What Does it Mean To Make Data Public?

Max Schrems filed a fascinating case, and leaves me puzzling over inference, again.

Black and white stylized image of a person manifestly making personal data public (DALL-E)

Those of you who have read this blog might recall that I see Max Schrems as largely as a chaos agent doing his damnedest to break the internet.¹ Others consider him an indefatigable privacy advocate, and I’m sure the answer falls somewhere in the middle of those two extremes. I’ve written about him here, for example:

Welcome to the Splinternet

Schrems is a frequent complainant before the Data Protection Commission in Ireland, and a frequent litigant before various courts, including the Court of Justice (CJEU). Last week, Mr. Schrems was arguing before the CJEU again, in the case of Schrems (Communication de données au grand public), C-446/21. This latest case, like many before it, is a sprawl of a complaint that substantively touched on four legal issues, or in CJEU-speak, ‘questions referred’. The first two covered lawfulness of Facebook/Meta’s processing and data minimisation. Those are, IMHO, not very interesting and worth skipping over. Instead, I want to focus on his last two questions he posed, which I’ve taken the liberty of summarizing into plain English from the legalese provided:

1. Whether controllers like Facebook are permitted to target advertising based on sensitive/special categories of data like a person’s political opinions or sexual orientation; and

2. Whether a person makes sensitive personal data (such as their sexual orientation) public for all, including targeted advertising, merely by sharing it certain contexts.

At this point, it’s important to note that Schrems is gay, and reasonably open about this fact. He has discussed his sexuality on public panels, usually in the context of the importance of data protection and fundamental rights. Thus, his sexuality is not a secret, but it’s also not something that he broadcasts a great deal. I wasn’t able to find any press where he shared this information. And it’s definitely not something he ever posted on or shared with Facebook. And yet, Facebook somehow figured this out, and began showing him ads and recommending connections based not only his sexual orientation, but his political opinions as well.

Schrems did what he usually does, and brought a case against Facebook in Austria. The Austrian court rejected his argument, and dismissed the action, on the grounds that no infringement of his rights under Article 9 GDPR had occurred, because he had manifestly made details of his sexuality public by discussing that fact on these public panels.

Article 9 covers how controllers and processors handle ‘special categories of personal data’. Here we’re talking about things like a person’s race or ethnic origin, religious beliefs, political opinions, sex life, and importantly, sexual orientation. Article 9 generally takes the view that a controller like Facebook cannot process sensitive information about a person without having a specific legal basis for doing so. There are a few different legal bases — ‘explicit consent’ of the data subject is one; another is when information is “manifestly made public by the data subject.” Facebook argued, and the lower court agreed, that Facebook hadn’t done anything wrong when it presented ads and suggested groups based on his sexual orientation and political affiliations, because Schrems had manifestly made this information public.

Easy in Theory, Wickedly Hard in Practice

This is a fascinating case, and touches on a few really interesting themes and potential privacy harms. And while I don’t usually predict decisions (it tends to anger the Law Gods when you do), I do think I know where the court is likely to go on this one: they’re probably going to agree with Schrems, and not Facebook in this case. I’m pretty confident in fact, because we’ve been here before, in the Meta v. Bundeskartellant decision. In that case, Facebook/Meta made a similar argument related to so-called Off-Facebook data.

Meta's Wakeup Call, and Big Tech's New Reality?

I wrote that:

the Court of Justice noted that Meta’s ability to link special categories data (based on Off-Facebook activity) to a user’s Facebook account, “may, in certain cases, reveal such information” about a user, without them directly providing it to Meta/Facebook (Para. 72). Therefore, this must be considered processing of special category data by Meta. Given that, there was no way to conclude that the user’s actions done outside of Facebook (e.g., by purchasing a product, or reading, liking, or sharing a page) meant that they intended to ‘explicitly and by a clear positive act’ to make that data public. Nor are they explicitly consenting to share that information with Meta (Para. 77-79).

Still, while the case might be predictable in principle, it might not be intuitively obvious to folks what such decision could mean in practice. Some questions to ponder:

• What’s the threshold of disclosure that needs to be met before something is ‘manifestly made public’? Do search results count?

• Does it require that the person in question be a public figure? Is there a different threshold for say, Schrems (who I’d argue, is in fact a public figure, at least in Europe) than me?

• Is there a distribution or reach threshold that needs to be met? If I share that I’m a catgirl on my Substack, have I manifestly made that detail public? Have I made it public to Facebook?² What if I author a piece on the benefits of being a catgirl in the New York Times?

• Do we have different standards when it comes to ‘processing’ ? For example, is an inference made by a machine learning model that I can be bucketed into some category the same as a targeted ad? What if I never learn about the inference? What about Google’s new Privacy Sandbox initiative? That’s arguably making loads of inferences at a group level by bucketing users into topics and categories. This is better than cookies, but arguably might still create challenges based on the output of the CJEU decision.

• Does the type of sensitive information matter? For example, is exposure of one’s race different than, say a person’s sex life?

