<前の日記(2025年11月20日) 最新 編集

高木浩光@自宅の日記

目次 はじめに 連絡先:blog@takagi-hiromitsu.jp
訪問者数 本日: 101   昨日: 6103

2025年12月06日

Digital Omnibus, Scope and Relevance: Are We Fighting the Wrong Battle?

In a previous post on the leaked “Digital Omnibus” draft, I played with the idea hidden in noyb’s name: None of Your Business – in other words, irrelevant. My point was that the GDPR already contains, in plain sight, a principle that should be at the core of the debate but is almost never mentioned: Article 5(1)(c)’s demand that data be adequate, relevant and limited to what is necessary in relation to a purpose. The real question is not how much “personal data” the law can cover in the abstract, but which data are any controller entitled to take into account for a given decision.

Since then, the Digital Omnibus proposal has been published, and noyb has issued a long report criticising it. The public debate has largely been framed as a familiar clash: on one side, a Commission accused of “watering down” the GDPR by narrowing the definition of personal data and creating new exceptions for AI and research; on the other, civil-society actors defending a maximal scope for “personal data” and a strong version of data minimisation.

In this post I want to argue that this is the wrong axis of conflict. Both sides are fighting over how broad the law’s coverage should be, while largely ignoring what the law was actually designed to control. The Omnibus debate is being treated as a war over definitions, when it should be treated as an opportunity to recover the original focus of data protection: the fairness of decisions about individuals, and the prevention of their being made on the basis of irrelevant data.

To make that case, I will first say something about the “maximal scope” paradigm that dominates current criticism of the Omnibus. I will then turn to the research and AI provisions, explain why both the Commission and noyb mis-locate the risks, and finally suggest a different way of looking at scope – one that re-reads “processing of personal data” in line with its directive-era function.

1. The “Maximal Scope” Paradigm and Its Limits

1.1 How the Omnibus Debate Frames Article 4(1)

A central target of current criticism of the Digital Omnibus proposal is the suggested rewording of Article 4(1) GDPR, the definition of “personal data”. In noyb’s reading, the new text would introduce a strongly controller-relative element into identifiability: whether information is “personal data” would depend on the specific controller’s perspective, resources or reasonable efforts. Pseudonymised data held by actors who lack re-identification keys might therefore fall outside the definition for them, even if another actor could identify the same records.

For noyb, this is not just an unfortunate drafting choice but a constitutional problem. The reasoning, roughly, is as follows:

  • the Charter of Fundamental Rights protects the “protection of personal data”;
  • the Court of Justice has repeatedly read “personal data” and “processing” very broadly;
  • therefore, any legislative move that appears to reduce that scope is a direct threat to fundamental rights.

On this view, the definition of “personal data” in Directive 95/46/E.C., as interpreted in landmark CJEU judgments, has been effectively “constitutionalised” by Article 8 of the Charter. And what is “constitutionalised” is understood as the breadth of the definition: the more generously we interpret “any information relating to an identified or identifiable natural person”, the more robust the protection of Article 8 must be.

From this vantage point, the only safe strategy is to insist that almost any information that can, in some chain of events, be linked to an individual must be treated as “personal data”, and that almost any operation involving that information must be treated as “processing”. The Omnibus proposal is thereby cast as an attempt to introduce controller-relative “subjectivity” into what ought to remain an objective, all-encompassing definition – a move that would undermine legal certainty, fragment enforcement and invite a race to the bottom.

I will call this way of thinking the maximal scope paradigm: the belief that the primary way to protect fundamental rights is to push the conceptual reach of “personal data” and “processing” as far as possible and then defend that maximal line against any retreat. Different actors occupy this paradigm in different ways – NGOs like noyb, some authorities and parts of the academic debate – but what they share is the assumption that “more situations legally qualifying as processing of personal data” almost automatically means “more effective protection.”

1.2 The Original Target of Data Protection: Decisions, Not “Data as Such”

Historically, however, the normative focus of European data protection instruments has never been “data as such”. It has been the use of data in systems of decision-making: credit scoring, employment screening, social security adjudication, policing, risk assessment, and so on. The early data protection laws, as well as their international precursors, were motivated by concerns that automated or computer-supported decisions about individuals could be made in opaque ways, based on incomplete, irrelevant or biased information, compiled in powerful, centralised files.

This orientation is visible in two features that are often overlooked in current debates.

First, the concept of processing was originally tied to what one might call an informational pipeline: a systematic sequence of operations through which data are collected, stored, combined, analysed and used to support decisions or actions. It was this pipeline – and the risks it created for individuals subjected to it – that justified a set of procedural and substantive constraints. The inclusion of “collection”, “use”, “disclosure” and “erasure” in the definition was meant to capture the different stages of that pipeline, not to transform any isolated mention of a person’s name into “processing” in the full legal sense.

Secondly, the much-cited triad of “adequate, relevant and not excessive” was framed as a data quality principle for decision-making. Its function was to ensure that decisions affecting individuals are based on information that has a rational link to the purpose of the decision, that is sufficiently complete and accurate for that purpose, and that does not include extraneous or prejudicial details. The idea was not to minimise data in the abstract, but to constrain the informational basis of decisions so that they are fair, non-arbitrary and non-discriminatory.

If one keeps this decision-oriented perspective in view, the relation between scope and protection looks different. What matters is not primarily how many pieces of information fall under the label “personal data”, but which informational pipelines and which decision contexts are brought within the reach of data protection law. A conception of “processing” that indiscriminately covers everything from one-off journalistic mentions to complex, automated profiling systems may be rhetorically expansive, but it risks obscuring precisely those structural features – centralisation, automation, iterative use – that generated the need for a specific legal regime in the first place.

1.3 A Preview: Two Layers of Protection

From this angle, the maximal scope paradigm is problematic in two ways.

It is over-broad, in that it extends the full conceptual and doctrinal apparatus of “processing of personal data” to contexts where the characteristic risks of data protection law are only marginally present. Treating every isolated reference to an individual, every casual photograph, or every transient log entry as “processing” subject to the GDPR may be logically possible under a literal reading of “any information relating to…”, but it stretches the concept far beyond the original risk profile it was designed to address. In doing so, it invites either pervasive non-compliance in everyday life or the proliferation of exceptions and derogations, which in turn make the law more complex and less intelligible.

At the same time, it is under-protective where protection is most urgently needed. By focusing energy on defending the outer conceptual perimeter of “personal data”, it diverts attention from the internal structure of those processing operations that truly matter: large-scale profiling, scoring and decision-support systems. Here, the crucial questions are not whether some fragment of information can be said to “relate” to an individual in the abstract, but whether the overall pipeline of processing is structured in such a way that decisions about individuals are based on adequate and relevant data, subject to meaningful contestation, and constrained by clear prohibitions on certain kinds of inferences.

Put differently, insistence on an ever-expanding notion of “personal data” risks functioning as a substitute for the harder work of identifying, and regulating, those decision contexts that require heightened protection. It can make it politically easier to denounce any perceived narrowing of scope – such as controller-relative identifiability in Article 4(1) – without having to say very much about which kinds of processing the fundamental right to data protection actually targets, and why.

To get out of this dead end, we need to recall that data protection law has always worked on two layers at once:

  • a decision layer, which governs how data are used in concrete decisions about individuals, and where “adequate, relevant and limited to what is necessary” should be understood as a genuine relevance requirement; and
  • a flow layer, which governs how data can move between contexts – purpose limitation, restrictions on re-use and onward disclosure – in order to prevent future decisions from being made on data that have no rational link to their purpose.

The rest of this post will unpack this two-layer structure and show how it helps to reframe the Omnibus debates about scientific research, AI and personal data definitions.

2. Scientific Research and AI: Data Minimisation without Decision Context

2.1 How the Omnibus and noyb Frame Scientific Research

A second focal point of the Omnibus debate is the proposed new definition of “scientific research” in Article 4(38) and the accompanying changes to purpose limitation and information duties, notably Article 13(5). The Commission’s draft defines scientific research, in essence, as “any research” that “can also support innovation”, supplemented by a long list of criteria referring to growth of society’s “knowledge and wellbeing”, adherence to “ethical standards in the relevant research area”, and similar elements.

noyb’s report offers a sharp and, in many respects, justified critique of this approach. It points out that:

  • the core term “research” is left so broad that it could cover almost anything, from “sifting through a library” or “using Google Search” to watching rats in a laboratory;
  • the phrase “any research which can also support innovation” extends the scope to mere possibilities (“can”) and by-products (“also”), so that virtually any data-driven activity in the name of innovation could qualify;
  • the requirement that research must be able to “support innovation” risks excluding areas of academic work not aimed at innovation in the economic or technological sense; and
  • references to “growth of society’s general knowledge and wellbeing” and to compliance with “ethical standards” are so indeterminate that they effectively outsource the definition of a Charter-limiting concept to private or sectoral ethics codes.

On this basis, noyb argues that the proposed definition, combined with an expanded framework for research exemptions, amounts to a “blanket allowance of further processing for a broadly and ill-defined ‘research purpose’”, which would likely fail the Charter’s requirements of necessity and proportionality. The Commission appears, in this reading, to be sacrificing the Article 7 and 8 rights on the altar of an amorphous “innovation” agenda.

At first glance, this critique appears squarely aligned with a rights-protective stance. Yet its internal logic again rests on two familiar premises: first, that the core of data protection is data minimisation as a direct emanation of Articles 7 and 8 and Article 52(1), and second, that the main threat posed by an overbroad research definition is the erosion of this minimisation and of purpose limitation as such. The questions of which types of research, and which decision contexts, ought to fall under stricter scrutiny remain largely implicit.

2.2 Research without Decisions vs Research Feeding Decisions

A decision-oriented perspective – one that takes seriously the role of adequacy and relevance in decision-making – suggests a different way to cut the problem.

On the one hand, there are forms of research that do not themselves feed into concrete decisions about identifiable individuals. Classic statistical and scientific uses of data, where the outputs are aggregate patterns, coefficients or models describing populations, may raise important issues of ethics, distributive justice and epistemic bias. But in terms of data protection, their risk profile is largely mediated by what happens after research: whether and how the resulting insights are embedded into operational decision systems.

On the other hand, there are forms of research that are structurally wired into treatment decisions: credit risk models developed on customer databases, hiring algorithms trained on past applicants, insurance pricing models, fraud detection tools. Here, the line between “research” and “operation” is thin. The same structured datasets used to develop the models are used, or readily usable, to support consequential decisions about individuals.

In this second family, what matters for data protection is not that “research” in the abstract uses more data than strictly necessary, but that the resulting decision systems may operate on data that are not relevant to the decision purpose. The core risk is arbitrary or discriminatory treatment: music taste influencing credit limits, social media activity affecting hiring, neighbourhood proxies standing in for ethnicity in insurance pricing. These are precisely the kinds of problems the adequacy and relevance requirement was originally meant to prevent.

From this angle, the Omnibus proposal’s changes to Article 4(38) and 13(5) are problematic, but not primarily because they offend an abstract ideal of data minimisation. They are problematic because they:

  • blur the distinction between research that merely generates statistical insight and research that is structurally embedded in decision pipelines; and
  • loosen purpose limitation and transparency duties in a way that weakens what I will later call the flow layer: the set of rules that restrict how data can move between contexts in order to prevent their use in irrelevant or discriminatory decisions.

noyb’s critique rightly attacks the vagueness of the definition and the risk of “innovation” being used as a carte blanche for further processing. But by casting the issue mainly in terms of Charter-compliant data minimisation, it misses the opportunity to ask the decisive question: which research-generated datasets are likely to flow into decision contexts, and how should their re-use be constrained to prevent irrelevant or discriminatory decision-making? This is the perspective I developed in my earlier post on the Omnibus debate, where I argued that both the Commission and noyb tend to overlook the GDPR’s explicit relevance requirement in Article 5(1)(c).

