable in future productivity, large enough to warrant close examination. It is not evidence that the only economically rational copyright system is one that minimises payments to creators.

The research behind the warning

The calculation was produced by Implement Consulting Group in cooperation with ifo, and was commissioned by CCIA Europe and Google, according to the institute’s own published disclosure. The collaboration gives the analysis useful contact with a market in which model builders operate. It also means readers should recognise the commercial interests behind the commission. Companies whose AI products depend on broad datasets can benefit from rules that make training easier. Commissioning does not invalidate a study, but disclosure is necessary when its findings enter a legislative dispute.

The ifo project description identifies Article 4 of the EU’s 2019 copyright directive as the principal object of the exercise. The research considered the potential economic effects of altering the reservation of rights, licensing arrangements and transparency obligations for commercial text and data mining. Its new October intervention is therefore not a general measurement of every AI regulation in Europe. It is principally an argument about the conditions under which developers can analyse and train on materials to which they have lawful access.

Why the estimate is not a budget forecast

Models of technological gains are particularly sensitive to the boundary between possible efficiency and realised production. A software tool may save time on a task without increasing the quantity of goods or services sold. A firm may redirect saved hours toward improving quality, shorten its working week, cut costs or expand output. These outcomes have different implications for value added, wages and tax receipts. The ifo scenario seeks to capture aggregate potential, but firms must still make investments and organisational changes for those possibilities to turn into recorded economic activity.

Nor should the €155 billion be added mechanically to a government’s deficit projections or subtracted from existing gross domestic product. It is a hypothetical reduction in an upside scenario for 2034, not a cash transfer from taxpayers or an annual loss already incurred in 2026. A responsible debate should ask what probability to assign to each scenario, how the benefits are distributed and how the figures compare with alternative estimates. The size of a number can draw attention; its meaning determines what action it supports.

Four channels, not one regulatory penalty

The researchers divide their €155 billion estimate into four effects: €45 billion from slower adoption of AI, €75 billion from less capable models, €25 billion from lost research potential and €10 billion from diminished AI development in Europe. These components add up to the reported total. They show that the argument goes beyond whether a German startup can scrape an additional web page. It rests on downstream consequences: poorer tools can become less attractive to customers, while uncertainty may reduce what researchers and companies attempt to build.

But the parts should not automatically be read as four independent losses that have already been observed. They arise within a shared model and could interact in practice. Less capable systems may slow adoption, for example; weaker demand may then make it less attractive to invest in local models. Conversely, specialised systems trained on high-quality licensed material can outperform larger general datasets in particular industrial tasks. The policy question is how these channels would operate under actual proposals, not simply whether the arithmetic is consistent.

Training data becomes an industrial input

Generative AI models are developed by exposing statistical learning systems to extensive collections of text, code, images and other material. Developers use those examples to infer patterns that later inform generated responses. Data access therefore resembles an input to production, although it is unlike steel or electricity: much of it may be reused, its ownership can be fragmented and its quality varies enormously. The ability to process more lawful material may help systems cover more languages, disciplines and kinds of expression, but volume alone cannot guarantee usefulness or accuracy.

For Europe, this creates an economic-security complication. Many firms may use overseas models regardless of how European law develops. A company that can buy a mature foreign service cheaply has little reason to finance a rival foundation model unless it expects meaningful performance or commercial advantages. The question is whether European developers can achieve that differentiation with available datasets, expertise, computing resources and market demand. The ifo intervention gives data access prominence, but it is only one element of the competitive calculation.

The token comparison has limits

The institute says the largest copyright-free dataset in Europe contains roughly two to three trillion tokens, while leading international models have been trained using approximately 15 to 36 trillion tokens, as stated in its press statement. Tokens are units into which language processing systems break text; they are not equivalent to individual documents, verified facts or valuable original works. The comparison conveys a scale problem, but it does not show the precise quality of either pool or prove that the entire difference can be filled only with material requiring permission.

