Limited model approach: the merits of methodological rigor in the European legal order concerning AI developments

Bruno Saraiva [master’s student in European Union Law and Digital Citizenship & Technological Sustainability (CitDig) scholarship holder])

Why the EU’s approach to AI development differs from that of the U.S. or China is a question that spans philosophy, sociology, geopolitics, and economics. But the simplest answer may be the following: they are different. Each polity carries distinct priorities, institutions, and constraints – and these differences translate into divergent AI trajectories.

In Europe, this divergence goes beyond regulation and economics; it extends to the very technical models being developed. While the U.S. and China pursue scale through ever-larger general-purpose systems, the EU has signaled a regulatory preference for limited models – special-purpose systems trained with curated data.

This post explores the methodological virtues of that approach. In a world where large models struggle with trust, reliability, and compliance with rights-based law, the EU’s strategy offers an alternative: models designed to minimise hallucinations, resist “model collapse”, and reduce opacity. By embedding rigor into training practices, the EU may not only advance trustworthy AI but also begin addressing its competitiveness woes, as underscored by the Draghi Report.[1]

Technical rationale for limited models

Before we reach that stage, however, we will address the EU’s focus on training limited AI models. Though reporting is still incipient, empirical research suggests that further general model development (i.e., training done with as much data as possible, eventually relying on internet scouring or large undifferentiated databases) has not produced more useful models.^[2]

Empirically, large language models can produce seemingly apparently real but false outputs (“hallucinations”), a reliability problem made worse when models are trained on their own synthetic content. A growing body of work shows that recursive training on AI generated data degrades models’ grasp of the true data distribution (hopefully one that closely mirrors reality) – a phenomenon termed model collapse^[3] – which is expected to increase factual errors unless countered with careful data curation and provenance filtering. In this context, talk of an interpretability deficit is apt:^[4] many high performing systems remain opaque,^[5] leaving users and developers without reliable insight into why an output is produced – this situation closely ties to “black-box” design and is known to encourage a circular over-reliance on synthetic data – an information death spiral.

This happens because large language models (LLMs, the systems most commonly referred to as “AI”) operate by learning to estimate “the conditional probability of a word given previous words”,^[6] producing outputs by sampling from a mathematically based learned distribution. As a result, their answers are approximations of patterns present in the data they were trained with, rather than the result of direct access to facts. Since this data is itself a set of human-filtered representations (be it in writing or in image form; be it tempered by satire, slander, error, or sincere but mistaken testimony), distinguishing truth from fiction becomes a non-trivial inference – much as it occurs in human experience. Understanding this is key to explain why LLMs can produce convincingly real but inaccurate content and why reliability is dependent on what data they “see” and how they are trained.^[7]

For highly specific and sensitive case-uses (as governance generally implies), general models might offer reasonably accurate outputs by providing additional elements, but guaranteeing reliable results by modeling alone becomes fraught.

This risk is compounded when models are trained and create outputs based on previous AI outputs (the “synthetic data” already alluded to). Empirical and theoretical results have proven that this recursive self-training causes model collapse: as the “tail ends” of mathematical distributions vanish, successive data generation leads to misperceived reality, polluting future training sets.^[8]

It is therefore nearly inevitable that general purpose models that pragmatically had to take a brute force approach (i.e., source data from the open internet), will eventually experience a deterioration of their capabilities.^[9] Their queries becoming less useful and their answers more prone to “hallucination” due to recursive training. A concern that becomes only more acute when one realizes the likelihood of training data including personal information under the GDPR’s definition. Because flawed quasi-realities generated by AI often possess high verisimilitude, they can trigger deeper grievance than outright falsified content. This is illustrated by what researchers have deemed the AI trust paradox, where plausible-sounding – but incorrect – AI outputs are more misleading due to their realism, something that lies at the heart of much feared deepfake techniques.^[10]

Though now self-evident, it bears reminding that LLMs are models of reality, not reality itself. Their outputs reflect compound approximations. This is why the EU’s approach proves valuable; when well executed, limited model approaches are built-in mitigation protocols. By curating data and producing provenance-aware datasets using trusted sources (as we hope will become standard practice), and the development of training and evaluation protocols that target factuality rather than fluency has the potential for a practical competitive edge in AI.