2.3 AI and Special Categories: Two Activities Confused under Article 9(2)(k) and (5)

The same pattern recurs, in sharper form, in the Omnibus proposal for Article 9(2)(k) and (5), dealing with AI and special categories of data. The draft introduces a new derogation for “processing in the context of the development and operation of an AI system … or an AI model”, accompanied by a paragraph 5 that requires controllers to implement “appropriate organisational and technical measures” to avoid the processing of special categories in training, testing and validation, and to remove such data where identified, or at least to prevent them from influencing outputs or being disclosed if removal would require disproportionate effort. The proposed recital 33 expressly frames this as a derogation from the Article 9(1) prohibition “in order not to disproportionately hinder the development and operation of AI”.

noyb criticises this structure on familiar grounds. It argues that AI is precisely the kind of opaque, large-scale processing that the data protection principles were originally designed to constrain; carving out a new derogation for AI development and operation is therefore backward-looking and threatens to hollow out the very safeguards that were written with such systems in mind. It also points out the loose drafting (“in the context of”, “appropriate measures”, “disproportionate effort”) and the risk that the derogation becomes a general permission for AI developers to retain special categories so long as they can claim to have implemented some minimisation efforts.

Again, this line of attack is understandable, but it presupposes that “AI development” is a unitary activity and that the relevant Charter risk can be captured by the volume and sensitivity of data processed. The actual situation is more complex.

As I argued in that earlier blog post, the label “AI development” currently covers two fundamentally different activities:

  • Treatment AI, such as credit scoring, hiring and insurance models, trains on genuine systematic datasets: customer or candidate databases that are prototypical “personal data” in the sense of data protection law. Here, database-oriented requirements make structural sense: minimise unnecessary special categories in training sets, remove them where inappropriately included, protect the sensitive training data against disclosure or misuse.

  • Generative AI training on public web text, by contrast, operates on scattered fragments of already public information to extract statistical patterns. It does not create systematic individual records; the resulting models do not, in normal operation, contain addressable records of specific persons’ sensitive characteristics. The main risks lie not in disclosure of special categories about identified individuals, but in the amplification of biases and stereotypes in generated content.

For the second family, Article 9(5)’s requirements are simply mis-specified. Instructing controllers to “remove” special categories from a corpus of heterogeneous web pages assumes a database-like structure that does not exist: what exactly is to be removed – sentences, paragraphs, entire documents? The “where possible” qualifier acknowledges the impracticality without providing a coherent standard. The language about preventing “disclosure” of special categories presupposes a context in which disclosure means revealing specific individuals’ sensitive data from a controlled database, not producing new text in which health or religion are mentioned in the abstract.

By lumping these structurally different activities together under a single derogation, the Omnibus proposal repeats the conceptual conflation already visible in the personal data definition. At the same time, and more importantly from a decision-oriented viewpoint, it misplaces regulatory attention:

  • Generative AI training on scattered public text, which normatively should not be treated as personal data processing at all because it never constructs systematic datasets suitable for decisions about individuals, becomes the direct target of impractical, database-oriented requirements.

  • Treatment AI operation, where structured datasets are used to make consequential decisions about individuals, receives relatively little targeted attention. Article 88c recognises AI operation as a legitimate interest, but provides no mechanism to enforce the relevance requirement of Article 5(1)(c) in this context. Irrelevant characteristics such as music preferences, social media activity or religious affiliation may thus be factored into decisions under the banner of “legitimate interest”, subject only to an individual right to object.

In short, both the Omnibus proposal and much of its criticism, including noyb’s, focus on how much and what kind of data AI systems may process, while leaving largely untouched the question whether the data used in AI-supported decisions are relevant to the decision purpose and how to prevent cross-context re-use of data in ways that generate arbitrary or discriminatory outcomes. That is precisely the gap that a relevance-based reading of Article 5(1)(c) is meant to fill.

2.4 So What Should Happen to the Research and AI Exceptions?

At this point the obvious question is: if both the Commission and noyb are framing the problem in the wrong way, what would a better treatment of research and AI actually look like? It seems clear that one political driver of both the new research definition and the AI derogation in Article 9(2)(k) – and of the attempt to redefine personal data – is the fear that, under the current reading of “processing of personal data”, even web-scale AI training and other large-scale analytics are automatically sucked into the full GDPR regime.

My answer is essentially the same as in my earlier post on the leaked Omnibus draft, but extended to the new definition and AI proposals.

  1. Keep the existing statistical/research framework for the learning phase.
    Europe already has a coherent way to deal with large-scale analysis that does not itself feed into decisions about individuals: the combination of

  2. the “compatibility” rule for statistical and scientific purposes in Article 5(1)(b) GDPR; and

  3. the older Council of Europe framework on statistical processing, which allows restrictions of individual rights as long as no decisions or measures concerning individuals are taken on the basis of the statistical results.

AI training belongs here. Training a model on large datasets is, in substance, statistical processing. In this phase there is, by definition, no concrete decision purpose yet, so it makes little sense to apply a relevance test at the level of the model or its inputs. The adequacy and relevance requirement is a constraint on the informational basis of decisions, not on the structure of models in the abstract. The core protected interest is not offended simply because a learning algorithm has “seen” certain variables; it is offended when those variables are actually used as part of a decision pipeline about individuals.

  1. Draw a hard line at the operational, decision-making phase.
    Once a trained model is used to support credit decisions, hiring decisions, insurance pricing, fraud detection and so on, we are no longer in the realm of “mere statistics”. We are in the realm of treatment AI, where the adequacy and relevance requirement must bite: every category of data that actually feeds into the decision pipeline must have a rational link to the decision purpose.

This distinction is crucial. A model may internally contain parameters influenced by many features, including features that would be irrelevant if used explicitly at decision time. What matters for data protection is which data are in fact taken into account when deciding about a person – which variables are read, combined and acted upon in the operational system. The legal architecture should therefore be very simple:

  • learning / training = statistical or research processing, covered by existing rules for such purposes, on the strict condition that the outputs are not used directly to decide about individuals;

  • operation = decision-making, subject to full data-quality scrutiny, including a serious enforcement of “adequate, relevant and not excessive” under Article 5(1)(c) with respect to the inputs actually used in the decision.

  • No special AI derogation is needed – and a broad one would be harmful.
    If we treat AI training as statistical processing and AI operation as ordinary (but highly consequential) decision-making, there is no need for a new, AI-specific derogation in Article 9(2)(k) or elsewhere. On the contrary, a blanket derogation “for AI development and operation” risks doing exactly what I criticised in my previous post: it invites controllers to invoke “AI” as a magic keyword to justify both phases at once, while never asking the simple question at the operational phase: are the data we actually use in these AI-supported decisions relevant to the decision purpose?

  • Special categories do not need an AI-only regime either.
    For the learning phase, the right question is not “how can we purge every trace of special categories from web-scale training corpora?”, but “how do we make sure that training corpora used for statistical purposes are not casually re-purposed into decision systems where special categories – direct or inferred – would be irrelevant to the decision purpose?”.

For the operation phase, if relevance is properly enforced, the distinction between directly revealed and inferred sensitive data becomes less important: both are ruled out wherever they are “none of your business” in light of the decision purpose. The problem, as I argued in the earlier post, is that we are currently trying to slim down Article 9 without first building the enforcement muscle around relevance at the decision layer.

In other words, instead of inventing AI-specific exceptions and research-specific shortcuts, we should put AI training back where it belongs – under the existing statistical/research framework – and concentrate regulatory energy on AI operation, where the relevance requirement finally has to be taken seriously at the point where decisions are made.

3. Decision Governance and Flow Governance: Re-reading “Processing of Personal Data”

The discussion so far already suggests that data protection operates on two interlocking layers:

  • decision governance, where the law constrains how data are used to make or support decisions about individuals; and
  • flow governance, where the law constrains how data can move between contexts, in order to prevent future misuse in irrelevant or discriminatory decisions.

We can now make this structure more explicit and ask how the Omnibus proposals affect each layer.

3.1 Two Complementary Layers of Protection

From a functional perspective, EU data protection law operates on two distinct but interdependent layers.

First, there is direct governance of decisions. Whenever personal data are actually used to make or support decisions about individuals, data protection law requires that the informational basis of those decisions meets the traditional data-quality standard: the data must be adequate and relevant to the purpose of the decision and not excessive. In modern language, there must be a rational link between the data items used and the legitimate decision purpose. This is where the classic “adequate, relevant and not excessive” formula belongs: it is a constraint on how decisions are made, not an abstract call to minimise data.

Second, there is preventive governance of data flows. Purpose limitation, restrictions on further use and onward disclosure, and the construction of “data silos” are not mysterious relics of a bygone era of database regulation. They are ex ante tools designed to prevent future violations of adequacy and relevance. The empirical intuition is simple:

Once data start to flow into new contexts, the probability that they will be re-used in decisions for which they are not relevant rises sharply.

If a dataset collected for one clearly defined purpose can be repurposed, combined and traded without structural constraints, it is only a matter of time before some actor uses it as input for decisions to which the data have no rational connection – credit scoring based on music taste, insurance pricing based on social media activity, hiring decisions based on political affiliation, and so on.

Seen in this light, purpose limitation is not a decorative recital about fairness. It is the central legal mechanism through which the law tries to shape the future graph of data flows: to determine which data can realistically end up feeding which decision systems. The decision layer governs the content of decisions (adequacy and relevance); the flow layer governs the routes through which data can travel into decision contexts in the first place.

Any attempt to redraw the boundary of “personal data” therefore has two effects:

  1. It changes which operations count as “processing” for the purpose of direct decision governance; and
  2. It determines where the preventive flow-governance layer applies at all. Data that fall outside “personal data” are not just exempt from information duties; they are exempt from purpose limitation and onward-transfer constraints, even if they later feed into decision systems.

The risk of the Omnibus proposal on personal data definitions is best understood at this second level.

3.2 How a Relative Personal Data Definition Undermines Flow Governance

The Omnibus proposal does not invent the idea that personal-data status may depend on the means of a particular controller; it codifies the Court’s relative approach, especially as expressed in more recent case law. The core idea is that information counts as personal data for a controller only if that controller has means that are “reasonably likely to be used” to identify the individuals concerned. Pseudonymised datasets held by actors who lack re-identification keys might therefore fall outside the definition for them, even if another actor could identify the same records.

On its face, this move is presented as a technical solution to a real problem: without some relativisation of identifiability, any data that someone could link to a person risks being treated as personal data for everyone. But once we look at the two-layer structure, a different picture emerges.

Consider a stylised scenario. A company collects information on music preferences – whether customers prefer jazz, rock or classical – linked to stable internal identifiers. At the time of collection, it pseudonymises the data and declares that they are used purely for market research. Under a relative definition, the company might argue:

  • it does not hold the mapping from identifiers to real-world identities;
  • it has no present means reasonably likely to be used to identify individuals;
  • therefore, the dataset is not “personal data” for it.

If accepted, this position has a crucial consequence: purpose limitation never attaches to the dataset. The data have been collected and processed in structured form, but as long as they are treated as “non-personal” for that controller, the entire flow-governance layer of data protection simply does not apply.

Fast-forward several years. The company changes strategy and enters the credit-scoring market. It discovers that music preference correlates, in its dataset, with repayment history. It now has strong incentives to:

  • re-establish or acquire a linkage between identifiers and individuals; and
  • re-use the existing music-preference dataset as an input for credit decisions.

At that moment, the company will insist that it now has identification means and that the data now constitute personal data. But purpose limitation will be invoked too late. Because the data were originally classified as “non-personal”, there was no legal obligation to tie their collection and retention to a specific, limited set of purposes, or to treat later re-use for credit scoring as a purpose change requiring a fresh legal basis and renewed guarantees for the individuals concerned.

This is not an exotic edge case. It describes a systematic pattern of “purpose drift through claimed non-identification”:

  1. At collection: claim non-personal status (no identification means), thereby avoiding purpose limitation and related constraints;
  2. In the interim: enjoy maximal freedom to replicate, analyse and trade the dataset as “non-personal”;
  3. At some later point, when a lucrative decision-application emerges: reclassify the data as personal and use them for consequential decisions, by which time the preventive layer has already failed.

The same logic applies not only to isolated companies, but to entire chains of actors. If upstream entities can treat rich, individual-level structured datasets as “non-personal” on the basis that they lack re-identification keys, those datasets can circulate freely until they reach actors who do have the means and incentives to link them to identified persons and embed them in decision systems. Purpose limitation was meant to constrain exactly such cross-context flows; a relative definition makes it easy to step outside its reach during the most critical stages of dissemination.

