Skip to content

Marres 2025 – How Tech Trials Put Society to the Test

In brief

Marres argues that trials of AI technologies in society – automated vehicles on UK streets, police facial recognition pilots, and DeepMind's Streams app at the Royal Free Hospital – turn society into a testing ground (pp. 1–2) [U]. These are ‘trials of implicitness’: they inscribe compute-intensive technologies into the background of everyday life and weaken the conditions for accountability (p. 1) [U]. She traces this to the paradigm of experimental innovation and calls it the ‘infrastructuralisation’ of testing (pp. 1, 8–9) [U]. As a counter-move she proposes a new ‘Durkheim test’: evaluating technologies against societal objectives by creating experimental situations (pp. 1, 14–15) [U].

Research question and main claim

  • Main claim: continuous technology testing in society changes the background conditions of social and public life in ways that limit everyday actors' capacity to practise accountability (p. 3) [U].
  • Paradoxically, even as tech trials insert testing technologies into society, they make innovations unavailable for critical evaluation by actors in society (p. 3) [U].
  • Seeing this requires shifting focus from AI's capabilities to the mechanisms by which it is introduced: tests, pilots, testbeds and trials (p. 3) [U].
  • Tech trials are not ‘trials of explicitness’ but ‘trials of implicitness’ (p. 13) [U].

Theoretical perspective

  • STS and sociology of testing: ‘experimentation beyond the laboratory’ (Van de Poel et al.), testbeds (Laurent and Tironi, Engels et al.), prototyping society (Dickel) (pp. 4–6) [U].
  • Muniesa and Linhardt's ‘trial of explicitness’, which she inverts (p. 9) [U].
  • Infrastructure studies (Harvey et al., Larkin) (p. 9) [U].
  • Critics of AI and accountability: Rouvroy, Amoore, Hildebrandt (pp. 2–3, 7, 13) [U]. She disagrees with Rouvroy that algorithmic systems ‘spare the burden of testing’ (p. 13) [U].
  • Susan Leigh Star's ‘Durkheim test’ (1989) (p. 14) [U].
  • Lukes on power, Schutz, Heller (pp. 14, 16) [U].

Empirical material and context

  • UK cases: automated vehicles on streets since at least 2016, including the Greenwich testbed; DeepMind's Streams app at the Royal Free Hospital (2016), where the company received medical records of 1.6 million patients without their knowledge or consent; police facial recognition trials (Met, South Wales); facial recognition in Scottish schools (2021) (pp. 1–2, 11–13) [U].
  • Sanctions: the ICO disciplined the Royal Free in 2017; the South Wales Police pilots were declared unlawful by the Court of Appeal in 2020 (p. 2) [U].
  • A freedom of information request to the Met about facial recognition cameras, which was refused (p. 11) [U].
  • Interviews with engineers and experts (pp. 7, 17, n. 3), and an earlier analysis of YouTube videos of street trials (p. 11) [U].

Method

  • Conceptual argument illustrated with UK cases, drawing on the author's Leverhulme fellowship research on intelligent vehicle testing (p. 17) [U].
  • She defines AI minimally: she follows public discourse in calling something ‘AI’ (p. 17, n. 2) [U].

Key concepts

Concept How the source defines or uses it Page U/T
tech trials / experimental innovation Introducing new technologies into society early in their development, relying on feedback from society 4 U
infrastructuralisation of testing Testing inscribed into everyday environments as part of the underlying apparatus 1, 9 U
trial of implicitness Trials that render innovation and evaluation implicit and weaken accountability 1, 13 U
trial of explicitness (Muniesa and Linhardt) Moments where relations between innovation, society and state become manifest 9 U
emergence testing Testing that triggers situational accountability by making unknown attributes explicit 13–14 U
Durkheim test (Star) Testing systems against their ability to meet community goals 14 U
instrumentalisation of society Society reduced to an environment serving innovation 5 U
assetisation Testbeds turn populations, communities and places into assets 5, 15 U

How privacy and private data are understood

  • Not a central concept, but data appear as part of how tech trials work: data from everyday environments are fed back to corporate data centres (p. 8) [U].
  • In the Royal Free case, patient data were moved into a tech company's infrastructure, and a for-profit third party took over management of patient data (pp. 2, 11) [U].
  • Surveillance is mentioned among the societal risks of tech trials (p. 2), and automated vehicle data are used for surveillance (p. 2) [U].
  • Data protection appears via the ICO's sanction against the Royal Free for breaching data protection laws (p. 2) [U].
  • Privacy is thus understood mainly indirectly: as part of accountability and of how environments become monitorable (pp. 8, 15–16) [T].

Arguments and findings

  • Experimental innovation configures society as stakeholders, as an experimental population and as a test environment (pp. 5–6) [U].
  • Testbeds resemble free-trade zones: regulation is relaxed, and environments are equipped with data infrastructure (p. 5) [U].
  • Machine learning is said to be a contextual method, but tech trials do not merely capture data; they modify environments to enable their own functioning (pp. 8–9) [U].
  • Most tech trials operate in the background and go unnoticed; even when disclosed, they do not necessarily enable engagement (p. 10) [U].
  • Tech trials move evaluation tasks in law enforcement, medicine and traffic into privately managed computational systems (p. 11) [U].
  • In Greenwich, the testbed led to fences, CCTV every 20 metres or so, and the exclusion of cyclists (pp. 12–13) [U].
  • Consultation and participation initiatives on AI often lift AI out of its social context as an object of deliberation rather than social experimentation (p. 15) [U].

Limitations the source itself mentions

  • She largely leaves generative AI such as ChatGPT out of account (p. 17, n. 2) [U].
  • Trials of implicitness may become trials of explicitness over time, as with counter-testing (p. 11) [U].
  • It is remarkably difficult to specify a method for evaluating such technology against society's collective objectives (p. 16) [U].

What the source does not cover

  • Norway or the Nordic countries [T].
  • Government documents as such, although the cases involve public bodies (NHS, police) [T].
  • Detailed legal analysis of data protection [T].

Possible relevance for the thesis

  • One of the few sources in the list that directly concerns the public sector's use of private data (health data, facial recognition) in practice [T].
  • The concept of trials of implicitness may be a lens for asking whether Norwegian documents make the state's data use visible or invisible to citizens [T].
  • Pilots, testbeds and ‘regulatory sandboxes’ may also appear in Norwegian documents; the article offers a critical vocabulary for these [T].
  • The UK context must be taken into account [T].

Connections to other sources

Key quotes

“Tech trials turn society into a testing ground, even as they dismantle the conditions for the evaluation of technology from the standpoint of social life.” (p. 1)

“Paradoxically, even as tech trials involve the insertion of testing technologies into environments in society, they render technological innovations unavailable for critical evaluation, or ‘testing’, by actors in society.” (p. 3)

“But this displacement of patient data into big tech company infrastructure remained invisible from within the care practices in relation to which it occurred” (p. 11)

“They constitute not trials of explicitness but of implicitness.” (p. 13)

“The Turing test should be replaced by the “Durkheim test,” that is, systems should be tested with respect to their ability to meet community goals” (p. 14)

Open questions for Rebekka

  • Do the Norwegian documents discuss pilots, testbeds or regulatory sandboxes for data use? How are they described?
  • Do the documents make the state's data use visible to citizens, or does it remain in the background?
  • Is the Royal Free case comparable to Norwegian health data cases you know of from your documents?
  • Is the Durkheim test – evaluating against societal goals – something you see in the documents, or its absence?