Taken together, these findings allow us to argue for a “last mile” of AI development that focuses less on brute-force data scale and more on meticulous dataset design (including filtering out AI-generated content), documentation, and evaluation protocols. Though obviously more expensive and human intensive, this is a check that the EU’s highly educated work force is well poised to cash.

Legal and institutional alignment

The EU’s approach is therefore methodologically salutary. The approach signaled by the Commission with its AI Factories initiative, the restrictions on the creation of general AI models after the 2^nd of August, 2025 (Article 113 of the AI Act), the creation of the AI Office and creating of European consortiums for the use of AI tools in governance have a not-so-subtle dirigiste approach to them. On the one hand these initiatives limit the creation of general-purpose AI models (and establish iterative compliance requisites for GPAI and high-risk model operators already in market), while on the other encouraging the creation of limited purpose and limited-trained models as an alternative.

Though again, experience in these subjects is lacking, the expectation is that the careful curation of data will limit the previously expressed shortcomings. Indeed, we would argue that the next step of AI development will not involve the brute force approach previously necessary for the creation of foundation code, but the careful selection of training data which will educate the LLMs in question. This “last step” is neither cheap nor trifling, it involves the creation of specific guidelines and the selection of training data appropriate for the intended context – an endeavor that requires collective support, pooling of expertise, and collaborative effort; the kind of endeavor that the European Union structurally excels at.

What is decisive here is that the determination of what constitutes an appropriate, limited dataset is not a purely technical or computational issue, but a fundamentally human one. Decisions regarding inclusion and exclusion, annotation, and representativeness carry normative weight: they embed judgments about fairness, proportionality, and relevance. In the EU context, this process must be measured against the binding framework of Article 22 GDPR, which restricts reliance on automated decision-making absent meaningful human oversight, and Article 5 of the AI Act, which explicitly prohibits practices deemed harmful to human dignity or fundamental rights. The very definition of a “limited dataset” thus becomes a safeguard mechanism: by curating data within human-defined boundaries, the EU legal order attempts to ensure that high-risk AI applications do not reproduce structural discrimination, undermine democratic participation, or erode individual autonomy.^[11]

Furthermore, not only does this development structure imply significant institutional oversight inherent in its foundation and funding apparatus, but the very case uses seem to suggest added oversight due to the sensitivity inherent in data relevant to governance. In practice, this means that the curation of datasets destined for applications in areas such as justice, social services, or public administration cannot be delegated solely to private actors or automated processes. Rather, they must be subject to layered accountability – combining Commission guidance, supervisory authority review, and judicial enforceability under EU law.^[12] Within the Single Digital Market, this integrated oversight ensures that dataset construction does not become a point of regulatory evasion, but instead reflects the harmonized standards of transparency, accuracy, and non-discrimination that the Union seeks to guarantee.

Thus, the fine human nature of dataset limitation is not an incidental feature of AI governance but a constitutive element of the EU’s regulatory strategy. Unlike the brute-force accumulation of data characteristic of early machine learning, the European model requires interpretive communities – lawyers, regulators, ethicists, and technical experts^[13] – to collaborate in shaping the very fabric upon which AI is weaved: data. This represents the extension of integrated digital constitutionalism to the epistemic foundations of AI.