The Omnibus proposal’s empowerment of the Commission to designate certain categories of pseudonymised datasets as “non-personal” through implementing acts would exacerbate this risk. Once such a designation is made, entire classes of structured data can travel as if they were mere technical exhaust, while remaining perfectly suited for later reintegration into decision pipelines. What disappears is not only some marginal transparency duty; it is the ex ante requirement to ask, at each step of the flow, “what purposes justify moving these data into this next context?”

From a decision-and-flow-oriented perspective, then, the danger of the Omnibus redefinition is not primarily that it makes the GDPR less “broad” in an abstract sense. It is that it removes structured, decision-capable datasets from the preventive flow-governance layer at precisely the stages where that layer does its most important work.

At the same time, it is worth recalling that the GDPR already contains a gate in the form of its material scope clause. Article 2 confines the Regulation to the “processing of personal data”. The popular complaint that “the GDPR applies to everything” does not stem from the absence of such a gate, but from the way in which “processing” has gradually been flattened into “any handling of any information about a person” – including, in the political imagination, web crawling for AI training or other large-scale analytics. The Omnibus proposals take this perception of overbreadth at face value and try to cure it at the level of the definition of personal data itself. The maximal scope paradigm reacts from the opposite direction and defends the broadest possible reading of “personal data” and “processing”. What both moves have in common is that they leave the structural notion of processing untouched – and with it, the original idea that the material scope of the GDPR was meant to capture systematic, file-like processing pipelines, not every scattered mention of an individual.

3.3 A Decision-and-Flow-Oriented Scope Test

Against that background, if we take seriously both layers of protection, an alternative way of framing the scope question suggests itself. Rather than asking how far we can conceptually stretch the notions of “personal data” and “processing”, we should ask in which situations EU data protection law must, at a minimum, apply with full force across both layers.

A decision-and-flow-oriented test could be formulated, in preview, along these lines:

  1. Systematic pipeline
    There must be a systematic informational pipeline in the original sense of “processing”: a structured sequence of operations on data organised according to stable categories or formats, not merely isolated mentions or ad hoc exchanges. This is the threshold that distinguishes “data” from scattered “information” and captures what early data protection instruments were designed to regulate.

  2. Decision context
    The outputs of this pipeline are, or are reasonably likely to be, used in decision contexts that affect individuals’ legal or social position: credit, employment, insurance, social benefits, access to services, law enforcement, and comparable domains. Here the decision layer kicks in: Article 5(1)(c) must be interpreted and enforced as a genuine relevance requirement, demanding a rational link between each category of processed data and the stated decision purpose.

  3. Cross-context flow risk
    The data in the pipeline can realistically flow into other decision contexts in which they would not be relevant, unless purpose limitation and onward-transfer rules apply. Where such cross-context drift is structurally plausible – as in large-scale analytics, AI development and platform ecosystems – the flow layer should apply from the moment the data are collected in structured form, regardless of whether the controller currently claims to have identification means.

On this view, the problem with the Omnibus redefinition of personal data is not that it abandons an abstract ideal of maximal scope. The problem is that it allows controllers to step outside this three-part test at will by declaring, at collection time, that they lack identification means – even when they are building exactly the kind of structured datasets that can later be fed into consequential decision systems. Once those datasets are born outside the legal category of “personal data”, the law’s ex ante capacity to shape their future flows is crippled.

The same test also helps to separate, conceptually, different families of AI-related processing. Generative-AI training on scattered public text corpora may fail the first and third prongs: it does not create systematic, individual-addressable datasets, and its outputs are not naturally wired into decision contexts. By contrast, treatment AI – credit scoring, hiring algorithms, insurance pricing models – falls squarely within all three: it operates on structured records, directly supports decisions about individuals, and its datasets have obvious cross-context re-use potential.

3.4 Re-reading “Processing of Personal Data” in Light of the Directive’s History

This interpretive proposal is not a call to rewrite the text of Article 4(1) GDPR. It is a call to re-read the existing concepts of “processing of personal data” and “personal data” in light of their directive-era function.

Historically, the structural understanding of “processing” was not an accidental by-product of outdated “file” concepts. The Commission’s 1990 proposal still spoke the language of files, but when the Parliament insisted on replacing “file” by “processing of personal data”, this was not meant to abandon the structural requirement altogether. It was meant to capture file-like processing pipelines, including on-the-fly decision systems that reproduce the same risks without a literal, centralised file. The shift from “file” to “processing of personal data” thus extended the concept along a functional line, rather than dissolving it into “any handling of any information”.

Some judgments – particularly in borderline cases – can be read as stretching the concept for specific factual constellations without turning “processing” into “any handling of any information” as a general rule. Where genuine tensions remain, they suggest that the Omnibus is premature in constitutionalising a maximalist scope at precisely the moment when the underlying concept still requires careful doctrinal clarification.

4. Conclusion: Not a War over Scope, but a Chance to Recover What Data Protection Was For

If you only read the open letters and press releases around the Digital Omnibus, it looks like a familiar drama. On one side, the Commission, accused of “watering down the GDPR” by narrowing the definition of personal data and creating new derogations for AI and research. On the other side, noyb and others, defending a maximal scope for personal data and a strong version of data minimisation.

My argument in this post, building on my earlier analysis, is that this is the wrong axis of conflict. Both sides are fighting over how broad the law’s coverage should be, while largely ignoring what the law was actually designed to control. From the 1970s onward, European data protection law was not about “data” in the abstract. It was about systems of decision-making: the pipelines through which structured data are collected, organised and used to allocate jobs, credit, insurance, social benefits and other advantages and burdens.

Seen from that angle, the core of the GDPR is not an abstract horror of “too much data”, but a pair of linked questions: which data may legitimately be taken into account in decisions about individuals, and how should we prevent data from flowing into decision contexts where they have no rational link to the purpose? That is what the data-quality principle and purpose limitation were originally built to address.

At this point, two natural questions may arise.

The first is that this may still sound like data minimisation by another name. After all, Article 5(1)(c) GDPR now literally speaks of data being “adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed”, and Directive 95/46/EC already spoke of “adequate, relevant and not excessive”. Hasn’t the law always been about minimising data?

My answer is that, historically, this triad was framed as a data-quality principle for decisions, not as a free-floating command to collect “as little data as possible”. “Not excessive” – and its later rephrasing as “limited to what is necessary” – was embedded in a standard that also demanded adequacy and relevance for a given decision purpose. Supervisory guidance often glosses “adequate” as “sufficient to properly fulfil your stated purpose”. But “sufficient” sometimes means more data, not less: more complete income information to avoid arbitrary denials of credit, more context to avoid misinterpreting a single data point, more nuance to prevent crude profiling. A pure minimisation reading, which treats “less data” as always better from a fundamental-rights perspective, erases this ambivalence and collapses adequacy and relevance into a one-dimensional “volume” metric. In that light, rebadging Article 5(1)(c) as the data minimisation principle is itself part of the problem: it obscures the original focus on relevance and sufficiency for decisions, and makes it easier to overlook situations where better protection in fact requires better – not thinner – data.

A second, closely related question is whether all this ultimately collapses into a call to strengthen Article 22 GDPR on automated decision-making. The recent battles around Article 22 are, in a way, a symptom of the same misframing. Because we forget that the data-quality principle already governs the informational basis of all decisions that take personal data into account, there is a tendency to load every concern about profiling and AI onto the narrow category of “solely automated” decisions in Article 22. One camp therefore demands an ever broader Article 22, hoping it will cure problems that actually belong in the general regime of Article 5(1)(c). Another camp concludes that Article 22 is an overreach that stigmatises even proper profiling that respects adequacy and relevance, and pushes to hollow it out or abolish it – a mood that is not entirely absent from the Omnibus debates.

What both sides risk forgetting is that Article 22 is a lex specialis for a very small subset of cases. The basic standard I have tried to reconstruct here is more general and more basic: it applies whenever personal data are used in decision-making at all, regardless of the mix of automation and human involvement. That standard lives primarily in Article 5(1)(c) and the surrounding data-quality and purpose-limitation provisions, not in Article 22. Strengthening Article 22 cannot substitute for taking the relevance requirement seriously across the board. Weakening Article 22 does not make decision systems benign if we continue to ignore adequacy and relevance at the decision layer and to relax flow governance so that irrelevant data can silently seep into consequential decisions.

Against that background, the real promise and danger of the Digital Omnibus become clearer. The promise is that the current controversy could have forced us to reopen the basic questions: what counts as “processing of personal data” in the structural sense? where do the characteristic risks of data protection actually arise? how should we distinguish statistical or learning phases from operational decision-making? The danger is that we instead constitutionalise a maximalist reading of scope at the very moment when the underlying concept of processing still needs careful doctrinal work – and, at the same time, quietly open back doors for exactly those structured datasets that most urgently require tight flow governance.

The alternative sketched in this post is modest in textual ambition but radical in effect. It does not call for rewriting Article 4(1); it calls for re-reading “processing of personal data” and Article 5(1)(c) in line with their directive-era function. It suggests that AI training should be treated as statistical processing under existing rules, and that the real regulatory energy should be directed at AI operation and other decision systems, where the adequacy and relevance requirement finally has to bite. And it suggests that both the Commission and its critics would do better to shift their attention from how far the legal concept of personal data can be stretched, to how the law governs the pipelines by which data flow into decisions. That, after all, was what data protection was for.

デジタル・オムニバス、スコープ、そして関連性──我々は間違った戦いをしているのではないか

本来であれば、学術論文の形で発表すべき(根拠となる文献をふんだんに引用して*1)ところであり、実際その執筆作業は一部進めているのだが、何ヶ月もかけて悠長にしていられる状況ではない。EUの政治は今まさに動いており、このままでは混迷を深めるばかりで、もはや今のEDPB・EDPSでは頼りになるとは期待しがたい*2。仮名化データを個人データとみなすかどうかは、GDPRの立法過程でも争点となったところだし、そもそも欧州評議会1997年勧告の段階で既に認識されていた課題である。CETS 108以来の「データ品質」原則が、GDPRの立法過程でドイツ的な誤訳を経た「データ最小化」原則へと捻じ曲げられ、それを当時の欧州理事会が辛うじて押し返していたことも、今となっては欧州でどれだけ共有されているのか心許ない。前回の投稿について思い切ってドイツのレジェンド教授にメールでお伝えしたところ、なんと、すぐに読んで頂けたようでお返事があり、称賛のお言葉とともに、わしはもう歳じゃから君がやりなさいとの励ましを頂いた。

そんなわけで、取り急ぎ大規模言語モデルの力を借りて、あちらの人たちに刺さる表現になるよう工夫して生成したのが上に掲載した英語ブログである。これは以下の手順で生成した。まず、前回の続きでClaudeに書かせていたのだが、イマイチあちらに刺さる感じに仕上がらなかったのでボツとし、ChatGPT(5.1)に一から書かせた。ChatGPTの調査力はかなり(特に8月のGPT-5以降)良く、ここ半年はChatGPTをメインに調査していたので、そのメモリー機能によってChatGPTは、私がどういう調査をしてどのような立場なのかを概ね把握しているようであり、何も言わなくても大筋の判断はできてしまうようだった。それでも材料が足りなかったので、ボツにしたClaude原稿(それは私の既発表論文群をソースにして生成されている)を読ませて観点を補充し、最終的にこちらから足りない要素を指摘して補完させ、完成に至った。

そのプロンプトを以下に開示しておく。完成版の日本語訳(Claudeによる翻訳)は最後に置いたので、結論から読みたい人は最後のターンまでスキップするとよい。


私:noybがレポートを出してきましたよ。これまで通りかな。 https://noyb.eu/sites/default/files/2025-12/noyb%20Digital%20Omnibus%20Report%20V1.pdf

ChatGPT:ざっと一通り目を通しましたが、「これまで通りのnoyb」という感じで、大枠のスタンスは完全に従来線ですね。

(略)

