Conflict of interest disclosure
This is meant as a disclosure of a situation that might be confusing or carry a conflict of interest. Starting Tuesday October 20th 2020, I will be joining for two months the COVID-19 contact tracing team of the Geneva Canton (at 30% FTE). This might conflict with my position as Director of PersonalData.IO. I think things are fine, but as is best practice in those situations a first step is to be transparent, particularly since I have been so far extremely vocal on Twitter. I explain here the journey that led to this situation.
Since 2018, I am Director of PersonalData.IO, a nonprofit dedicated to “making individual data rights actionable and collectively useful”. In complete coherence with this mission, I have been very visible and vocal regarding the deployment and use of systems that do not respect individual rights.
Making sense of corona
When the coronavirus crisis happened, I very quickly mobilized many actors around the PersonalData.IO community to look in depth at various aspects of the crisis. One of the main learnings, very early on (March), was of the significant impact that personal data would (need to) have in the response to the crisis. We also expanded on other themes, like the sociology questions around such a pandemic. We participated in the VersusVirus hackathon and hosted a separate workshop ourselves. I gave (lots of) talks.
Contact tracing apps
In parallel to all this, I made a serious effort to understand the risks and efficacy of Bluetooth-based contact tracing applications, and in particular the SwissCovid application since it was at the forefront of a wave of European apps. My hypothesis was always two-fold: the risks are higher and the efficacy lower than the public assessment. This felt important, as both considerations should be at the core of any proportionality assessment, a legal obligation prior to deploying such a system.
Concerning the risks of SwissCovid, I have:
- disclosed to the Swiss CyberSecurity authority, following proper procedures, a serious security shortcoming in their deployment of the SwissCovid central server;
- written two papers on risks of attacks on SwissCovid stemming from the commercial surveillance industry (the second peer-reviewed);
- questioned the legal aspects of the deployment of those apps in various countries;
- deconstructed the dependencies introduced by the Google/Apple API on the public health response;
- intervened extensively in the media on those topics.
Concerning the efficacy of SwissCovid, I have:
- made a serious effort at understanding the various sources of error in the notification process, particularly around Bluetooth at first;
- made a serious effort at understanding the benefit of the app in light of the epidemic dynamics.
Obviously, this took a huge amount of time. My strategy was always to question the security/privacy first since it felt more directly aligned with the mission of PersonalData.IO. I figured the efficacy would be more quickly addressed by others, first and foremost epidemiologists. That hasn’t really happened though: the systems collect very little data, and it is therefore very hard for epidemiologists to directly evaluate the efficacy.
Mathematics Fundamentals
At the same time, I started reading very interesting papers coming out of the mathematics community. They sounded one of three horns:
- most epidemiologists conducting predictive modeling are overlooking a significant characteristic of the virus (more stochasticity in infection patterns than most think!);
- epidemiologists conducting predictive modeling are overconfident in the quality of their own simulations (cue subprime crisis);
- epidemiologists conducting predictive modeling are misinterpreting their own results;
The last paper, combined with other information that has trickled since, essentially answers the app utility question.
Alternative
This is all a rather pessimistic outcome. Fortunately, in parallel, I came to understand progressively more and more of an alternative (many others have, but that debate has only trickled in Switzerland now). It was a long reasoning process, but by now it is easier to explain, thanks in particular to an article of Zeynep Tufekci in The Atlantic. It all derives from the first mathematical point above, which concerns the underestimated stochasticity of the virus (*).
The alternative, in its extreme formulation, is about trying to bust clusters instead of breaking chains of infections (things are never absolute — and both statements are equally unquantified — but this is more a matter of allocation of resources). Starting mid-August, I co-developed a prototype for assisting doctors with contact tracing of clusters within the Canton of Geneva. I will regret for a while things have not moved faster, but essentially I am now tasked by the Canton to deploy this system operationally. This is a paid job, but for a low percentage (30%) and short time period (two months).
This of course puts me in a dicey situation, both criticizing data collection systems and actively contributing to a very intrusive system.
One argument for joining this effort is to say that, since the data is being collected anyways, it might as well be made more useful, in order to limit what other data health authorities might be tempted to access. Another argument is to say that there is a severe health crisis looming in Europe and particularly in Switzerland, and I (=individual) can’t just spend my time criticizing (of course that argument always paves the road with good intentions).
The Conflict
I don’t think there is fully a conflict so far with PersonalData.IO’s mission: data protection rights are complex and never fully absolute, and this is precisely a moment where “data [] rights [can be] collectively useful”. In fact, that was one of the reasons to approach the Canton in the first place.
Of course, the devil is in the details, and who knows where this is going?
Still, I feel a need to make clear that my work for the Canton will be subject to the following clauses:
These clauses, which I will definitely respect, might conflict more directly with my position as PersonalData.IO’s Director. They will certainly explain my sudden silence on Twitter on certain matters.
(*) In short, the importance of superspreading events is even greater than expected. In technical terms, the overdispersion factor k is higher than expected, at around values 5–10 (instead of 2–3), such that fewer infected people actually transmit the virus than we thought and the others transmit it to more people than we thought, leading to so-called “viral clusters”.
