Christian Thönnes, research assistant at the Department of Public Law of the Max
Planck Institute for the Study of Crime, Security and Law in Freiburg.
Christian Thönnes ist wissenschaftlicher Mitarbeiter in der Abteilung Öffentliches Recht des Freiburger Max-Planck-Instituts zur Erforschung von Kriminalität, Sicherheit und Recht.
13 July 2021 was a
potentially fateful day for the balance of privacy and security in the European
Union. The Court of Justice of the European Union (CJEU) held an oral hearing
in its preliminary ruling procedure C-817/19. Following a legal challenge undertaken
by the Belgian NGO Ligue des Droits Humains, the Constitutional Court of
Belgium had submitted ten preliminary questions to the CJEU (arrêt
n°135/2019 du 17 octobre 2019). These questions concern the interpretation
of Directive (EU)
2016/681 of 27 April 2016 on the use of passenger name record (PNR) data for
the prevention, detection, investigation and prosecution of terrorist offences
and serious crime (in short: PNR Directive) and its compatibility with EU
primary law. The hearing did offer important insights into the court’s
thinking. Especially judge-rapporteur Mr. von Danwitz asked the EU Commission
many critical questions which do not appear to portend good news for proponents
of this unprecedented surveillance measure. After having worked on a similar case for the Berlin-based
strategic litigation NGO Gesellschaft
für Freiheitsrechte, I attended the oral hearing. I am providing this entry
as a brief summary and analysis.
This report
hopefully encapsulates the high stakes of the case in question: The PNR
preliminary ruling procedure could result in a landmark decision, laying the
doctrinal groundwork not only generally for the mass retention of travel data
in the name of defending security, but more specifically for the deployment of
self-learning algorithms in order to pre-emptively detect presumptively
suspicious movement patterns. As such, the CJEU has to deal with one of the
first EU-wide, large-scale use cases of predictive policing. If the court were
to essentially approve of this paradigm shift, a radical expansion of
technology-driven surveillance to all sorts of ordinary human behavior,
regardless of individual prior suspicion or imminent threat, could ensue. In
its national transposition law, the Belgian parliament, for instance, decided
to expand the PNR’s scope of application beyond only aviation to international
trains, busses and ferries (Doc. parl., Chambre, 20152016, DOC 54-2069/001,
p.7).
The PNR Directive: An unprecedented tool of
technology-driven mass surveillance
The PNR Directives
obliges EU Member States to require air carriers to transmit a set of data for
each passenger to national security authorities, so-called “Passenger
Information Units”. PNR (passenger name record) datasets contain unverified
information provided by passengers to the airlines or to a travel agency in
order to facilitate the processing of each flight. The exact content of these
PNR datasets depends on the commercial needs of each airline. All necessary
data categories to be transmitted are defined in the PNR Directive’s Annex I.
They encompass clearly-defined data items, for example, date of birth, details
of accompanying persons, travel itineraries, or baggage information, but also
loosely-defined items such as “general remarks” made by the aviation staff.
After their
reception by the PIUs, PNR datasets are then automatically compared against
databases “relevant for the purposes of preventing, detecting, investigating
and prosecuting terrorist offences and serious crime” (Art. 6 § 3 letter a), as
well as against “pre-determined criteria” (Art. 6 § 3 letter b). The latter are
used “to identify ‘unknown’ suspects” (EU
Commission PNR Directive Proposal, SEC(2011) 132, p. 12). What exactly
these pre-determined criteria are, is not really defined in the PNR Directive’s
text. Art. 6 § 4 only lays out that they ought to be “targeted, proportionate
and specific”, as well as not based on “a person's race or ethnic origin, political
opinions, religion or philosophical beliefs, trade union membership, health,
sexual life or sexual orientation”. Their general purpose, however, is to
extrapolate suspicious patterns passengers’ flight behaviors. Research (Korff, Passenger Name Records,data mining
& data protection:the need for strong safeguards, p. 4; Rademacher,
Predictive Policing im deutschen Polizeirecht, AöR 2017, 366, 410-415; Sommerer,
Personenbezogenes Predictive Policing, S. 96-98), policy
papers and the
GFF generally discuss “pre-determined criteria” as a likely product of
self-learning algorithms
According to Art. 6
§ 5, once a “hit” occurs – meaning an automatically generated match between a
PNR dataset and either a database or pre-determined criteria – it is “reviewed
by non-automated means”. Verified hits can then be transmitted to other
national competent law enforcement authorities on a “case-by-case basis” (Art.