もし必要なら、あなたの立場(processing 概念やスコープ論)から見て「ここは使える/ここはずれている」という仕分けも、条文ごとにやっていきましょう。

私:おお、どこまで自動的にできるのかな? それをやってみてください。

ChatGPT:やってみました。
PDF 全体をざっとなめて、noyb が章立てしている 14 本の条文ごとのコメントを拾って、「あなたのいつもの立場(処理概念/スコープ論/決定指向モデル)」から見てどれくらいそのまま使えそうかをざっくり自動分類してみました。

記号はこんな感じで付けています:

  • ◎:かなりそのまま使える(立場がかなり近い)
  • ○:基本方向は同じだが、前提理論は違うので要フィルタ
  • △:一部の指摘だけつまみ食いならアリ。基本線はズレ気味
  • ×:あなたのスコープ論と正面から衝突しそう

(略)

私:うーん、いや、「使える」というのは、逆方向で、noybがダメな批判をしているぞ本当はこっちだぞ!ということを言うのに使えるかを基準にしてほしいw。

ChatGPT:なるほど、「noybすごい!使える!」じゃなくて
「noybの批判がズレてる例として使えるかどうか」が基準ですねw 了解です。

ざっくり言うと、この Digital Omnibus レポートは「GDPR=Charter 7/8条の最大限のスコープを死守する」という立場に振り切れていて、あなたの「決定指向+スコープを整理する」路線から見ると、ツッコミどころの宝庫です。

以下、「こういう点で n​oyb がダメ、本当はこう考えるべきだ」という形で、論点ごとに“使い方”を書きます。

(略)

私:おお、よくわかってくれているね、素晴らしい。しかしここから欧州方面の一般読者に説得力のある論述を構成するのはなかなかにタフだと思う。しかしやらねばならぬし、チャンスだよなあ。

ChatGPT:そう、「ここから欧州読者を口説き落とす」のが本番ですよね。
でも、今のタイミングでそれをやるの、本当にチャンスだと思います。

とりあえず「どういうストーリーで書けば欧州方面に刺さるか」を、かなり具体的に組み立ててみます。

(略)

私:なるほどうまいなあ。ではそれをやってみてください。

ChatGPT:(略)

私:おお、上手ですね。その調子で続きまで書けちゃいますか?

ChatGPT:(略)

私:maximal scope paradigm、Maxだけにねw。このパートは答えを示さず終わっている感があるので、次と次を終えた後で解決策を示すのかな?

ChatGPT:(略)

私:セクション3はそんな感じで良さそうですが、セクション4の答えは弱いと思う。決定の適切性確保のための直接規制(原則遵守)だけではなく、不適切な決定を予防するための目的外の利用・提供の禁止がある。目的外利用が禁止されるのは一般的にデータが目的外に決定利用されれば元の目的とは異なって決定の目的に関連性のないデータとなる蓋然性が高いから。流通を制限するのも同じ理由から。今回のpersonal data定義変更はそこを直撃することを言った方がいい。この観点を中心に別途Claudeに書かせた原稿案があるのだけど、Omnibus提案の中身を正確に捉えきれていないなど先走った内容で失敗作になった(それで今ここで作り直している)んだけど、参考になる要素を拾ってみる?

ChatGPT:そうそう、まさにそこですよね。
「決定の適切性の直接規制」だけだと半分しか描けてなくて、目的外利用・提供禁止という「予防の層」が本丸のかなり大きい部分を占めている、というのは絶対に前面に出した方がいい。

(略)

別途Claudeに書かせた原稿案がある … 参考になる要素を拾ってみる?

ぜひやりたいです。
今こちらで組んだ 4章の骨格があるので、(略)

私:こちらがそのClaude原稿です。

ChatGPT:うん、このClaude原稿、めちゃくちゃ素材が詰まってますね。
「決定レイヤー」と「予防レイヤー(目的外利用・流通制限)」という話は、ほぼ全部ここから組み立てられそうです。

(略)

私:まずはいいでしょう。ではその調子でセクション3を書いてみてください。

ChatGPT:いきます、そのまま英語でセクション3を書きますね。(略)

私:OK、細かな問題がいくつかある。
・フローの話をセクション3でしれっと使っているけど、初出はセクション4だよね。作成した順と並べる順が逆転しているよ。
・「As the earlier draft argument reconstructed in Section 7 of my previous paper」とあるけどそれってさっきのClaudeのボツ原稿のことだよね。それは参照できないよ。
・「前回述べたように」のような記述が数箇所あるのは、すでに発表した前哨戦の私のブログのことを指しているのだと思うけど、それは望むところなのだけど、どんなブログだったか把握してる?以下のURLにあるので見ておいて。それを踏まえて、そこにリンクするつもりで書き方を調整したい。
https://takagi-hiromitsu.jp/diary/20251120.html

あとは、セクション4だけど、さっきの内容だと、まだ尻切れ感がある。結論セクションを追加するのかな?

ChatGPT:いいポイントの突き方してくるなあ…w
ご指摘3点、その通りなので、順番と書き方を整理し直します。

(略)

私:概ねそれで良さそうですが、「personal data 定義を flow-governance の観点から再設計」の具体的な中身はどうするのかな。定義を変更せずとも解釈でデータ保護指令の際の意義を再確認するという方向性がいいのではないかなあ。1990年の欧州委員会提案ではfile概念を用いていたけど欧州議会で「fileは古い」と言われて「processing of personal data」に置き換えた経緯がある。それは完全に構造要件を捨てたわけではなく、蓄積を伴わないオンザフライ決定を含める必要性が認識されていたからだと思うのよね。そこの根拠は論文(日本語だが)で発表済みだがそこを参照するほど深入りせずにサクッと断言してしまうのが吉かも。判例と矛盾すると一蹴される予感もするが、それも本当は判決を精査しないといけないと思うのよねえ。

ChatGPT:(略)

私:概ねいいですが、いくつか気になるところが以下です。
・これは論文ではなくブログポストなので、注はしっかり入れてないで割り切っていいです。ブログ用だからと表現を平易にする必要はありませんが、表記が矛盾しないように気をつけて。
・結論として、定義をどうするかの話は盛り込んだが、研究とAIの例外をどうするかが足りてないのでは? 私の立場は要するに学習段階は統計目的として扱えばOKというものです。
・前哨戦の内容をやっぱり見ておいてもらった方が良さそうなので、印刷したPDFを置きます。これを既に読んだ読者向け(未読ならリンク先を見る想定)でうまく簡潔にできる部分があると思う。

ChatGPT:了解です、だいぶ全体像がクリアになりました。
前哨戦ブログもざっと把握しました(あの「None of Your Business=irrelevant」ネタ、今回も美味しく使えますね)。

(略)

私:3.5のところ、惜しいのが、関連性が求められるのは決定段階でありモデル自体(モデルの入力)には求められないのであるということがうまく書けていない。
5.のところ、惜しいのは、「Within that framework, research and AI do not need ……」のところで、生成AIの方のトレーニングではfile的な処理ではないという話を書いてないこと。

ChatGPT:いいところ突いてきますね、その2点はちゃんと直した方がいいやつです。

(略)

私:OK、これであとは全体を生成してみることだけど、その前にちょっと検討。さっき添付したClaudeのボツ原稿では、クローリングに着目して、検索エンジンだって同じなのに問題にしてこなかったじゃないかという話から始めていて、それはそれでキャッチーな入り方だと思っていたのだけど、Google Spein判決の話が入ってきてややこしくなる上、今回のOmnibus提案はそこの部分についてさほど悪くはない案になっているようだから、正当な批判にならなかったと思ってボツにした。その評価は妥当だと思う? 今回のバージョンをそこと比べるとどう評価できる?

ChatGPT:感覚として持ってた違和感、かなり正確だと思います。

(略)

私:OK、小ネタにするのはやめておきましょう。では、全体を生成してください。その際に、全体の表現の統一感を確認しながらでお願いします。canvasに出力できるのでしたっけ?

ChatGPT:できました。

(略)

(略)

私:OK、noybは答えを見つけようとはしていない(できないからかもしれないし、そうする立場でもないからかもしれないが)から、言えちゃう批判をしているだけなので、結果的にmaximal scope paradigmの一員になっているというところかな。もう一つは、欧州委員会のような定義を狭めようとする策というのは典型的に「体系的」processing限定をわかっていない人たちが疑問視するGDPRの定義の広すぎさ論であり、我々の論法は、それに対するカウンターであり、生成AIの学習段階での障害を回避するために定義までいじってきたのだろうという想定で批判しているので、そのこととの関係が今のドラフトでは見えにくくなっている気がしてきたが、どうか。

ChatGPT:そうそう、その整理でだいたいピタッと来てますね。

(略)

私:あ、新たな問題点に気づいた。「scope」の語だけども、定義のscopeのことを言っているのだろうけど、GDPRの規定では、material scopeとして現れる言葉で、そこがまさしく、processing of personal dataに限定すると規定している箇所なので、用語の混乱があるなあ。

ChatGPT:うわ、そのポイントに気づくのめちゃくちゃ鋭いです。
これはたしかに EU 法クラスタに読ませるならケアした方がいいやつですね。

(略)

私:ラベルは生かすのは同感として、しかし、Art.2のscopeとは別と注意書きを入れるのは上手くない(ややこしくなって読者はついてこない)と思うなあ。むしろ、欧州委員会側がmaterial scopeによる限定を無視しているだろ、という指摘を入れることによって、scope概念の交差を中盤か後ろあたりで拾うのはどうだろうか。

ChatGPT:なるほど、その発想いいですね。
「scope という言葉を冒頭で定義で潰す」のではなく、(略)

私:その件とさっきの欧州委員会側の「定義を狭めたい動機」の件を合わせて直すとして、加筆で済ますというよりも周辺の見直しも含めて全体を再検討してもいいよ。

ChatGPT:いいですね、「全体をもう一度組み直してもいい」という一言で、むしろ構造がきれいになりました。

いまやった主な調整はこんな感じです:

(略)

私:ブログ原稿を通して再読した。この内容だと、だいぶ早い段階で「それGDPR 22条を強化すべきという話だよね?」と思ってしまう読者がかなり多いと思う。上の方で単純に否定しても下まで読まないとわからないだろうから、上の方で何か入れて、下の方でそれとの違いを論じた方がいいと思われる。横道に逸れる感じではあるが、どこにどう入れるのがいいだろうか。

ChatGPT:そうそう、その「22条じゃないの?」という読みは、かなり早い段階で大勢の人が頭に浮かびそうなので、そこを先回りしておいた方がいいですね。

いま入れた修正はこんな方針です:

(略)

私:先ほど追加した22条のところ、冒頭で言うにはちょっと重いかなあ。まだそこの段階では「決定」の話をしているとは感じられないのでは?

ChatGPT:うん、それはその通りだと思います。
今の入れ方だと、

(略)

(略)

私:本当は脚注に書くような話なのよね(脚注は今回使わない)。なので確かに、22条パラグラフを上に移動すると思考があっちゃこっちゃ行ってしまうので、この辺に置いた方がいい。というところかなあ。改めて通して読んでみたけど、やっぱりこの場所でも異物感があるなあ。いっそ結論パートに移してはどうだろうか?