Training also involves decisions about duplication, contamination, language representation, filtering and evaluation. A model that ingests many repeated low-quality pages may learn less useful information than a carefully constructed collection. At the same time, access to specialised European language, legal, cultural and scientific material can be unusually valuable precisely because generic collections are uneven. Policymakers should separate the need for diverse, reliable inputs from any suggestion that every item reachable on the public internet should automatically be available on identical terms.

The copyright regime already distinguishes uses

The relevant legal foundation is Directive (EU) 2019/790 on copyright in the Digital Single Market. Article 3 creates a mandatory exception for text and data mining undertaken by research organisations and cultural heritage institutions for scientific research, subject to its conditions. Article 4 is broader and covers text and data mining, including commercial contexts, but its application depends on lawful access and whether the relevant rights have been expressly reserved in an appropriate manner. These provisions were developed before generative AI became the present mass-market phenomenon.

In practice, the distinction matters for universities, laboratories, publishers and commercial model builders. Research institutions are not simply subject to the same framework as every firm selling a generative tool. Equally, the commercial exception is not an unlimited licence to ignore copyright or technological restrictions. It sits within a larger legal system that includes ownership, contractual terms, access conditions and enforcement. Any account of a regulatory threat that omits the existing rights reservation mechanism risks making the policy choice sound more binary than it is.

An opt-out is a right, not a technical footnote

Under the commercial mining framework, a rightsholder can reserve its rights, with machine-readable means especially significant for publicly available online material. That can allow publishers and creators to signal that their works should not be used in particular data-mining contexts. How a developer identifies and honours those signals is a serious implementation problem, particularly when information passes through intermediaries, archives and datasets assembled over time. A clear standard could make lawful compliance easier without removing the underlying choice from the owner of the work.

The economics of opt-outs cut both ways. If declarations are invisible or inconsistent, smaller developers face legal uncertainty and larger companies may have an advantage because they can maintain compliance teams. If reservations are ineffective, rightsholders have little negotiating power even where their work contributes to commercial systems. A better mechanism would reduce ambiguity, help distinguish authorised access from prohibited uses and make it feasible for both modest publishers and modest technology firms to participate. The dispute is over its design, not the desirability of uncertainty.

The AI Act adds separate responsibilities

The European Commission’s guidance for providers of general-purpose AI models describes obligations under the AI Act that coexist with copyright law. Providers must develop policies to comply with EU copyright requirements and respect reservations of rights. They must also make a sufficiently detailed public summary of the content used for training. Technical documentation and information for downstream providers have separate roles. These are disclosure and compliance requirements; they should not be confused with an automatic change to the underlying ownership of creative works.

The Commission has published a template and explanatory guidance for training-content summaries, intended to give a common basis for transparency. The fact that a model provider describes broad categories of training material does not, by itself, settle every dispute about permission or payment. Nor does disclosure automatically establish infringement. What it can do is narrow the information gap that has made negotiation and litigation difficult. The more useful the disclosure standard, the easier it becomes to assess whether the legal framework works for both developers and rightsholders.

A 2026 review reopened the debate

On May 18, the European Commission launched a review of EU copyright rules, seeking evidence on how the 2019 framework has worked and on challenges associated with generative AI, including licensing and enforcement. That review forms a tangible setting for the October research. The institute says uncertainty around possible changes is already discouraging investment, a contention that should be recognised as the authors’ assessment rather than treated as a quantified finding of capital flight. Investors normally evaluate regulatory predictability alongside expected demand, costs and competitors’ progress.

The Commission has also consulted on protocols used to reserve rights from text and data mining, including mechanisms relevant to the general-purpose AI code of practice. The technical details are politically consequential: an opt-out that is universally readable can lower compliance costs, while incompatible protocols can generate confusion for users and publishers alike. The debate now reaches from abstract intellectual-property principle into software engineering, metadata, contract administration and cross-border enforcement. Those details may determine the workable compromise more than broad declarations about being pro-innovation or pro-copyright.