Obviously, if I selectively tell friends and loved ones that I’m a catgirl, it would feel like a gross violation of my autonomy if Google started blasting ads for kitten kink sites and gear, or Facebook recommended me to catgirl appreciators. Exposure of that information could have some very serious consequences to me, including substantial chilling effects on my future choices and behaviors. As Professors Dan Solove and Danielle Citron note in their seminal law review article Privacy Harms:

Chilling effects have an impact on individual speakers and society at large as they reduce the range of viewpoints expressed and the nature of expression that is shared. … Consider the impact of news that the gay dating app Grindr had shared subscribers’ HIV status with analytics firms. Subscribers expressed profound dismay. Individuals told the press that they would no longer share that information on that app or any dating app—it was simply not worth the possibility that employers or others could find out their HIV status and hold it against them.

There’s no question that harms to our autonomy and choices cause us to hold back. For example, if I share nude photos with a significant other, and they leak those photos to a WhatsApp group chat, my autonomy has been violated. I may choose in the future never to share that information with anyone else out of fear of something similarly bad happening. Not only would that impact my rights to autonomy and choice, but it could also negatively impact my relationships with others. I might trust people less. It would be appalling then, if a court determined that such a violation nonetheless meant that I made my nudes or other confidential information public.

But, we also have to consider the very real question of how we enforce this sort of thing practically? It’s all well and good to declare that Max Schrems has not manifestly made his sexuality public. It’s an easy academic exercise to state that such an outcome would be a net good, and tech companies just shouldn’t leak that sort of thing in any way, but it’s another thing entirely to implement that in the bowels of Facebook’s targeted ads or OpenAI’s LLM. We are increasingly relying on AI, algorithms, and automation, even though we don’t have a goddamn clue about how to keep said machines from intuiting or inferring details about us.

In the Bundeskartellant article, I offered a number of inadvertent inferences that are already being made about us daily based on fairly benign stuff — Google Maps and location information can intuit your sexuality and recommend gay bars, your Substack subscriptions might lead to an inference about your religious or political affiliations, race or ethnic origin can be inferred from what you watch on Netflix. The Budeskartellant case points all of this out, but it throws a wrench in the machinery by also preventing Facebook/Meta from relying on other common legal bases for processing special or sensitive data (such as explicit consent or contract). The end result seems to be leading towards just don’t infer sensitive information, without any clear guidance about how.

The AI Act Adds More Complexity

The AI Act also doesn’t offer any help here. The Act is set to be finalized any day now, explicitly calls out inference dozens of times. For example, in Recital 6:

A key characteristic of AI systems is their capability to infer. This inference refers to the process of obtaining the outputs, such as predictions, content, recommendations, or decisions, which can influence physical and virtual environments and to a capability of AI systems to derive models and/or algorithms from inputs/data.

The AI Act bans inference about emotions and ‘intentions’ of an individual based on biometric and other special categories data, including what might be seen as benign details like hair color, tattoos, and interests which go well beyond what’s already recognized under the GDPR. The law is also riddled with some rather squishy exceptions.³

To reign in the worst abuses, the law sets a high bar on so-called “High-Risk AI Systems” (as defined under Title III of the Act and Annex III). In addition to outright bans in some cases, permissible High-Risk systems now have new obligations to meet. Most of these are pretty reasonable (providing details about the AI itself, including some insights as to how it arrives at decisions, what it’s designed for, details about the input data used, and the identity of the provider), but other obligations might prove especially tricky to implement. For example, the obligation to document things like the types of individuals or groups ‘likely to be affected by use of the AI system’, as well as “known or foreseeable circumstance[s], related to the use of the high-risk AI system in accordance with its intended purpose or under conditions of reasonably foreseeable misuse …”.⁴

This language, and the related obligations, strike me as vague and in some places, completely infeasible. As I think I’ve hammered home a few times now, we humans are shit at identifying our own inferential biases, or inferences made by relatively dumb algorithms and ‘smart’ devices. Applying that to generative AI and other forms of machine learning that are designed to infer things seems like it’ll be even more of a challenge. I really don’t know how all of this will work. And unless the CJEU has a firmer idea of how to unravel this mess, I suspect in the end, the Schrems decision is unlikely to offer much clarity either.

Note: I am probably going to move off of Substack in the near future. It shouldn’t have much of an impact to subscribers (I think I’ll be moving to Medium.com) but will also continue cross-posting on LinkedIn if you’re not a Medium.com fan.

1. I should note that I generally don’t agree with Max Schrems, or the organization he remains heavily involved with Noyb, on much of anything. In fact, I personally think Max is a bit of an egotistical asshole. So, when I came across his latest suit against Facebook Ireland, I read it assuming I’d have something new to rail about. Except, this time, he’s actually got a point.

2. FYI to Meta and anyone else: I am not, in fact a catgirl. Not that I’m yucking someone else’s yum, here, but that’s just not my kink.

3. See Recital (8a) of the AI Act: The notion of emotion recognition system for the purpose of in this regulation should be defined as an AI system for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data. … It does not include physical states, such as pain or fatigue. It refers for example to systems used in detecting the state of fatigue of professional pilots or drivers for the purpose of preventing accidents. It does also not include the mere detection of readily apparent expressions, gestures or movements, unless they are used for identifying or inferring emotions.

4. Article 13 (3)(b)(iii), AI Act.