ChatGPT:うん、その感覚けっこうよくわかります。
3.1 の流れは「二層構造の提示 → それぞれの役割 → Omnibus でどこが壊れるか」に集中していて、22条はやっぱり「脚注ネタ」に近いので、そこに入るとどうしても異物感が出ますね。

(略)

私:OK、あと、結論パートはまとめだから必然かもしれないけどそれにしても同じ話の繰り返しだなあという重複感は感じていたところ。それに加えてもう一つ、改善の必要なポイントを思いついた。それは、全体を通して、データ品質原則がDPDまでの「adequate, relevant and not excessive」とGDPR以降の「adequate, relevant and limited (minimisation)」とが両方使われているので、混乱しそうでうです。noybがminimisation原理主義になっていると批判しながら、「ate, relevant and not excessive」と持ち正しても「同じものでは?」と思う人がいそう。それは仕方ないが、DPDからGDPRで変えられてしまったのだということを言っておきたいところ。それも合わせて結論パートをそういう疑問視への回収パートにすればいいのではないかな。

ChatGPT:うん、その2点(22条の位置と DPD→GDPR の文言変更)を結論パートで「回収する」方向はかなり良さそうです。

(略)

私:「反論」と書くより、「疑問が湧くかもしれない」的な言い回しの方がいいなあ。

ChatGPT:わかる、そのニュアンスの差はけっこう大きいですよね。「objection」って書くと、読者を敵に回してる感じがちょっと出る。

(略)

私:うーん、1点目は、「adequate」の意味について触れるといいのでは?ICOのガイダンスではsufficientと説明しているよね、これは量を増やすという意味なので、data minimisationというラベルは完全に矛盾しているのだけど、GDPR作った時にadequateもrelevantも意味を取り違えたのでは?という指摘ができる。2点目は、「Strengthening Article 22 may well be desirable」としているけど、むしろ、そこも誤解ゆえの争いになっているところで、自動決定じゃない通常決定も規制しているのがデータ品質原則だということを忘れているが故に問題を22条の強化でカバーしろと言う人たちがいる一方で、これはデータ品質原則に適合する正しいプロファイリングをも規制するものなので規制が不要だという勢力もいて今回のような案が出ているわけですよね。その辺をうまく言いたいな。

ChatGPT:いいポイントが2つきれいに刺さっているので、それをそのまま結論に反映させる形で書き換えてみました。

(略)


これで完成。これをClaudeで日本語訳させたのが以下である。


私:これは先ほどChatGPTに書かせた文章です。日本語訳を作成してもらいます。直訳調より読みやすさを優先してください。専門用語については統一をお願いしたい。まず要注意な用語を洗い出してください。先にその訳を決めましょう。

(略)

デジタル・オムニバス、スコープ、そして関連性──我々は間違った戦いをしているのではないか

以前の投稿で、デジタル・オムニバス草案のリークを取り上げた際、私はnoybという名前に隠された意味──None of Your Business、すなわち無関係──を手がかりに議論を展開した。私が言いたかったのは、GDPRには、本来この議論の中心に据えられるべき原則が明文で存在しているにもかかわらず、ほとんど言及されていないということだ。それは5条1項(c)が求める、データは目的との関係で「十分であり、関連性があり、必要なものに限定される」べきだという要件である。本当の問題は、法がどれだけ多くの「個人データ」を抽象的にカバーできるかではない。ある決定に際して、管理者がどのデータを考慮に入れることが許されるのか──これこそが問われるべき問いなのだ。

その後、デジタル・オムニバス提案が公表され、noybはこれを批判する長大なレポートを公開した。公の議論は、おなじみの対立構図として描かれている。一方には、個人データの定義を狭め、AIや研究のための新たな例外を設けることでGDPRを「骨抜きにしている」と非難される欧州委員会。他方には、「個人データ」の最大限のスコープと、データ最小化の強いバージョンを擁護する市民社会のアクターたち。

本稿で私が主張したいのは、これが間違った対立軸だということである。両陣営とも、法の射程をどこまで広げるべきかをめぐって争っているが、法が本来何をコントロールするために設計されたのかという点はほとんど無視されている。オムニバスをめぐる議論は、定義をめぐる戦争として扱われているが、本来はデータ保護の原点──個人に関する決定の公正さ、そして無関係なデータに基づく決定の防止──を取り戻す機会として扱われるべきなのだ。

この主張を展開するために、まずオムニバス批判を支配している「最大スコープ」パラダイムについて述べる。次に研究とAIに関する規定に目を向け、なぜ欧州委員会もnoybもリスクの所在を見誤っているのかを説明する。最後に、スコープについての別の見方──「個人データの処理」を指令時代の機能に即して読み直すアプローチ──を提案したい。

1. 「最大スコープ」パラダイムとその限界

1.1 オムニバス論争における4条1項の位置づけ

デジタル・オムニバス提案に対する現在の批判の中心的な標的は、GDPR 4条1項、すなわち「個人データ」の定義の改正案である。noybの読み方によれば、新しい条文は識別可能性に強く管理者相対的な要素を導入するものだという。すなわち、ある情報がpersonal dataであるかどうかは、当該管理者の視点、リソース、あるいは合理的な努力に依存することになる。再識別のための鍵を持たない主体が保有する仮名化データは、たとえ別の主体が同じレコードを識別できるとしても、その主体にとっては定義の外に落ちる可能性がある。

noybにとって、これは単なる起草上の不備ではなく、憲法上の問題である。その論理は概ね次のようなものだ。

  • EU基本権憲章は「個人データの保護」を保障している
  • EU司法裁判所はpersonal dataとprocessingを繰り返し広く解釈してきた
  • したがって、そのスコープを縮小するように見えるいかなる立法的措置も、基本権への直接的な脅威である

この見方によれば、データ保護指令におけるpersonal dataの定義は、EU司法裁判所の画期的判決を通じて解釈されてきたとおり、EU基本権憲章8条によって事実上「憲法化」されている。そして「憲法化」されたものは、その定義の広さとして理解される。「識別された又は識別可能な自然人に関するあらゆる情報」をより寛大に解釈すればするほど、8条の保護はより堅固になるはずだ、と。

この観点からすれば、唯一の安全な戦略は、何らかの事象の連鎖によって個人と結びつきうるほぼすべての情報をpersonal dataとして扱うべきだと主張し、その情報を含むほぼすべての操作をprocessingとして扱うべきだと主張することになる。オムニバス提案は、客観的で包括的であるべき定義に、管理者相対的な「主観性」を持ち込もうとする試みとして描かれる。それは法的安定性を損ない、執行を分断し、底辺への競争を招くものだ、と。

私はこのような考え方を最大スコープ・パラダイムと呼ぶことにする。基本権を保護する主要な方法は、personal dataとprocessingの概念的到達範囲を可能な限り押し広げ、その最大線をいかなる後退からも守ることだという信念である。このパラダイムには様々なアクターが様々な形で位置している──noybのようなNGO、一部の当局、そして学界の一部──しかし彼らが共有しているのは、「個人データの処理として法的に該当する状況が増える」ことがほぼ自動的に「より効果的な保護」を意味するという想定である。

1.2 データ保護の本来の標的:「データそのもの」ではなく「決定」

しかし歴史的に見れば、欧州のデータ保護法制の規範的焦点は「データそのもの」に向けられたことはなかった。それは意思決定システムにおけるデータの利用に向けられていた。信用スコアリング、雇用審査、社会保障の裁定、警察活動、リスク評価などである。初期のデータ保護法、およびその国際的な先行規範は、個人に関する自動化された、あるいはコンピュータ支援による決定が、不透明な方法で、不完全、無関係、あるいは偏った情報に基づいて、強力で中央集権的なファイルにまとめられた形で行われうることへの懸念を動機としていた。

この方向性は、現在の議論ではしばしば見過ごされている二つの特徴に見て取れる。

第一に、processingの概念は、本来、いわば情報パイプラインと結びついていた。データが収集され、保存され、結合され、分析され、決定や行為を支援するために使用される、という一連の体系的な操作のことである。一連の手続的・実体的制約を正当化したのは、このパイプライン──そしてそれにさらされる個人に対してパイプラインが生み出すリスク──であった。定義に「収集」「利用」「開示」「消去」が含まれていたのは、このパイプラインの各段階を捉えるためであり、ある人物の名前への孤立した言及をすべて完全な法的意味でのprocessingに変換するためではなかった。

第二に、よく引用される「十分であり、関連性があり、過度でない」という三要素は、意思決定のためのデータ品質原則として構成されていた。その機能は、個人に影響を及ぼす決定が、決定目的と合理的な関連を持つ情報に基づいて行われ、その目的にとって十分に完全かつ正確であり、無関係な詳細や偏見を含まないことを保証することであった。その趣旨は、データを抽象的に最小化することではなく、決定の情報的基盤を制約して、決定が公正で、恣意的でなく、差別的でないようにすることだった。

この決定志向の視点を念頭に置けば、スコープと保護の関係は違って見える。重要なのは、主として「個人データ」というラベルの下にどれだけ多くの情報が入るかではなく、どの情報パイプラインどの決定文脈がデータ保護法の射程内に入るかである。一回限りのジャーナリスティックな言及から複雑な自動プロファイリングシステムまで、あらゆるものを無差別にカバーするprocessingの概念は、レトリックとしては拡張的かもしれないが、そもそも特別な法的レジームを必要とした構造的特徴──中央集権化、自動化、反復的使用──をまさに覆い隠してしまうリスクがある。

1.3 保護の二つの層──見取り図

この観点から見ると、最大スコープ・パラダイムは二つの点で問題がある。

それは過剰に広い。「個人データの処理」という概念的・教義的な装置一式を、データ保護法の特徴的なリスクがごくわずかしか存在しない文脈にまで拡張してしまう。個人へのあらゆる孤立した言及、あらゆる何気ない写真、あらゆる一時的なログエントリをGDPR上の「processing」として扱うことは、「…に関するあらゆる情報」という文言の字義通りの解釈として論理的には可能かもしれないが、それは概念を、本来対処すべきリスク・プロファイルをはるかに超えて拡張してしまう。そうすることで、日常生活における広範な遵守不能か、例外と適用除外の増殖を招き、後者は法をより複雑で理解しにくいものにする。

同時に、それは最も保護が切実に必要とされる場面では保護が不十分である。personal dataの概念的外縁を守ることにエネルギーを集中させることで、真に重要な処理操作──大規模プロファイリング、スコアリング、決定支援システム──の内部構造から注意をそらしてしまう。ここで決定的な問いは、ある情報の断片が抽象的に個人と「関連する」と言えるかどうかではなく、処理のパイプライン全体が、個人に関する決定が十分かつ関連性のあるデータに基づいて行われ、有意義な異議申立ての対象となり、一定の種類の推論に対する明確な禁止によって制約されるような形で構造化されているかどうかである。

言い換えれば、「個人データ」概念の際限のない拡張への固執は、より困難な作業──高度な保護を必要とする決定文脈を特定し規制すること──の代替物として機能するリスクがある。それは、データ保護の基本権が実際にどのような種類の処理を標的としているのか、そしてなぜなのかについてあまり語らなくても、4条1項における管理者相対的な識別可能性のようなスコープの縮小と見えるものを非難することを政治的に容易にしてしまう。

この袋小路から抜け出すためには、データ保護法が常に二つの層で同時に機能してきたことを思い出す必要がある。

  • 決定層。個人に関する具体的な決定においてデータがどのように使用されるかを規律し、「十分であり、関連性があり、必要なものに限定される」は真正な関連性要件として理解されるべきである。
  • 流通層。データが文脈間をどのように移動しうるかを規律する──目的制限、再利用と転送の制限──将来の決定が、その目的と合理的な関連を持たないデータに基づいて行われることを防ぐために。

本稿の残りでは、この二層構造を詳しく解きほぐし、それがオムニバスをめぐる科学研究、AI、およびpersonal dataの定義についての議論をどのように再構成するのに役立つかを示す。

2. 科学研究とAI:決定文脈なきデータ最小化

2.1 オムニバスとnoybが科学研究をどう捉えているか

オムニバス論争の第二の焦点は、4条38項で提案されている「科学研究」の新しい定義と、それに伴う目的制限と情報義務の変更、特に13条5項である。欧州委員会の草案は、科学研究を本質的に「イノベーションをも支援しうるあらゆる研究」と定義し、社会の「知識と福祉」の成長、「当該研究分野における倫理基準」の遵守、および類似の要素を参照する長いリストの基準を付け加えている。

noybのレポートは、このアプローチに対して鋭く、多くの点で正当な批判を展開している。それは次のように指摘する。p>

  • 中核用語である「研究」があまりに広く放置されており、「図書館で資料を調べる」「Google検索を使う」から実験室でネズミを観察することまで、ほとんど何でもカバーしうる
  • 「イノベーションをも支援しうるあらゆる研究」というフレーズは、スコープを単なる可能性(「しうる」)と副産物(「も」)にまで拡張しており、イノベーションの名の下に行われるほぼすべてのデータ駆動活動が該当しうる
  • 研究が「イノベーションを支援しうる」という要件は、経済的・技術的意味でのイノベーションを目指さない学術研究領域を排除するリスクがある
  • 「社会の一般的な知識と福祉の成長」や「倫理基準」の遵守への言及は不確定であり、憲章を制限する概念の定義を私的または分野別の倫理規範に事実上外部委託している