European lawmakers defend the other side of the ledger

The European Parliament did not approach the question only as a potential cost to AI developers. In a March 10 resolution approved by 460 votes to 71, with 88 abstentions, members called for transparency about protected works used in training, fair remuneration and ways for rightsholders to refuse use of their material. They also advocated licensing frameworks accessible to smaller creators and enterprises. These are parliamentary recommendations, not a new binding copyright law. Nevertheless, they demonstrate the political weight of an argument that is largely absent from a model centred on the benefits of cheaper data access.

A newspaper, independent artist, technical publisher or specialist database may incur substantial costs creating material that later becomes useful for model development. If AI services substitute for subscriptions, licensing income or visits, losses can arise even as productivity improves elsewhere. The correct economic comparison should therefore consider changes to the production of valuable new content, not only the marginal convenience of reusing existing content. A system that drains investment from reliable original material could eventually weaken the information environment on which advanced models depend.

Who benefits and who bears the costs?

Value added across an economy is not the same as profit for AI vendors. Savings might accrue to manufacturers, retailers, public administrations and consumers that use AI to reduce costs or offer better services. Some could appear as wages or additional output; some might be retained by platform operators. On the other side, artists, media businesses, researchers and data curators supply inputs that may have substantial commercial worth. The distribution of gains affects whether the ecosystem sustains the upstream investment that AI developers require.

This is why a simple choice between total access and total prohibition is misleading. Licences can permit access with payment; research exceptions can preserve noncommercial inquiry; technical standards can make lawful uses easier to identify. Different models of compensation may have different costs and benefits, and not every dataset warrants the same treatment. What Europe must avoid is a framework in which only the largest actors can afford compliance and bargaining. Such an outcome would leave the continent with both a less diverse creative economy and fewer competitive model developers.

The commissioner's identity belongs in the story

Industry-funded modelling is common in regulatory debates, from energy markets to banking and telecommunications. It often identifies channels that officials ought to examine. But a sponsor’s economic incentives can influence which questions a research programme prioritises and which counterfactuals receive attention. Google has extensive interests in AI systems and digital information markets; CCIA Europe represents technology-industry interests. Those facts are relevant context for evaluating the assumptions, not a substitute for evaluating the work on its merits.

The ifo release is unusually explicit in stating that its numbers are estimates of potential rather than forecasts and in identifying the commission. The next step for public scrutiny is therefore methodological comparison: what happens under a different rate of adoption, a viable licensing market or improved European access to high-quality data under enforceable rights reservations? A balanced policy assessment would also estimate gains to creators from compensation and losses from any decline in original production. Without such comparisons, the most dramatic outcome of a model can eclipse alternatives that policymakers have not yet tested.

German companies are already adopting the tools

Germany is not waiting for 2034 to experiment. An ifo survey released on June 5 reported that 54.5% of German businesses were already using AI in their processes, up from 40.9% the previous year. The survey put adoption at 58.7% in industry and 56.2% in services, with retail near 45%. These are survey responses about use, not evidence that each firm has recorded a measurable productivity gain. Nevertheless, they establish that the technology’s potential effect is no longer limited to laboratories or a few large digital companies.

The widening adoption base changes the public-policy stakes. Rules aimed at model builders can indirectly affect thousands of businesses that never train foundational systems, because those businesses depend on tools, prices and quality set upstream. Yet the connection is not automatic. A company purchasing an established product may have access to multiple competing suppliers, and changes to European data access may affect products unevenly. It is important to measure actual effects on service quality, customer choice and cost rather than treating every model-level constraint as a directly equivalent loss for every user.

Buying AI is not the same as building it

The same German business survey found that nearly three-quarters of AI-using companies rely on paid external solutions; 48.4% use free offerings, while only 18.7% develop their own AI systems. Those categories are not mutually exclusive, but the pattern is revealing. For much of German industry, the immediate challenge is choosing, governing and integrating an application rather than training a model from scratch. Access to training data may influence what suppliers can build; purchasing decisions also hinge on security, reliability, integration and demonstrable business value.