6 § 6), or to other Member States (Art. 9) or, under additional conditions, to
third countries (Art. 11). These authorities can then decide to take further
action under their respective national law. The PNR Directive itself only
requires the collection of PNR data for all flights entering and exiting the EU
but recognizes Member States’faculty to extend its scope to intra-EU flights. Almost
all Member States proceeded to do just that. PNR data are stored for six months
in raw form and, after that, for another four-and-a-half years in a
(reversibly) pseudonymized form (Art. 12).
EU Member States scramble to defend the PNR Directive’s
proportionality
As always, the
hearing opened with opening statements by the parties, Member States and EU
institutions. This turned out to be a rather one-sided affair, given the fact
the plaintiff’s representative faced a united front of eleven delegations from
Member States and EU institutions, all in favor of the PNR Directive. They all
defended against the charge that the PNR Directive was a disproportionate
affair mainly by providing anecdotal evidence of its usefulness and by
characterizing the relevant interferences with fundamental rights as not
particularly severe.
First of all, many
Member States pointed to the CJEU’s recent Opinion
1/15 of 26 July 2017 on the Draft PNR Draft agreement between Canada and the
European Union. In that opinion, they argued, the CJEU had not generally
prohibited the mass retention of PNR data; neither did it, in principle, object
to a five-year retention period. And, they asked, were PNR data not much safer
within the realms of the PNR Directive than in Canada, they suggested, as data
stored on European servers remain protected by the GDPR and other legal
regimes? Member States’ representatives failed to mention, however, that the
CJEU considered the Draft PNR Agreement to be incompatible with Articles 7, 8 ,
21 and Article 52 § 1 CFR. It did so due to the Draft Agreement’s lack of
clear and precise criteria, substantive and procedural, for the automated
processing of PNR data. It also criticized the lack of a required link, such as
new circumstances of threat, between the Agreement’s stated objective (averting
terrorist threats) and the prolonged retention and processing of PNR data (n°
232). As pointed out above, the PNR Directive perpetuates this dissatisfying
state of vagueness (see its Art. 6 § 4) and even extends it beyond the Draft
Agreement’s scope of serious cross-border crimes to any crime of some (but not
necessarily exceeding) gravity, such as fraud (Annex II, Number 7) or the
facilitation of unauthorised entry and residence (Number 11).
Second, Member
States pointed to evidence of the PNR Directive’s effectiveness in combatting
crime. The Belgian government, for example, mentioned the interception of human
trafficking victims on Belgian airports. The French government lauded the PNR
Directive for enabling it to stop and arrest a person running an illegal
prostitution ring who was en route to Bangkok, accompanied by several minors.
The Polish government talked about detecting illegal imports of cigarettes from
Poland to Germany, and about Ukrainian nationals attempting to use flights for
illegal immigration. It was striking, however, that all this evidence remained
purely anecdotal. Just like the EU Commission in its
Evaluation Report, Member States never provided any detailed statistical
evidence on the PNR Directive’s contribution to the prevention, detection,
investigation or prosecution of terrorism or serious crime. It also remained
unclear, throughout the hearing, whether and, if so, how many of the cases
provided by Member States could have been detected by previously existing data
processing methods.
Third, most Member
States contended that the PNR Directive did not result in a particularly severe
interference with the right to respect for private and family life (Article 7
CFR), and the right to the protection of personal data (Article 8 CFR). The PNR
Directive, they opined, prohibited the processing of sensitive personal data
(see Article 13 § 4 and Recital 37). Some Member States, like Germany, Ireland,
Spain, and Cyprus, leaned into that claim by forcefully asserting that all
pre-determined criteria in use were assembled by humans, not by self-learning
algorithms. Their opening statements were characterized by emphatic
renouncements of such algorithms as processing tools. The Dutch government even
went so far as to claim that the PNR Directive prohibited the use of artificial
intelligence or self-learning algorithms for the creation of pre-determined
criteria. This assertion, however, appears rather implausible at the least. In
academic literature, the PNR Directive was seen as a blueprint for the use of
self-learning algorithms (see references above). The Directive contains no
explicit prohibition of artificial intelligence. The European Parliamentary
Research Service mentions the PNR Directive in its
report on the use of “Artificial Intelligence at EU borders” (pages 18-19).