この分析に基づき、noybは、提案された定義は、研究適用除外の拡大された枠組みと組み合わさって、「広範かつ不明確に定義された『研究目的』のためのさらなる処理を一律に許容する」ことになり、憲章の必要性と比例性の要件を満たさない可能性が高いと主張する。この読み方によれば、欧州委員会は無定形の「イノベーション」アジェンダの祭壇に7条と8条の権利を犠牲として捧げているように見える。

一見すると、この批判は権利保護的な立場と完全に一致しているように見える。しかしその内部ロジックは、やはり二つのおなじみの前提に依拠している。第一に、データ保護の核心は、7条・8条および52条1項の直接的な発現としてのデータ最小化であるという前提。第二に、過度に広い研究定義がもたらす主要な脅威は、このデータ最小化と目的制限そのものの浸食であるという前提。どの種類の研究が、そしてどの決定文脈が、より厳格な審査の対象とされるべきかという問いは、ほとんど暗黙のままである。

2.2 決定を伴わない研究と決定に供される研究

決定志向の視点──意思決定における十分性と関連性の役割を真剣に受け止める視点──は、この問題を切り分ける別の方法を示唆する。

一方には、識別可能な個人に関する具体的な決定に直接供されない形態の研究がある。古典的な統計的・科学的データ利用、すなわち産出物が集団を記述する集計的パターン、係数、モデルであるような利用は、倫理、分配的正義、認識論的バイアスの重要な問題を提起しうる。しかしデータ保護の観点からは、そのリスク・プロファイルは主として研究のに何が起こるかによって媒介される。すなわち、得られた知見が運用上の決定システムにどのように、どの程度組み込まれるかである。

他方には、処遇決定に構造的に接続されている形態の研究がある。顧客データベースで開発される信用リスクモデル、過去の応募者で訓練される採用アルゴリズム、保険価格モデル、不正検知ツールなどである。ここでは、「研究」と「運用」の境界は薄い。モデルを開発するために使用されるのと同じ構造化されたデータセットが、個人に関する重大な決定を支援するために使用される、あるいは容易に使用可能である。

この第二の類型において、データ保護にとって重要なのは、抽象的な「研究」が厳密に必要な以上のデータを使用することではなく、結果として生じる決定システムが決定目的と関連しないデータに基づいて運用されうることである。核心的なリスクは恣意的または差別的な処遇である。音楽の趣味が与信限度額に影響を与える、ソーシャルメディア活動が採用に影響する、近隣の代理変数が保険価格において民族性の代わりに使われる。これらはまさに、十分性と関連性の要件が本来防止するはずだった種類の問題である。

この観点からすれば、オムニバス提案の4条38項および13条5項の変更は問題があるが、それは主としてデータ最小化という抽象的な理想を侵害するからではない。それらが問題なのは、以下の理由による。

  • 単に統計的知見を生成する研究と、決定パイプラインに構造的に組み込まれた研究との区別を曖昧にしている
  • 後に流通層と呼ぶもの──無関係または差別的な決定におけるデータ利用を防ぐために、データが文脈間をどのように移動しうるかを制限するルールの集合──を弱める形で、目的制限と透明性義務を緩和している

noybの批判は、定義の曖昧さと「イノベーション」がさらなる処理の白紙委任として使われるリスクを正しく攻撃している。しかし、問題を主として憲章に適合するデータ最小化の観点から捉えることで、決定的な問いを発する機会を逃している。どの研究生成データセットが決定文脈に流入する可能性が高く、無関係または差別的な意思決定を防ぐためにその再利用はどのように制約されるべきか? これは、以前のオムニバス論争に関する投稿で私が展開した視点であり、そこで私は欧州委員会もnoybも5条1項(c)のGDPRの明示的な関連性要件を見落としがちだと主張した。

2.3 AIと特別カテゴリー:9条2項(k)および5項の下で混同された二つの活動

同じパターンが、AIと特別カテゴリーのデータを扱う9条2項(k)および5項に関するオムニバス提案において、より鮮明な形で繰り返されている。草案は「AIシステム…またはAIモデルの開発および運用の文脈における処理」のための新たな適用除外を導入し、5項では管理者に対し、訓練、テスト、検証において特別カテゴリーの処理を避けるための「適切な組織的・技術的措置」を講じ、特定された場合はそのようなデータを削除するか、削除に不均衡な努力を要する場合は少なくとも出力への影響や開示を防止することを求めている。提案された前文33は、これを「AIの開発と運用を不均衡に妨げないために」9条1項の禁止からの適用除外として明示的に位置づけている。

noybはこの構造をおなじみの論拠で批判している。AIはまさにデータ保護原則が本来制約するために設計された、不透明で大規模な処理の一種であり、AI開発・運用のための新たな適用除外を設けることは後ろ向きであり、そのようなシステムを念頭に置いて書かれたまさにその保障を骨抜きにする恐れがある、と。また、緩い起草(「の文脈における」「適切な措置」「不均衡な努力」)と、何らかの最小化努力を講じたと主張できる限りAI開発者が特別カテゴリーを保持することを一般的に許可することになるリスクも指摘している。

ここでも、この批判路線は理解できるが、それは「AI開発」が単一の活動であり、関連する憲章上のリスクは処理されるデータの量と機微性によって捕捉できるという前提に立っている。実際の状況はより複雑である。

以前のブログ投稿で私が主張したように、「AI開発」というラベルは現在、根本的に異なる二つの活動をカバーしている。

  • 処遇AI、例えば信用スコアリング、採用、保険モデルは、真正な体系的データセット──データ保護法の意味でのプロトタイプ的な「個人データ」である顧客または候補者のデータベース──で訓練される。ここでは、データベース志向の要件が構造的に意味を持つ。訓練セットにおける不必要な特別カテゴリーを最小化し、不適切に含まれている場合は削除し、機微な訓練データを開示や悪用から保護する。

  • 生成AIの公開ウェブテキストに対する訓練は、対照的に、すでに公開されている情報の散在する断片を処理して統計的パターンを抽出する。それは体系的な個人レコードを作成しない。結果として生じるモデルは、通常の運用において、特定の人物の機微な特性についてアドレス可能なレコードを含まない。主要なリスクは、識別された個人についての特別カテゴリーの開示ではなく、生成されたコンテンツにおけるバイアスやステレオタイプの増幅にある。

第二の類型について、9条5項の要件は単に的外れである。管理者に対し、異質なウェブページのコーパスから特別カテゴリーを「削除」するよう指示することは、存在しないデータベース的構造を前提としている。正確には何を削除すべきなのか──文、段落、文書全体か?「可能な場合」という修飾語は、首尾一貫した基準を提供することなく実行不可能性を認めている。特別カテゴリーの「開示」を防止するという文言は、開示が管理されたデータベースから特定の個人の機微データを明らかにすることを意味する文脈を前提としており、健康や宗教が抽象的に言及される新しいテキストを生成することではない。

これらの構造的に異なる活動を単一の適用除外の下にまとめることで、オムニバス提案はpersonal dataの定義においてすでに見られた概念的混同を繰り返している。同時に、決定志向の観点からより重要なことに、それは規制上の注意を誤った場所に向けている

  • 散在する公開テキストに対する生成AIの訓練は、個人に関する決定に適した体系的データセットを構築することがないため、規範的にはpersonal dataの処理として扱われるべきではないにもかかわらず、実行不可能なデータベース志向の要件の直接の標的となる。

  • 処遇AIの運用──構造化されたデータセットを使用して個人に関する重大な決定を行う場面──は、比較的少ない的を絞った注目しか受けない。88c条はAI運用を正当な利益として認めているが、この文脈で5条1項(c)の関連性要件を執行する仕組みを提供していない。したがって、音楽の好み、ソーシャルメディア活動、宗教的所属といった無関係な特性が、「正当な利益」の旗印の下で決定に織り込まれうることになり、個人の異議申立権のみが制約となる。

要するに、オムニバス提案もその批判の多く(noybを含む)も、AIシステムがどれだけの、どのような種類のデータを処理しうるかに焦点を当て、AI支援決定に使用されるデータが決定目的と関連しているかどうか、そして恣意的または差別的な結果を生む形でのデータの文脈横断的再利用をどのように防ぐかという問いは、ほとんど手つかずのまま残している。これがまさに、5条1項(c)の関連性に基づく解釈が埋めるべきギャップなのである。

2.4 では、研究とAIの例外はどうあるべきか

この時点で明白な問いが浮かぶ。欧州委員会もnoybも問題を間違った形で捉えているなら、研究とAIのより良い扱いは実際にはどのようなものになるのか? 新しい研究定義と9条2項(k)のAI適用除外──そしてpersonal dataを再定義しようとする試み──の両方に共通する政治的な動因の一つは、現在の「個人データの処理」の解釈の下では、ウェブスケールのAI訓練やその他の大規模分析さえもGDPRの完全なレジームに自動的に吸い込まれてしまうという恐れであることは明らかだろう。

私の答えは、リークされたオムニバス草案に関する以前の投稿と本質的に同じだが、新しい定義とAI提案にまで拡張される。

  1. 学習フェーズには既存の統計・研究の枠組みを維持する
    欧州にはすでに、個人に関する決定にそれ自体が供されない大規模分析を扱う首尾一貫した方法がある。それは以下の組み合わせである。

  2. GDPR 5条1項(b)における統計・科学目的のための「両立性」ルール

  3. 欧州評議会の統計処理に関する古い枠組み。これは、統計結果に基づいて個人に関する決定や措置が取られない限り、個人の権利の制限を認めている。

AI訓練はここに属する。大規模データセットでモデルを訓練することは、実質的に統計処理である。この段階では、定義上、まだ具体的な決定目的がないので、モデルやその入力のレベルで関連性テストを適用することにはあまり意味がない。十分性と関連性の要件は、決定の情報的基盤に対する制約であり、抽象的なモデルの構造に対する制約ではない。保護される核心的利益は、学習アルゴリズムが特定の変数を「見た」というだけで侵害されるのではない。それは、それらの変数が実際に個人に関する決定パイプラインの一部として使用されるときに侵害される。

  1. 運用、すなわち意思決定フェーズに明確な線を引く
    訓練されたモデルが与信決定、採用決定、保険価格設定、不正検知などを支援するために使用されるようになれば、我々はもはや「単なる統計」の領域にはいない。処遇AIの領域にいるのであり、そこでは十分性と関連性の要件が効力を発揮しなければならない。決定パイプラインに実際に投入されるすべてのデータカテゴリーは、決定目的と合理的な関連を持たなければならない。

この区別は決定的に重要である。モデルは内部的に、決定時に明示的に使用されれば無関係となる特徴を含む多くの特徴に影響されたパラメータを含んでいるかもしれない。データ保護にとって重要なのは、ある人物について決定する際に実際にどのデータが考慮されるか──運用システムにおいてどの変数が読み取られ、結合され、行動の基礎とされるか──である。したがって、法的アーキテクチャは非常にシンプルであるべきだ。

  • 学習・訓練=統計または研究処理。そのような目的のための既存ルールでカバーされ、産出物が直接に個人について決定するために使用されないことを厳格な条件とする

  • 運用=意思決定。完全なデータ品質審査の対象となり、決定に実際に使用される入力に関して5条1項(c)の「十分であり、関連性があり、過度でない」の真摯な執行を含む

  • 特別なAI適用除外は不要であり、広範なものは有害である
    AI訓練を統計処理として、AI運用を通常の(しかし極めて重大な)意思決定として扱うなら、9条2項(k)その他における新たなAI固有の適用除外は必要ない。むしろ、「AI開発・運用のための」一律適用除外は、以前の投稿で私が批判したまさにそのことを招くリスクがある。管理者が「AI」をマジックワードとして援用し、両フェーズを同時に正当化することを可能にし、運用フェーズにおけるシンプルな問い──このAI支援決定に実際に使用するデータは、決定目的と関連しているか?──を問うことがなくなる。

  • 特別カテゴリーもAI専用レジームを必要としない
    学習フェーズについて、正しい問いは「ウェブスケールの訓練コーパスから特別カテゴリーのあらゆる痕跡をどのように除去できるか?」ではなく、「統計目的で使用される訓練コーパスが、特別カテゴリー──直接的であれ推論されたものであれ──が決定目的と無関係となる決定システムにカジュアルに転用されないようにするにはどうすればよいか?」である。