This forward-looking approach is likely to strengthen the EU’s position in the competitive global market by promoting a more carefully structured model – one that carries the added advantage of being adaptable to all 27 Member States’s languages. This will hopefully combat any one language’s supremacy in AI aided creation, furthering the goals constant in Article 3 of the Treaty on European Union and Article 21 of the Charter of Fundamental Rights of the European Union.^[14]

Socioeconomic implications – bridging the unseen gap

Nonetheless, we consider this approach to be constructivist, in the sense that it seeks to embed AI into currently existing systems – whether educational or productive – keeping the value of digital interaction largely unchanged. This expectation of continuity, however, is less than certain. The expected results of stronger online moderation, aimed at creating a “cleaner” digital scape^[15] are likely to alter not only the way users interact with the digital realm but also fundamentally reshape how they perceive worth. Indeed, merging empirical research suggests consumers perceive AI-generated content negatively, something that affects both digital and physical product valuations.^[16]

Despite this, one should not throw in the towel and assume AI’s contributions to the Single Digital Market are inherently negative. A research brief examined the impact of AI-generated product summaries and arrived at the conclusion that summaries can speed up consumer decision-making and reduce hesitation.^[17] Other real-world research suggests generative AI in marketplaces lowers prices and boosts volume. This study conducted in China’s anime art market is particularly promising in the sense that not only is it a use-case regarding cultural content, but results show that most of the boosted demand was for lower-end personalised orders (though existing creators held market share).^[18]

This suggests that while AI can lower price points and alter market dynamics, it is not necessarily devaluing physically created art – rather it disrupts pricing structures in digital marketplaces. This is because it is unlikely that the change in the value of digital goods and services will be limited to itself, with a distorting effect being felt even on the value of physical goods, services, and experiences. If properly articulated, the potential for increasing the value of art and digital content will aid Europe’s quest for dematerialisation and full circular economy cycles.  We hold that this is the true disruptive potential of AI in the digital revolution, providing the communicational missing link that a green transition requires.^[19]

Though the already mentioned initiatives have enormous potential as regards the technical development of AI technology in the EU, they do not present a fully articulated case use scenario to meaningfully achieve the vast goals proposed.

One of the great advantages of large language models lies, predictably, in their inhumanity: the ability to detect patterns that human consciousness often cannot. This limitation arises from the constraints of data input – our minds can only process so much before perception becomes blurred by emotion or disposition – or from habituation to the “normality” of recurring processes. Just as humans rarely perceive the biotic cycles that sustain life, we also struggle to discern economic patterns, particularly in ways that disproportionally disadvantage administratively marginalised regions. What is “out of sight” often becomes “out of mind” – economically, politically and institutionally – an invisibility that correlates strongly with non-urban or non-capital areas.

The possibility of identifying hitherto unquantified productive activity therefore holds significant value.[20] Portugal’s Alto Minho region offers a telling example. A prime example of what has been termed “deep periphery”, these regions exhibit persistently low GDP per capita (in purchasing power standards), limited growth in productivity, and stagnation despite broader national trends.^[21] Despite this, it nonetheless sustains a dense fabric of subsistence farming practices and interdependent cultural and productive systems. These activities remain only partially captured by official statistics, owing to a deliberate reliance on parallel economy, permissive neglect by central governance, and pragmatic taxation limits – since subsistence generates and maintains life, not monetary wealth.^[22]

Though this case is modest compared to broader critiques of GDP calculations, this case shows the potential for redress and integration within the transparent economic cycle sought by the EU. Quantifying such activity would both strengthen local economies and offer a fuller account of contributions within the Single Market – digital or otherwise.

Any political project that grounds legitimacy in results must demonstrate them. Without adequate quantification, outcomes remain elusive. This absence of accounting does not automatically signal economic backwardness, but it does reflect a loop of invisibility at odds with the Union’s circular economic ambitions and push for deeper integration.

The EU’s limited model approach is therefore more than a technical choice – it is a normative project. By embedding curated models into education and governance, the Union strengthens not only competitiveness but also public trust in the digital tools shaping daily life. This strategy goes beyond regulation, weaving fairness, human rights, and sustainability into the very foundations of AI. In doing so, the EU advances digital constitutionalism, aligning technological progress with democratic values and collective responsibility.

