Procurement is progressively put in the position of regulating what types of artificial intelligence (AI) are deployed by the public sector (ie taking a gatekeeping function; see here and here). This implies that the procurement function should be able to verify that the intended AI (and its use/foreseeable misuse) will not cause harms—or, where harms are unavoidable, come up with a system to weigh, and if appropriate/possible manage, that risk. I am currently trying to understand the governance implications of this emerging gatekeeping role to assess whether procurement is best placed to carry it out.
In the context of this reflection, I found a very useful recent paper: M E Kaminski, ‘Regulating the Risks of AI’ (2023) 103 Boston University Law Review forthcoming. In addition to providing a useful critique of the treatment of AI harms as risk and of the implications in terms of the regulatory baggage that (different types of) risk regulation implies, Kaminski provides an overview of a very interesting legislative proposal: Washington State’s Bill SB 5116.
Bill SB 5116 is a proposal for new legislation ‘establishing guidelines for government procurement and use of automated decision systems in order to protect consumers, improve transparency, and create more market predictability'. The governance approach underpinning the Bill is interesting in two respects.
First, the Bill includes a ban on certain uses of AI in the public sector. As Kaminski summarises: ‘Sec. 4 of SB 5116 bans public agencies from engaging in (1) the use of an automated decision system that discriminates, (2) the use of an “automated final decision system” to “make a decision impacting the constitutional or legal rights… of any Washington resident” (3) the use of an “automated final decision system…to deploy or trigger any weapon;” (4) the installation in certain public places of equipment that enables AI-enabled profiling, (5) the use of AI-enabled profiling “to make decisions that produce legal effects or similarly significant effects concerning individuals’ (at 66, fn 398).
Second, the Bill subjects the procurement of the AI to approval by the director of the office of the chief information officer. As Kaminski clarifies: ‘The bill’s assessment process is thus more like a licensing scheme than many proposed impact assessments in that it envisions a central regulator serving a gatekeeping function (albeit probably not an intensive one, and not over private companies, which aren’t covered by the bill at all). In fact, the bill is more protective than the GDPR in that the state CIO must make the algorithmic accountability report public and invite public comment before approving it’ (at 66, references omitted).
What the Bill does, then, is to displace the gatekeeping role from the procurement function itself to the data protection regulator. It also sets the specific substantive criteria the regulator has to apply in deciding whether to authorise the procurement of the AI.
Without getting into the detail of the Washington Bill, this governance approach seems to have two main strengths over the current emerging model of procurement self-regulation of the gatekeeping role (in the EU).
First, it facilitates a standardisation of the substantive criteria to be applied in assessing the potential harms resulting from AI adoption in the public sector, with a concentration on the specific characteristics of decision-making in this context. Importantly, it creates a clear area of illegality. Some of it is in line with eg the prohibition of certain AI uses in the Draft EU AI Act (profiling), or in the GDPR (prohibition of solely automated individual-decision making, including profiling — although it may go beyond it). Moreover, such an approach would allow for an expansion of prohibited uses in the specific context of the public sector, which the EU AI Act mostly fails to tackle (see here). It would also allow for the specification of constraints applicable to the use of AI by the public sector, such as a heightened obligation to provide reasons (see M Fink & M Finck, ‘Reasoned A(I)dministration: Explanation Requirements in EU Law and the Automation of Public Administration‘ (2022) 47(3) European Law Review 376-392).
Second, it introduces an element of external (independent) verification of the assessment of potential AI harms. I think this is a crucial governance point because most proposals relying on the internal (self) assessment by the procurement team fail to consider the extent to which such approach ensures (a) adequate resourcing (eg specialism and experience in the type of assessment) and (b) sufficient objectivity in the assessment. On the second point, with procurement teams often being told to ‘just go and procure what is needed’, moving to a position of gatekeeper or controller could be too big an ask (depending on institutional aspects that require closer consideration). Moreover, this would be different from other aspects of gatekeeping that procurement has progressively been asked to carry out (also excessively, in my view: see here).
When the procurement function is asked to screen for eg potential contractors’ social or environmental compliance track record, it is usually at arms’ length from those being reviewed (and the rules on conflict of interest are there to strengthen that position). Conversely, when the procurement function is asked to screen for the likely impact on citizens and/or users of public services of an initiative promoted by the operational part of the organisation to which it belongs, things are much more complicated.
That is why some systems (like the US FAR) create elements of separation between the procurement team and those in charge of reviewing eg competition issues (by means of the competition advocate). This is a model reflected in the Washington Bill’s approach to requiring external (even if within the public administration) verification and approval of the AI impact assessment. If procurement is to become a properly functioning gatekeeper of the adoption of AI by the public sector, this regulatory approach (ie having an ‘AI Harms Controller’) seems promising. Definitely a model worth thinking about for a little longer.