運用フェーズについて、関連性が適切に執行されれば、直接明らかにされた機微データと推論された機微データの区別はさほど重要ではなくなる。両方とも、決定目的に照らして「あなたには無関係」である場合には排除される。以前の投稿で私が主張したように、問題は、我々が決定層における関連性についての執行力をまず構築することなく、9条をスリム化しようとしていることである。

言い換えれば、AI固有の例外や研究固有のショートカットを発明する代わりに、AI訓練を本来の場所──既存の統計・研究の枠組み──に戻し、規制上のエネルギーを、関連性要件が決定がなされる時点でついに真剣に受け止められなければならないAI運用に集中させるべきなのである。

3. 決定ガバナンスと流通ガバナンス:「個人データの処理」を読み直す

ここまでの議論は、データ保護が相互に連関する二つの層で機能していることをすでに示唆している。

  • 決定ガバナンス。個人に関する決定を行う、または支援するためにデータがどのように使用されるかを法が制約する。
  • 流通ガバナンス。将来の無関係または差別的な決定における悪用を防ぐために、データが文脈間をどのように移動しうるかを法が制約する。

この構造をより明確にし、オムニバス提案が各層にどのように影響するかを問うことができる。

3.1 保護の二つの相補的な層

機能的な観点から見ると、EUのデータ保護法は二つの別個だが相互依存的な層で機能している。

第一に決定の直接的ガバナンスがある。個人データが実際に個人に関する決定を行う、または支援するために使用される場合、データ保護法は、それらの決定の情報的基盤が伝統的なデータ品質基準を満たすことを要求する。データは決定の目的にとって十分であり関連性があり、過度であってはならない。現代の言葉で言えば、使用されるデータ項目と正当な決定目的との間に合理的な関連がなければならない。古典的な「十分であり、関連性があり、過度でない」という定式が属するのはここである。それは決定がどのように行われるかに対する制約であり、データを抽象的に最小化せよという呼びかけではない。

第二にデータ流通の予防的ガバナンスがある。目的制限、再利用と転送の制限、「データサイロ」の構築は、過ぎ去ったデータベース規制時代の不可解な遺物ではない。それらは十分性と関連性の将来の違反を防止するために設計された事前のツールである。経験的な直観はシンプルである。

データがいったん新しい文脈に流入し始めると、それが関連性のない決定に再利用される確率は急激に上昇する。

ある明確に定義された目的のために収集されたデータセットが、構造的制約なしに転用、結合、取引されうるなら、ある主体がそれを、データと合理的な関連のない決定への入力として使用するのは時間の問題である──音楽の趣味に基づく信用スコアリング、ソーシャルメディア活動に基づく保険価格設定、政治的所属に基づく採用決定など。

このように見れば、目的制限は公正さについての装飾的な前文ではない。それは、どのデータが現実的にどの決定システムに供されうるかというデータ流通の将来のグラフを法が形成しようとする中心的な法的メカニズムである。決定層は決定の内容(十分性と関連性)を規律し、流通層はデータがそもそも決定文脈にたどり着きうる経路を規律する。

したがって、「個人データ」の境界を引き直そうとする試みは二つの効果を持つ。

  1. 直接的な決定ガバナンスの目的でどの操作がprocessingとしてカウントされるかを変える
  2. 予防的な流通ガバナンス層がそもそもどこに適用されるかを決定する。personal dataの外に落ちるデータは、情報義務から免除されるだけではない。たとえ後に決定システムに供されるとしても、目的制限と転送制約からも免除される。

オムニバス提案のpersonal dataの定義に関するリスクは、この第二のレベルで最もよく理解される。

3.2 相対的なpersonal data定義が流通ガバナンスをどのように損なうか

オムニバス提案は、personal dataの地位が特定の管理者の手段に依存しうるという考えを発明したのではない。それは裁判所の相対的アプローチ、特により最近の判例法で表明されたものを法典化している。核心的な考えは、情報がある管理者にとってpersonal dataとしてカウントされるのは、その管理者が当該個人を識別するために「合理的に使用される可能性がある」手段を持っている場合に限られるというものである。再識別のための鍵を持たない主体が保有する仮名化データセットは、たとえ別の主体が同じレコードを識別できるとしても、その主体にとっては定義の外に落ちる可能性がある。

表面上、この措置は現実の問題に対する技術的解決策として提示されている。すなわち、識別可能性を何らかの形で相対化しなければ、誰かが人物と結びつけうるあらゆるデータがすべての者にとってpersonal dataとして扱われるリスクがある、という問題である。しかし二層構造を見れば、異なる像が浮かび上がる。

単純化した仮想事例を考えよう。ある会社が、顧客がジャズ、ロック、クラシックのどれを好むかという音楽の好みに関する情報を、安定した内部識別子に紐づけて収集している。収集時点で、データを仮名化し、純粋に市場調査のためにのみ使用すると宣言する。相対的な定義の下では、会社は次のように主張しうる。

  • 識別子から現実世界のアイデンティティへのマッピングを保持していない
  • 個人を識別するために合理的に使用される可能性がある手段を現在持っていない
  • したがって、このデータセットは自社にとってpersonal dataではない

もしこの立場が認められれば、決定的な帰結が生じる。目的制限がデータセットに一切付着しない。データは構造化された形式で収集・処理されているが、その管理者にとって「非個人」として扱われる限り、データ保護の流通ガバナンス層全体が単純に適用されない。

数年後に早送りしよう。会社は戦略を変更し、信用スコアリング市場に参入する。そのデータセットにおいて、音楽の好みが返済履歴と相関することを発見する。今や会社には、以下の強いインセンティブがある。

  • 識別子と個人との紐付けを再確立または取得する
  • 既存の音楽の好みデータセットを与信決定への入力として再利用する

その時点で、会社は今は識別手段を持っており、データは今はpersonal dataを構成すると主張するだろう。しかし目的制限は遅すぎる形で援用される。データが当初「非個人」として分類されていたため、その収集と保持を特定の限定された目的のセットに結びつける法的義務も、後の信用スコアリングへの再利用を、新たな法的根拠と当該個人への新たな保証を必要とする目的変更として扱う義務もなかった。

これは奇抜なエッジケースではない。それは「識別手段がないとの主張を通じた目的漂流」という体系的パターンを記述している。

  1. 収集時:非個人の地位を主張し(識別手段がない)、それによって目的制限と関連する制約を回避する
  2. 中間期:「非個人」としてデータセットを複製、分析、取引する最大限の自由を享受する
  3. 収益性の高い決定用途が現れた後のある時点:データを個人として再分類し、重大な決定に使用する。その時点で予防層はすでに失敗している

同じロジックは、孤立した会社だけでなく、主体の連鎖全体にも適用される。上流の主体が、再識別の鍵を持っていないことを根拠に、豊かで個人レベルの構造化されたデータセットを「非個人」として扱えるなら、それらのデータセットは、識別された人物と結びつけ、決定システムに組み込む手段とインセンティブを持つ主体に到達するまで、自由に流通しうる。目的制限はまさにそのような文脈横断的流通を制約するはずだった。相対的な定義は、普及の最も重要な段階においてその射程から外れることを容易にする。

欧州委員会に対し、実施法を通じて仮名化データセットの特定のカテゴリーを「非個人」として指定する権限を与えるオムニバス提案は、このリスクを悪化させるだろう。そのような指定がなされれば、構造化されたデータのクラス全体が、後に決定パイプラインに再統合されるのに完全に適している間も、単なる技術的な排気ガスであるかのように移動しうる。消失するのは、何か周辺的な透明性義務だけではない。流通の各段階において「これらのデータをこの次の文脈に移動させることを正当化する目的は何か?」と問う事前の要件が消失するのである。

決定と流通志向の観点からすれば、オムニバスの再定義の危険は、主としてGDPRを抽象的な意味で「広くなくする」ことではない。それは、構造化され、決定に使用可能なデータセットを、予防的な流通ガバナンス層がその最も重要な仕事をするまさにその段階において、その層から除外することである。

同時に、GDPRはその適用対象範囲条項の形で、すでにゲートを持っていることを思い出す価値がある。2条は、規則を「個人データの処理」に限定している。「GDPRは何にでも適用される」という広く聞かれる不満は、そのようなゲートが存在しないからではなく、processingが次第に「ある人物に関するあらゆる情報のあらゆる取扱い」へと平坦化されてきた──政治的想像力においては、AI訓練のためのウェブクローリングやその他の大規模分析を含めて──からである。オムニバス提案はこの過広認識を額面通りに受け取り、personal dataの定義そのもののレベルでそれを治療しようとしている。最大スコープ・パラダイムは逆方向から反応し、personal dataとprocessingの可能な限り広い解釈を擁護している。両方の動きに共通しているのは、processingの構造的概念を手つかずのまま放置していることである──そしてそれとともに、GDPRの適用対象範囲は体系的でファイル的な処理パイプラインを捉えるためのものであり、個人への散在するあらゆる言及を捉えるためのものではなかったという、本来の考えも放置している。

3.3 決定と流通志向のスコープ・テスト

この背景に照らして、保護の両方の層を真剣に受け止めるなら、スコープの問いを構成する別の方法が示唆される。personal dataとprocessingの概念をどこまで概念的に拡張できるかを問う代わりに、どの状況においてEUのデータ保護法が、最低限、両方の層にわたって完全な力で適用されなければならないかを問うべきである。

決定と流通志向のテストは、予告的に、次のように定式化できよう。

  1. 体系的なパイプライン
    processingの本来の意味での体系的な情報パイプラインが存在しなければならない。安定したカテゴリーやフォーマットに従って組織化されたデータに対する、構造化された一連の操作であり、単なる孤立した言及やアドホックなやり取りではない。これが、「データ」を散在する「情報」から区別し、初期のデータ保護法制が規制しようとしたものを捉える閾値である。

  2. 決定文脈
    このパイプラインの産出物が、個人の法的・社会的地位に影響を与える決定文脈──信用、雇用、保険、社会給付、サービスへのアクセス、法執行、およびそれに類する領域──で使用される、または合理的に使用される可能性がある。ここで決定層が作動し、5条1項(c)は、処理される各データカテゴリーと表明された決定目的との間に合理的な関連を要求する、真正な関連性要件として解釈・執行されなければならない。

  3. 文脈横断的流通リスク
    パイプライン内のデータが、目的制限と転送制限が適用されない限り、それらが関連性を持たない別の決定文脈に現実的に流入しうる。このような文脈横断的な漂流が構造的にあり得る場合──大規模分析、AI開発、プラットフォームエコシステムにおけるように──流通層は、管理者が現時点で識別手段を持っていると主張するかどうかにかかわらず、データが構造化された形式で収集された瞬間から適用されるべきである。

この見方によれば、オムニバスのpersonal data再定義の問題は、最大スコープという抽象的な理想を放棄することではない。問題は、それが管理者がこの三部構成のテストから任意に離脱することを許容する点にある。収集時点で識別手段を持たないと宣言するだけで──たとえ後に重大な決定システムに投入されうるような構造化されたデータセットをまさに構築している最中であっても──テストから外れることができてしまう。そのようなデータセットがpersonal dataという法的カテゴリーの外側で生まれてしまえば、その将来の流通を事前に形成する法の能力は損なわれる。

同じテストは、概念的に、AI関連処理の異なる類型を区別するのにも役立つ。散在する公開テキストコーパスに対する生成AIの訓練は、第一と第三の要件を満たさないかもしれない。それは体系的で個人宛(individual-addressable)のデータセットを作成せず、その産出物は自然に決定文脈に接続されない。対照的に、処遇AI──信用スコアリング、採用アルゴリズム、保険価格モデル──は三つの要件すべてに該当する。構造化されたレコードを操作し、個人に関する決定を直接支援し、そのデータセットには明らかな文脈横断的再利用可能性がある。

3.4 指令の歴史に照らした「個人データの処理」の再解釈

この解釈的提案は、GDPR 4条1項の条文を書き換えることを求めるものではない。既存のprocessing of personal dataおよびpersonal dataという概念を、その指令時代の機能に照らして読み直すことを求めるものである。