Artificial intelligence only works when fed with gigantic amounts of data. That
is why Article 6 § 2 letter c of the PNR Directive allows “analysing PNR data
for the purpose of updating or creating new criteria to be used in the
assessments carried out“ [through pre-determined criteria].
Judge von Danwitz’ questions
Much more
interesting than the (quite repetitive) jubilant opening statements was the
oral hearing’s second part: questions by the court. The majority of this part
was characterized by an exchange between judge-rapporteur Professor von Danwitz
and the EU Commission’s representative. Judge von Danwitz structured his
questioning into four topics: The statistical reliability (or fallibility) of
the PNR system (1), the severity of fundamental rights interferences produced
by the PNR Directive (2), discriminatory effects of the PNR system (3), and the
overall proportionality of the system (4).
(1) Judge von
Danwitz began by referencing concerningly high false-positive rates mentioned
in the Member States’ respective statements. In its Evaluation
Report, the EU Commission writes on page 28 that, in 2019, “0.59% of all
passengers whose data have been collected have been identified through
automated processing as requiring further examination”. Only 0.11% of all data,
however, were verified by humans and then transferred to law enforcement
authorities. This suggests, as Judge von Danwitz emphasized, a false-positive
rate of more than 81 %. Moreover, it remains uncertain whether the remaining 19
% of datasets were legitimately process, thus making a definitive assessment of
the full false-positive rate impossible.
Referencing the COVID pandemic in questioning the PNR system’s
suitability for its intended purpose, Judge von Danwitz quipped: “If a PCR test
operated with a sensitivity of 19%, I doubt it would be welcomed with open
arms” (The original words were spoken in French. Translations are my own).
While these numbers may intuitively feel small, consider this: When the German
transposition law was adopted, the German government expected roughly 170
million yearly affected flight passengers (Gesetzesbegründung, BT-Drs.
18/11501, S. 23). This would mean that, in Germany alone, 187,000 people could
be subjected to false automated suspicion every year.
In fact, these EU
numbers are no anomaly. Similar false-positive rates have been reported by
Member States: In a GFF case before the Administrative Court of Wiesbaden
(docket number 6 K 806/19.WI), the Bundeskriminalamt, which functions as the
German PIU, reported that 31,617,068 processed PNR datasets yielded 237,643
automatic matches. After human review, only 910 matches remained, resulting in
a false-positive rate of 99.6%. Out of these 910 matches, 396 investigations
went dry because the affected flight passengers were not identical with the
actual wanted persons –leading to even more serious false suspicions. This
rate, mind you, only pertains to database matches – one is left to imagine the
error rate of the much more volatile matching procedure against pre-determined
criteria.
As Judge von
Danwitz pointed out, there is a statistical reason for such high rates: Base
rate fallacy. Base rate fallacy denotes the phenomenon that when you are
looking for very rare incidents in very large datasets, even extremely
sophisticated detection tools will likely yield more false-positives than
true-positives. Out of all EU flight passengers, extremely few will be actual
terrorists or serious offenders – European law enforcement is looking for the
proverbial needle in a haystack. Adding more hay to the stack will not allow
them to find more needles – needles remain just as rare and elusive.
Confronted with
this criticism, the EU Commission pointed to the legislator’s limited
responsibility for mathematical limits: "There are mathematical limits and
errors, but the legislator is not required to conduct mathematical demonstrations.".
(2) Judge von
Danwitz then proceeded to refute the Member States’ and EU institutions’ claim
that, since the PNR Directive does not result in the processing of particularly
sensitive data, it does not constitute particularly severe interferences with
Articles 7 and 8 CFR. In so doing, he drew an explicit comparison with
telecommunications data which were the subject of Digital
Rights Ireland, another landmark CJEU decision on mass data retention.
