Leveraging technically feasible practices

Documentation Practices

Leveraging established documentation practices such as “model cards” and “datasheets” was a recurring topic throughout the interviews (Candidates 4, 5, 7). The establishment of these practices would “enable better oversight mechanisms on the market as well as create a kind of economic incentive for market actors to improve the quality of documentation of existing practices in order to comply with the AI Act” (Candidate 5).

Auditing

Auditing emerged through most interviews as a key established methodology that “will be at the core of compliance with the AI Act” (Candidate 5) and that can address the current issues of understanding the impact of the systems through the development phase.

Assurance Practices

Another important topic that emerged through two interviews is the “necessity of establishing assurance practices” (Candidate 3), which are “methodologies to improve oversight mechanisms by verifying the validity of the claims made” (Candidates 3, 7). Assurance practices provide a framework to conduct “controls to ensure a certain technology does not produce wrong results” (Candidate 7).

Data Quality

Moreover, developing methodologies to assess data quality emerged throughout three interviews. An interviewee suggested that a “solution could be to apply to data the same regulations and requirements that are enacted in the supply chain to disincentivise the development of practices that explore how the data was sourced, similarly to what happens in the trade of raw materials” (Candidate 2). Data quality practices for instance should not merely cover the statistical properties of the training sets but also “guarantee that no infringements to human rights were conducted in the sourcing phase” (Candidate 2).

Cross-disciplinary training and collaboration

An important gap that the interviews identified is the lack of cooperation between the relevant stakeholders in AI systems; there is a lack of interdisciplinarity, both in research and jurisdiction (Candidates 1, 2, 4, 5, 6, 9, 10). The development and deployment of AI is “a multi-stakeholder subject, so one needs somebody who knows about these legal issues and how to translate them” into requirements for an AI product, this person should also know how legal departments work” (Candidate 6). Another interview candidate explained the necessity to: “I guess it would be good to reshape the organisation that every team has then one legal guy as an interface expert. In order to have the linkage in each team to all of these different regulations.” (Candidate 10)

Tooling and systemic practices

Finally, one interview highlighted the “necessity of applying tools implemented at a system level” (Candidate 6). In general software systems, part of the quality assessment practices is implemented at the system level, documentation is mostly generated automatically from code, tests are executed by the Continuous Integration system, and versioning is an integral part of the daily work, to give a few examples.