That distinction helps explain why Europe could benefit from more than one industrial strategy. It can encourage competitive local model development while also ensuring smaller firms can use reliable tools from many sources. It can facilitate lawful datasets without assuming all value is captured in the foundation model. It can finance education and process redesign rather than concentrating resources only on computing clusters. The ifo estimates highlight the importance of the upstream layer, but growth in the broad economy will depend on the practical adoption decisions made by ordinary companies.

Smaller firms need workable rules

Large multinational businesses can negotiate bespoke data licences, employ copyright specialists and commission legal reviews. A smaller publisher or software developer may lack those resources. For the developer, uncertainty can mean abandoning a potentially useful specialised model. For a creator, negotiation costs can make an otherwise valuable claim uneconomic to pursue. Uniform technical standards and standardised terms could help both sides, provided they do not eliminate individual rights or impose a disproportionate fixed administrative burden on the smallest participants.

There is also a market-design risk in complicated compliance. If each work must be investigated through a proprietary register, dominant platforms may become the only firms capable of managing the process. But if there is no credible way to identify ownership or opt-outs, smaller publishers could be forced to tolerate uncompensated use. The task for regulators is to make lawful transactions easier, not merely to make permission cheaper for a particular industry. The success of any framework should be judged partly by how many independent participants can use it.

Factory data is not a public web archive

German manufacturing has a particular interest in AI applications built around maintenance records, engineering documents, production measurements and supply-chain information. These datasets can contain trade secrets, personal information, contractual restrictions and security-sensitive details. The economic value of analysing them may be high, but the legal obstacles are not identical to copyright reservations on openly accessible web pages. A firm seeking a predictive-maintenance tool may need secure data integration and dependable access inside its own organisation more urgently than a larger collection of online prose.

This distinction matters because the overall value of AI in manufacturing could rise even if general-purpose model training becomes more difficult, particularly where companies control the relevant industrial data. Conversely, strong general-purpose tools may improve the interfaces and reasoning features of specialised applications. The two layers reinforce each other without being interchangeable. A serious competitiveness agenda should therefore address industrial interoperability, confidentiality, cybersecurity and workers’ training alongside text and data mining. Narrowing every obstacle to the copyright question would leave important sources of lost productivity unaddressed.

Research needs clarity but not a false exemption

Scientific discovery offers another route through which AI might generate economic value. Models can help researchers identify patterns, search literature and organise complex information, while laboratories assess results through conventional methods. The ifo calculation allocates €25 billion of its modeled downside to reduced research potential. That makes it important to distinguish the special framework for scientific text and data mining from commercial data access under Article 4. Removing this distinction would exaggerate how directly one policy option applies to every research activity.

The research exception cannot, however, resolve all scientific data problems. Institutions still face issues of confidentiality, data stewardship, international collaboration, reproducibility and the cost of computing. Publishers and universities may disagree over particular uses or licences. And promising AI-assisted research still needs independent validation before becoming reliable knowledge or a marketable product. Policy should protect legitimate scientific mining and the creation of authoritative research collections while asking for evidence on where existing law is actually preventing useful work. That is a more precise question than whether regulation is simply holding back science.

The labour-market effects cannot be ignored

The productivity narrative also intersects with employment and pay. In an August 12 ifo survey of companies already using AI, many respondents expected downward pressure on wages over the following five years, especially for less experienced workers in certain occupations. Those answers are expectations, not measured wage reductions; they do not establish how the whole labour market will develop. They nevertheless challenge the assumption that additional economic value will automatically raise the incomes of employees whose work is reorganised around AI.

For governments, the distinction between higher output and shared prosperity is fundamental. A company may adopt a tool to increase capacity, reduce routine work, improve customer service or cut headcount. Workers may benefit from training and better tasks, or confront weaker bargaining positions and the disappearance of entry-level pathways. The ifo data-access study addresses one potential constraint on the supply of powerful models. It cannot by itself answer how those models will alter pay, job quality or the distribution of gains. Public policy must examine that question separately.

The economy already faces other constraints

German businesses operate within a wider setting of energy costs, fiscal policy, investment and demand. The ifo September economic forecast expected a recovery in Germany in 2026 even as higher energy prices weighed on households and parts of the economy. Regulatory arrangements for AI data will matter over a longer horizon, but they do not replace the familiar drivers of business cycles. A firm facing weak orders or high electricity bills may delay a digital investment despite generous access to data or technically impressive tools.