While acknowledging that telecommunications data may per se contain more
sensitive information than passenger data, he pointed out that, when
determining the severity of interferences, one must also take into account the
scale and method of processing: Firstly, while Directive 2006/24/EC intended
that the majority of telecommunications data be accessible but remain
unscrutinized, the PNR Directive provides for the automated analysis of every
single PNR dataset. Secondly, Judge von Danwitz emphasized that the deployment
of data mining through self-learning algorithms intensified the severity of the
interference (thus also casting aside the renouncements put forth by many
Member States).
The EU Commission
agreed with Judge von Danwitz’ analysis in principle (“Oui, certainement, la
gravité de l’ingérence est déterminée dans la manière où le Data Mining se
fait.") but contended that there were different degrees of Data Mining
("Il y a Data Mining et Data Mining."). They claimed that the
safeguards included in the PNR Directive rendered the Data Mining at hand
rather minor.
(3) Judge von
Danwitz then turned to the lack of clear criteria for the individual review of
automated matches by non-automated means, as per Article 6 § 5. He pointed out
that this vagueness opened room for discrimination. When the EU Commission
responded that discrimination was prohibited under Article 6 § 4, Judge von
Danwiz replied that indirect discriminations certainly remained possible – which
the EU Commission admitted (“un risque de discrimination indirecte existe
toujours"). Judge von Danwitz then asked why the EU legislator did not
include more provisions in the text in order to mitigate the risk of indirect
discrimination – given the extremely high false-positive rate of over 80
%. The EU Commission responded that there were limits to the specificity that
can reasonably be expected from any legislator (“tout législateur a des limites
lorsque l’on doit réglementer une activité minutieuse et détaillée").
Responding to this, the European Data Protection Supervisor proposed a reversal
of the burden of proof as a possible solution, but this was rejected by the EU
Commission because, according to them, that would presuppose that “everyone is
automatically a victim”.
(4) Judge von
Danwitz’ fourth line of questioning turned on the overall proportionality of
the PNR system. His questions mainly focused on the (lack of a) link between
the occasion for mass data retention – taking a flight – and the PNR
Directive’s stated purpose – combatting terrorist offences and serious crime.
The lack of clear criteria buttressing this link was one of the CJEU’s main
sources of concern in its Opinion 1/15
(see n°217). The EU Commission responded that criminals specifically use the
convenience of international air travel to orchestrate their crimes. Judge von
Danwitz proceeded by pointing out that the notion that locations and behaviors
suitable for crime should be subjected to mass surveillance was stretchable to
an almost unlimited extent. “Why not rock concerts?”, he asked. “Why not museum
visits?”. Surprisingly, the EU Commission basically agreed with him, saying
that yes indeed, rock concerts could be prone to drug-related offenses (“I
don’t have any police experience, but I could imagine that there could be much
drug-related crime occurring at rock concerts.”).
Lingering doubts regarding the PNR system’s
proportionality
By no means can
this report paint a full picture of what was said during the oral hearing or
what is to be considered when assessing the mass surveillance of flight
passengers. For example, other questions raised by the Advocate General
Pitruzzella concerned the vagueness around which databases would be “relevant”
and could therefore be used for comparison (von Danwitz asked whether Facebook
databases could be used), whether the unanimous extension of the PNR
Directive’s scope to intra-EU flights was warranted or excessive, and whether
the five-year-retention period was disproportionate. Advocate General
Pitruzzella also inquired about external oversight of pre-determined criteria
and whether false-positives were systematically used in order to improve
algorithms (to which some Member States replied in the affirmative).
In my opinion
though, the hearing shone a light on the PNR Directive’s manifold
constitutional weaknesses. The EU Commission and Member States did not succeed
in dispelling my lingering doubts about its underdeterminacy, its questionable
suitability and effectiveness, its unchecked potential to produce large-scale
discrimination. But chief among these weaknesses is the Directive’s sheer
excessiveness: It is just disproportionate to take ordinary human behavior and
use it as an opportunity to unleash an unprecedented degree of technology-fueled
surveillance upon hundreds of millions of European flight passengers and to
sift through mountains of useless, potentially discriminatory data, just to –
maybe, kind of – detect a handful of criminals.
Photo credit: Juke
Schweizer, via Wikimedia
Commons