歴史的に、processingの構造的理解は、時代遅れの「ファイル」概念の偶然の副産物ではなかった。欧州委員会の1990年提案はなおファイルの言葉で語っていたが、欧州議会が「ファイル」を「個人データの処理」に置き換えることを主張したとき、それは構造的要件を完全に放棄することを意味していなかった。それは、文字通りの中央集権的ファイルなしに同じリスクを再現するオンザフライの決定システムを含め、ファイル的な処理パイプラインを捉えることを意図していた。「ファイル」から「個人データの処理」への移行は、概念を機能的な線に沿って拡張したのであり、「あらゆる情報のあらゆる取扱い」へと溶解させたのではなかった。

一部の判決──特に境界事例において──は、processingを「あらゆる情報のあらゆる取扱い」という一般則に変換することなく、特定の事実関係のために概念を拡張したものとして読むことができる。真の緊張が残る場合、それは、基礎にある概念がなお慎重な教義的明確化を必要としているまさにその時に、オムニバスが最大主義的スコープを憲法化するのは時期尚早であることを示唆している。

4. 結論:スコープをめぐる戦争ではなく、データ保護の本来の目的を取り戻す機会

デジタル・オムニバスをめぐる公開書簡やプレスリリースだけを読めば、おなじみのドラマに見える。一方には、個人データの定義を狭め、AIや研究のための新たな適用除外を設けることでGDPRを「骨抜きにしている」と非難される欧州委員会。他方には、個人データの最大スコープとデータ最小化の強いバージョンを擁護するnoybその他のアクター。

本稿で私が以前の分析を踏まえて主張してきたのは、これが間違った対立軸だということである。両陣営とも、法の射程をどこまで広げるべきかをめぐって争っているが、法が本来何をコントロールするために設計されたのかという点はほとんど無視している。1970年代以降、欧州のデータ保護法は抽象的な「データ」を対象としていたわけではない。それは意思決定システムを対象としていた。構造化されたデータが収集され、組織化され、雇用、信用、保険、社会給付その他の利益と負担を配分するために使用されるパイプラインである。

この観点から見れば、GDPRの核心は「データが多すぎる」ことへの抽象的な恐怖ではなく、一対の関連する問いである。個人に関する決定においてどのデータが正当に考慮されてよいのか、そしてデータが目的と合理的な関連を持たない決定文脈に流入することをどのように防ぐべきか? これこそが、データ品質原則と目的制限が本来対処するために構築されたものである。

ここで、二つの自然な疑問が浮かぶかもしれない。

第一に、これは結局、名前を変えたデータ最小化ではないのか、という疑問である。結局のところ、GDPR 5条1項(c)は今や文字通り、データは処理目的との関係で「十分であり、関連性があり、必要なものに限定される」べきだと述べており、データ保護指令もすでに「十分であり、関連性があり、過度でない」と述べていた。法は常にデータを最小化することを目指してきたのではないか?

私の答えは、歴史的に見れば、この三要素は決定のためのデータ品質原則として構成されていたのであり、「できるだけ少ないデータを収集せよ」という宙に浮いた命令ではなかった、というものである。「過度でない」──そして後に「必要なものに限定される」と言い換えられたもの──は、所与の決定目的にとっての十分性関連性をも要求する基準の中に埋め込まれていた。監督当局のガイダンスはしばしば「十分」を「表明された目的を適切に果たすのに足りている」と解説している。しかし「足りている」は時として、より少ないデータではなく、より多くのデータを意味する。恣意的な信用拒否を避けるためのより完全な収入情報、単一のデータポイントの誤解を避けるためのより多くの文脈、粗雑なプロファイリングを防ぐためのより多くのニュアンス。「より少ないデータ」を基本権の観点から常により良いと扱う純粋な最小化の読み方は、この両義性を消し去り、十分性と関連性を一次元的な「量」の指標に押しつぶしてしまう。この観点からすれば、5条1項(c)をデータ最小化原則とリブランドすること自体が問題の一部なのである。それは決定にとっての関連性と十分性という本来の焦点を覆い隠し、より良い保護が実際にはより薄いデータではなくより良いデータを必要とする状況を見落としやすくする。

第二の、密接に関連する疑問は、これは結局、自動化された意思決定に関するGDPR 22条の強化を求めることに帰着するのではないか、というものである。22条をめぐる最近の戦いは、ある意味で、同じ誤った枠組みの症状である。データ品質原則がpersonal dataを考慮に入れるすべての決定の情報的基盤をすでに規律していることを忘れているがゆえに、プロファイリングやAIに関するあらゆる懸念を、22条の「専ら自動化された」決定という狭いカテゴリーに背負わせる傾向がある。そこで一方の陣営は、実際には5条1項(c)の一般的レジームに属する問題を治癒してくれることを期待して、22条をますます広げることを要求する。もう一方の陣営は、22条は十分性と関連性を尊重した適切なプロファイリングさえも烙印を押す行き過ぎだと結論づけ、それを骨抜きにするか廃止することを推進する──この雰囲気はオムニバス論争においても皆無ではない。

両陣営が忘れがちなのは、22条は非常に小さなサブセットのケースのための特別法(lex specialis)だということである。私がここで再構成しようとしてきた基本的な基準は、より一般的でより基本的なものである。それは、自動化と人間の関与の組み合わせがどうであれ、personal dataが意思決定に使用されるすべての場合に適用される。その基準が主として根拠を置くのは5条1項(c)および周辺のデータ品質・目的制限規定であり、22条ではない。22条を強化することは、関連性要件を全面的に真剣に受け止めることの代替にはなりえない。22条を弱体化しても、決定層における十分性と関連性を無視し続け、無関係なデータが重大な決定に密かに浸透できるように流通ガバナンスを緩和するなら、決定システムが無害になるわけではない。

この背景に照らせば、デジタル・オムニバスの真の約束と危険がより明確になる。約束とは、現在の論争が基本的な問いを再び開かせる契機となりえたはずだということである。構造的な意味で「個人データの処理」として何がカウントされるのか? データ保護の特徴的なリスクは実際にはどこで生じるのか? 統計・学習フェーズと運用上の意思決定をどのように区別すべきか? 危険とは、processingの基礎概念がなお慎重な教義的作業を必要としているまさにその時に、最大主義的なスコープ解釈を憲法化してしまうこと──そして同時に、最も緊急に厳格な流通ガバナンスを必要とするまさにその構造化されたデータセットのために、密かにバックドアを開けてしまうことである。

本稿で素描した代替案は、条文上の野心は控えめだが、効果においてはラディカルである。それは4条1項を書き換えることを求めていない。「個人データの処理」と5条1項(c)を、その指令時代の機能に即して読み直すことを求めている。AI訓練は既存のルールの下で統計処理として扱われるべきであり、真の規制上のエネルギーは、十分性と関連性の要件がついに効力を発揮しなければならないAI運用やその他の決定システムに向けられるべきだと提案している。そして、欧州委員会もその批判者たちも、personal dataという法的概念をどこまで引き伸ばせるかから、データが決定に流入するパイプラインを法がどのように規律するかへと注意を移した方がよいと提案している。結局のところ、それこそがデータ保護の存在理由だったのだから。


*1  例えば、ドイツが歴史的にデータ品質原則を一度も理解した様子がないことを裏付けるには、1973年のSpiros Simitisの初期の構想「Datenschutz - Notwendigkeit und Voraussetzungen einer gesetzlichen Regelung」(DVR 2巻138頁以下)で「Relevanz」概念がどう書かれていたか、1977年のOtto Mallmannが「Zielfunktionen des Datenschutzes: Schutz der Privatsphäre, korrekte Information; mit einer Studie zum Datenschutz im Bereich von Kreditinformationssystemen」で、信用情報システム分野を題材に、被影響者の差別のリスクに言及しながら「正確性」「完全性」に並び「文脈依存性の維持」を挙げて問題の本質に迫っていたにもかかわらず、その出口を必要性(Erforderlichkeit)に求めてしまったこと、その後もSimitisは一貫して何らかの具体的基準の必要性を訴え続けるものの、その答えは示されず、2000年ごろから出口が見えないままDatensparsamkeitに傾倒して行き、ついにはHans Peter Bullに、データ禁欲(Datenaskese)とか正気か?と窘められるという展開があったこと、Bull自身は、データ保護の趣旨を正しく理解し、より多くのデータ項目を要することもある旨を指摘していた(しかしDPDのデータ品質原則がそれをカバーしていることには触れていない)こと、ドイツでも労働法分野では古くから関連性(sachlicher Zusammenhang)を要件とする判例と学説が確立していたが、データ保護法制のデータ品質原則がそれをカバーしているとは気づいていない様子がある(Peter GolaのBDSGコンメンタールなど)こと、国内法のBDSGは長らくDPDのデータ品質原則を無視して(比例原則と勘違いして)導入しなかった(対応済みだと勘違いして)こと、それが原因でSCHUFAの住所を用いたGeoscoringに対処できず、BDSG 2010で個別の規制条項を導入していること、それにもかかわらず問題は解決せず、結局GDPR 22条に活路を見出そうとして現在も迷走していること、Alexander Roßnagelが、1990年代から2010年代にかけて常にDatensparsamkeitを強調し、GDPRの成立に際しても、Datensparsamkeitが徹底されなかったことを指摘し、22条の自動決定規制をもっと強化すべきとGDPRの改正を主張していたこと、などなどを示す必要があるだろう。情報的自己決定権を持ち出してDatensparsamkeitを強調する者には、国勢調査判決以来の「Informationelle Selbstbestimmung」がドイツの学説でどう論じられてきたか、特に2010年のGabriele Britz「Informationelle Selbstbestimmung zwischen rechtswissenschaftlicher Grundsatzkritik und Beharren des Bundesverfassungsgerichts」(Offene Rechtswissenschaft: ausgewählte Schriften von Wolfgang Hoffmann-Riem mit begleitenden Analysen、561頁以下)がどう結論づけているかなどを挙げるのがいいし、ドイツ以外でのデータ保護法制のrelevant要件の意味が、本当に私の言うような意味のものだったのかについては、Paul SieghartやJon Bingの文献の他、英国Lindopレポートの記述と、米国FCRA 1970立法過程の公聴会の記録を参照するのがよいだろう。

*2  このことは日本にとっても他人事ではない。生成AIの台頭に伴って日本の個人情報保護委員会も一昨年、要配慮個人情報の収集についてOpenAI社に対して無理のある注意喚起を出し、保護の方向性について自民党から疑問視される事態になっていたのは、欧州とよく似た状況である。ちょうどその頃始まっていた日本の個人情報保護法の3年ごと見直しでは、この問題に対処するべく、私の提出意見が通り、生成AIの学習段階を統計処理と捉えて、要配慮個人情報が含まれてしまうことを問題視しないようにする改正が予定(昨日の新聞報道)されている。


<前の日記(2025年11月20日) 最新 編集

最近のタイトル

2000|01|
2003|05|06|07|08|09|10|11|12|
2004|01|02|03|04|05|06|07|08|09|10|11|12|
2005|01|02|03|04|05|06|07|08|09|10|11|12|
2006|01|02|03|04|05|06|07|08|09|10|11|12|
2007|01|02|03|04|05|06|07|08|09|10|11|12|
2008|01|02|03|04|05|06|07|08|09|10|11|12|
2009|01|02|03|05|06|07|08|09|10|11|12|
2010|01|02|03|04|05|06|07|08|09|10|11|12|
2011|01|02|03|05|06|07|08|09|10|11|12|
2012|02|03|04|05|06|07|08|09|
2013|01|02|03|04|05|06|07|
2014|01|04|07|09|11|12|
2015|01|03|06|07|10|11|12|
2016|01|02|03|04|06|07|08|10|11|12|
2017|01|02|03|04|05|06|07|10|12|
2018|03|05|06|10|12|
2019|02|03|05|06|07|08|10|
2020|08|09|
2021|07|08|10|12|
2022|01|04|06|12|
2023|03|
2024|03|04|07|11|12|
2025|01|02|03|04|05|06|10|11|12|
<前の日記(2025年11月20日) 最新 編集