It would be equally misleading to dismiss AI on the grounds that it cannot solve everything. General-purpose technologies often spread through thousands of incremental changes, many of which are invisible in their earliest stages. A realistic assessment asks how productive these investments become once firms reorganise around them and whether competing uses of scarce capital yield a better return. The policy challenge is to remove avoidable barriers while accepting that technology and macroeconomic performance cannot be read from the same single variable.

A credible licensing market is an economic policy

Parliament’s March recommendations envision voluntary collective licences capable of serving individual creators and small businesses, not just large publishers negotiating with global AI companies. That model could reduce transaction costs if rights are clear, prices are predictable and participants can check whether their works were used. The difficulty is implementation. A licence must establish which rights are covered, for which uses, how revenue is allocated, and what happens when the relevant information is unavailable. Poorly designed schemes can merely replace one uncertainty with another.

A sustainable market also needs a way to recognise that different content has different values. Highly specialised reference material, original investigations, local-language archives and creative catalogues may be expensive to produce and hard to replace. A flat transaction structure might not compensate them appropriately. Yet a regime in which every tiny use requires an individual negotiation could make legitimate experimentation impractical. Mechanisms that combine standard technical permissions with a range of voluntary commercial arrangements offer one route forward, but their effects should be tested rather than assumed.

European competition extends beyond the Atlantic

The competitive backdrop stretches from American model companies to Asian semiconductor suppliers and globally distributed cloud infrastructure. Europe risks dependence if domestic businesses cannot acquire or build capable technology at sustainable prices. But creating a local champion is not automatically the same as creating a contestable market. A permissive data regime might benefit foreign incumbents as much as European entrants if the former have superior computing capacity, distribution and customer relationships. The identity of who can exploit an exception matters as much as the formal neutrality of the wording.

Competition authorities and legislators must consider whether compliance requirements reinforce market concentration, whether public datasets are accessible on fair terms, and whether smaller firms can switch providers without losing operational control. The same scrutiny applies to licensing intermediaries that could become essential gateways to copyrighted material. Europe may need open standards and multiple business models more than a single declaration for or against training data access. A rule can be formally equal yet economically unequal when the capacity to comply differs dramatically among market participants.

What evidence should be collected next

An evidence-led evaluation would compare several real regulatory designs rather than one extreme of unrestricted access with another extreme of prohibitive restrictions. Policymakers could measure the cost of respecting machine-readable reservations, the availability and price of licensed collections, and the performance differences between models trained on alternative datasets. They could also look at how much compliance costs vary by company size and whether publication and research output changes after agreements are reached. These are observable outcomes, even when perfect causal identification remains difficult.

The economic gains should be examined with the same discipline. Companies may say they use AI while failing to achieve lasting productivity improvements. Independent case studies can distinguish time saved from additional value created and assess costs of errors, verification and security. Measurements should also extend to downstream effects on media revenue, creative employment and investment in European cultural production. A policy that increases model output in the short term but weakens the supply of trustworthy training material could be self-defeating over time.

The next decision is about balance, not a single figure

The October 13 ifo study provides a striking estimate at a moment when Europe is deciding how its copyright framework should serve a rapidly expanding AI economy. Its central warning is plausible in direction: access to diverse material can influence model capability, investment choices and economic adoption. Its financial magnitude is a modelled conditional outcome whose assumptions deserve scrutiny. The €155 billion should prompt policymakers to test proposals carefully, not settle the debate before creators, publishers and smaller competitors have been heard.

Europe does not have to choose between an AI industry with no workable access to data and a creative economy without enforceable rights. It has to construct a system where research can proceed, commercial developers can obtain lawful inputs, and those who create valuable work retain meaningful choices and opportunities for remuneration. The ultimate measure will not be which side wins a headline estimate. It will be whether the rules support competitive technology and continued production of the knowledge, journalism, culture and expertise that make the technology useful in the first place.