Finally, regarding user value: top-down, business-focused developments cannot resolve citizen distrust of digital systems. Expanding digital production means little if confidence in digital goods declines,^[23] risking not sustainable growth, but speculative bubbles. Yet failing to act risks market irrelevance and eroding public distrust – the EU’s twin perils.

[1] European Commission, ed., The future of European competitiveness: part A: a competitiveness strategy for Europe (Publications Office, 2025), 23–27, https://doi.org/10.2872/9356120.

[2] As one AI researcher succinctly put it, “the largest models were generally the least truthful.” See Stephanie Lin et al., “TruthfulQA: measuring how models mimic human falsehoods,” version 2, preprint, arXiv, 2021, https://doi.org/10.48550/ARXIV.2109.07958.

[3] Ilia Shumailov et al., “AI models collapse when trained on recursively generated data,” Nature 631, no. 8022 (2024): 755–57, https://doi.org/10.1038/s41586-024-07566-y.

[4] We prefer this term to “hallucination”, as it reframes the issue to on methodical and technical grounds. For a closer analysis of the issue and techniques for overcoming it, see Catherine B. Hurley et al., “interactive slice visualization for exploring machine learning models,” version 2, preprint, arXiv, 2021, https://doi.org/10.48550/ARXIV.2101.06986.

[5] See, Lin et al., “TruthfulQA.”

[6] Daniel Jurafsky and James H. Martin, Speech and language processing: an introduction to natural language processing, computational linguistics, and speech recognition with language models, third edition draft (Pearson, 2025), 39.

[7] For an explanation of the very processes that determine that dependence and its “autoregressiveness”, see Tom B. Brown et al., “Language models are few-shot learners,” version 4, preprint, arXiv, 2020, https://doi.org/10.48550/ARXIV.2005.14165.

[8] For an in-depth view of this purportedly rare phenomena, see Shumailov et al., “AI models collapse when trained on recursively generated data.”

[9] To that end, the EU’s AI Act’s Article 51 (in tandem with Article 53, 54 and 56) establishes guidelines and a general-purpose AI code of practice that came into effect in August 2, 2025. This choice limits the creation of new general-purpose models by demanding more stringent reporting and supervision of models “trained with over 10^23 FLOP and capable of generating language.”; European Commission, “EU Rules on general-purpose AI models start to apply, bringing more transparency, safety and accountability | Shaping Europe’s digital future,” Press Release, 1 August 2025, accessed September 3, 2025, https://digital-strategy.ec.europa.eu/en/news/eu-rules-general-purpose-ai-models-start-apply-bringing-more-transparency-safety-and-accountability. For the guidelines, see European Commission, “Guidelines on the scope of obligations for providers of general-purpose AI models under the AI Act | Shaping Europe’s digital future,” Policy and Legislation, 18 July 2025, accessed September 3, 2025, https://digital-strategy.ec.europa.eu/en/library/guidelines-scope-obligations-providers-general-purpose-ai-models-under-ai-act.

[10] For a technical oversight of these issues, see Trisha Ray, “The paradox of innovation and trust in artificial intelligence,” Observer Research Foundation, 22 February 2024, https://www.orfonline.org/expert-speak/the-paradox-of-innovation-and-trust-in-artificial-intelligence; Roger Vergauwen and Rodrigo González, “On the verisimilitude of artificial intelligence,” Logique Et Analyse 190, no. 189 (2005): 323–50; Federica Lago et al., “More real than real: a study on human visual perception of synthetic faces”, version 2, arXiv, 2021, https://doi.org/10.48550/ARXIV.2106.07226.

[11] Judgment CJEU Nowak, 20 December 2017, C-434/16, ECLI:EU:C:2017:582 establishes a particularly broad interpretation of “personal data”, emphasising that even exam answers qualify. This is demonstrative of how inclusive the EU’s notion of relevant datasets is, placing emphasis on the human element in determining data scope.

