In collaboration with UNESCO's Ethics of AI team and the AI Expertise Centre of Digital Flanders, the Knowledge Centre for Data & Society used a co-creation process to develop a shared vision on AI governance within the Flemish Government. This work is part of UNESCO’s AI-Ready Flemish Administration-project and is generously financed through SG REFORM’s Technical Support Instrument. In this blog, we outline the key findings.
From hallucinated content that is copied without question, to fears of privacy breaches and discriminatory algorithms on learning and job platforms, to apocalyptic predictions that AI will wipe out humanity: the news about artificial intelligence in Flanders does not always make for cheerful reading.
Yet the remedy for all these problems is well known: AI governance. This concerns all sorts of things that ensure AI systems work the way they are supposed to, by:
AI governance can be approached differently by organisations. The bar is set higher for public services because their impact is considerable, they have an exemplary role to play, they are expected to embody trustworthiness and inclusion and, on top of that, they are subject to strict (European) rules.
AI governance starts with a clear and widely supported vision. The Knowledge Centre Data & Society has collaborated to develop such a thing, together with UNESCO’s Ethics of AI team and the AI Expertise Centre of Digitaal Vlaanderen (Digital Flanders). What is the Flemish Government’s vision for the responsible use of AI within its own administration?
Drawing on conversations with experts and input from three interactive workshops, we examined various facets of AI governance within the Flemish Government. Six points make up this vision.
The Flemish Government has drawn up clear principles for the use of AI: democratic, trustworthy and human-centred AI, deployed by self-aware and competent staff, with careful data use and management, and with consideration for people and the planet. Whether AI is trustworthy depends, in turn, on eight conditions:
privacy and data protection.
The Flemish Government has drawn up clear principles for the use of AI: democratic, trustworthy and human-centred AI, deployed by self-aware and competent staff, with careful data use and management, and with consideration for people and the planet.
Good principles and conditions, but how do you translate them into a concrete governance structure? Which rules should staff follow? How do we keep track of whether these principles are being put into practice? And how do we put things right if they are not? Governance does not stop at proclaiming principles; it needs to be given concrete institutional form. That includes guidelines, procedures, responsibilities, awareness-raising, and monitoring mechanisms.
Like Flanders itself, the Flemish Government is highly diverse. It comprises entities of widely varying sizes, with more or less experience with AI and differing appetites for risk. Some entities are taking their first tentative steps with AI experiments, while others are already scaling up large projects. Moving at different speeds is not a problem in itself, as long as every citizen enjoys the same minimum level of protection.
Moving at different speeds is not a problem in itself, as long as every citizen enjoys the same minimum level of protection.
Guaranteeing that minimum level of protection requires a shared foundation, developed centrally and applied locally. The AI Expertise Centre has already issued guidelines, which are being supplemented on an ongoing basis. Individual entities should be given the freedom to put these guidelines into practice in their own organisations. In turn, the knowledge these entities gain in practice is indispensable for developing new guidelines or improving existing ones.
The governance of an innovation project faces a standard dilemma:
Moreover, the balance does not look the same for every project (e.g. a transcription tool versus a fraud detection model) or for every phase (testing versus implementation). Finding the right weight and the right form of AI governance for each phase of an AI project takes a great deal of thought. At the outset, put together a project team to map out the governance needs throughout the project: keep it as light as possible and as heavy as necessary. Revisit this plan regularly.
AI systems are often unpredictable because of their autonomy or because they adapt. This means that stage-gate governance is not always appropriate (checks at a fixed point, after which the project moves on to the next phase). Think of phenomena such as model or data drift. Regular, sustained and recurring checks work better. A wide range of tools is available for this. For internal AI projects (back-end), internal checks or audits may suffice; for projects that interact with citizens, consider calling on external partners (for example for real-world testing, external audits or red-teaming).
Every AI system should have an owner, supported by a multidisciplinary team with technical, legal and policy expertise.
AI projects are often so complex that (ultimate) responsibility can hardly rest with a single individual. At the same time, shared responsibility leads to ambiguity and a weaker sense of ownership. Every AI system should have an owner, supported by a multidisciplinary team with technical, legal and policy expertise. That expertise is not yet present everywhere and needs to grow. New roles such as an AI Compliance Officer will be far from a luxury in the future.
AI systems that can have a significant impact on citizens must be subject to meaningful human oversight. This can take many forms, such as proofreading texts, checking input or carrying out periodic checks. Take measures to ensure that this human oversight remains meaningful: avoid rubber-stamping, provide time, resources and training for those carrying out oversight, but also optimise your system so that meaningful human oversight remains possible. And – last but not least – make sure AI projects do not take the humanity out of decision-making.
Make sure AI projects do not take the humanity out of decision-making.
Allow staff to deviate from AI suggestions (without any additional burden) and preserve that important discretionary space, so that the impact of a decision remains proportionate. Keep investing in staff and their skills, too, so that they continue to play a meaningful role in the future.
AI governance with a backbone: steered centrally and implemented locally, proportionate in weight, sustained and recurring, with every responsibility in the right place and people – and humanity – at the heart of it all. That is the vision. That is the first step. This vision now needs to be developed further, applied and kept up to date. The AI Expertise Centre of Digitaal Vlaanderen has a key role to play here, together with all stakeholders across the Flemish Government.
Image: Clarote & AI4Media / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/