[12] Judgment CJEU Schrems II, 16 July 2020, C-311/18, ECLI:EU:C:2019:1145 ruled that under Articles 7 and 8 of the Charter of Fundamental Rights of the European Union strict safeguards as regards the transferring and processing of personal data are required. Under this interpretation even the advantages granted by limited datasets do not exempt fundamental rights safeguards, having to be assured from the ground up.

[13] Judgment CJEU Glawischnig-Piesczek v. Facebook, 3 October 2019, C-18/18, ECLI:EU:C:2019:458 held that platforms can be required to remove identical or equivalent illegal content globally. Useful to tie dataset curation and oversight with DSA/DMA obligations in the Single Digital Market.

[14] For a technology-based oversight on this hot button subject see Gábor Bella et al., “towards bridging the digital language divide,” version 1, preprint, arXiv, 2023, https://doi.org/10.48550/ARXIV.2307.13405, and Paula Helm et al., “Diversity and language technology: how language modeling bias causes epistemic injustice,” Ethics and Information Technology 26, no. 1 (2024): 8, https://doi.org/10.1007/s10676-023-09742-6.

[15] Judgment CJEU Glawischnig-Piesczek v. Facebook, as cited before, creates a legal obligation for online moderation and “cleaner” digital spaces by demanding platforms to remove infringing content. Judgment CJEU Republic of Poland v. European Parliament and Council of the European Union, 26 April 2022, C-401/19, ECLI:EU:C:2022:297 also briefly touches on this subject as it weighs moderation versus freedom of expression, something that irremediably impacts user interaction.

[16] See Gedas Kučinskas, “Negative effects of revealing AI involvement in products: mediation by authenticity and risk, moderation by trust in AI and familiarity with AI,” preprint, 2024, https://doi.org/10.2139/ssrn.4805186.

[17] Tung X. Bui, ed., Proceedings of the 58th Hawaii International Conference on System Sciences: Hilton Waikoloa Village, January 7-10, 2025 (Department of IT Management, Shidler College of Business, University of Hawaii, 2025).

[18] See Kaichen Zhang et al., “The impact of generative artificial intelligence on market equilibrium: evidence from a natural experiment,” version 2, preprint, arXiv, 2023, https://doi.org/10.48550/ARXIV.2311.07071.

[19] To that end, see Nicholas Stern et al., “Green and intelligent: the role of AI in the climate transition,” Npj Climate Action 4, no. 1 (2025): 56, https://doi.org/10.1038/s44168-025-00252-3.

[20] As signaled by European Commission, ed., Science, research and innovation performance of the EU: a competitive Europe for a sustainable future (Publications Office, 2024), 283, https://doi.org/10.2777/965670.

[21] For a telling analysis of so called “bioeconomies” and their systemic invisibility see Lilian Pungas, “Invisible (bio)economies: a framework to assess the ‘blind spots’ of dominant bioeconomy models,” Sustainability Science 18, no. 2 (2023): 689–706, https://doi.org/10.1007/s11625-023-01292-6.

[22] Economic studies of the EU‑15 find that in agriculture about 20% of Gross Value Added (GVA) arises from unrecorded or informal economic activity, including subsistence work, family labor, and seasonal off-the-books contributions. Notably, Portugal is among those Member States with the highest shares of informal agricultural activity, indicating that significant portions of productive activity remain invisible to formal accounting systems. Friedrich G. Schneider et al., “Measuring the immeasurable: the evolution of the size of informal economy in the agricultural sector in the EU-15 up to 2019,” SSRN Electronic Journal, ahead of print, 2021, 23–25, https://doi.org/10.2139/ssrn.3805463.

[23] Kučinskas, “Negative effects of revealing AI involvement in products.”

Picture credit: by Google DeepMind on pexels.com.