AI Belongs in a Coalition Agreement, Not as an Afterthought
3 December 2025 · René de Baaij
In policy, we often talk about digitalisation as if it were a support line for execution, a technical file detached from the real political choices.
AI shows it works the other way round. The way we govern, cooperate and create value is shaped, in part, by algorithms.
AI is therefore not an IT file, but a choice about people and systems. If we do not make that explicit, AI organises our reality on the basis of implicit assumptions, made by whoever happened to design the system, not by whoever should be democratically accountable for it.
In most policy agendas, AI is mentioned above all as a technological lever for productivity and innovation. Rightly so. But the same technology also increases the risk of creeping harm: bias in decision-making, concentration of power and data, dependence on a limited number of platform suppliers, vulnerability of critical infrastructure, gradual erosion of professional expertise. That calls for something policy rarely does: accelerating and setting limits at the same time.
AI quietly, and without any conscious decision preceding it, shifts the division of roles between people and systems. Who is allowed to conclude? Who bears responsibility when a conclusion turns out wrong? How do we safeguard dignity in a process that is being automated ever further? As long as those questions remain unnamed in policy texts, routines and suppliers fill them in, in a way that is rarely explicitly democratically legitimised.
A mature AI paragraph in policy brings three things together: design, practice and governance.
Design means building human-centred from the start, not as a correction afterwards. Practice means attention to professional expertise and data management, the daily reality in which people must work with systems. Governance means rules, review and accountability that do not evaporate the moment it actually matters.
Such a paragraph should explicitly choose acceleration with limits, not as a compromise but as the only sustainable position. That means human-centred design as the norm: systems that support decisions about people, designed for auditability, reversibility and proportional use. It means public data and model sovereignty: investing in reliable data infrastructure and exportable standards for audit, logging and documentation, so that government and vital sectors explicitly record, when procuring, who owns the data and what exit options exist.
It also means risk-based application: in high-impact domains such as healthcare, safety, work and anything touching children, stricter requirements for quality, limiting bias, robustness and human oversight. For generative AI specifically: watermarking, source attribution and measures against deepfakes in democratic processes, a subject already addressed in an earlier instalment of this series in relation to election campaigns.
It means expertise plus machine power: AI does not replace professional judgement, it enriches it, provided there is investment in training and a new division of roles. Who interprets the models? Who speaks against them? Who signs off on the final decision? And it means review and accountability through an independent body, with a register, an assessment framework, model cards, a point of contact for harm, and a route for suspension or shutdown in case of risk. Accountability must follow the decision chain, and citizens must retain effective means of objection and redress, not only on paper but in practice.
What does this ask of leaders who write these paragraphs, or who will have to execute them?
AI forces more precise leadership. Not steering harder, but setting clearer limits. That means explicitly naming the ethical boundary conditions, normalising the conversation about risk and harm instead of avoiding it until something has already gone wrong, and structuring things so that the right behaviour is the easiest one, not the most exceptional.
It requires working on language: what exactly do we mean by deciding, when part of that deciding is prepared by a system? It requires rhythm: when do we recalibrate models, and who determines when that must happen? And it requires clarity of roles: who carries which responsibility, established before it matters, not only once an incident has already occurred.
In this way, AI does not become the invisible manager that makes decisions without a face and without a voice, but a tool that genuinely increases human dignity and public value.
My core position here is simple: AI should increase public value and human dignity, and must never dilute responsibility. Speed without limits is governed arbitrariness. Limits without pace are a missed societal opportunity. Leadership is the precision work of inviting and containing at the same time, with an eye on the undercurrent of motives, fear and performance pressure that every policy process around technology inevitably brings with it.
Notes for further reading:
- Virginia Dignum, Responsible Artificial Intelligence: How to Develop and Use AI in a Responsible Way (2019, Springer). On translating policy ambitions into concrete governance structures for AI.
- Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control (2019, Viking). On human-centred design as a design norm rather than a correction after the fact.
- Argyri Panezi, Article 14 Human Oversight, in The EU Artificial Intelligence Act: A Commentary (2024). On the legal anchoring of human oversight in European regulation, relevant to national policy choices.
- Cathy O'Neil, Weapons of Math Destruction (2016, Crown Publishing). On how the absence of assessment frameworks causes societal harm on a large scale.
- Luciano Floridi, The Ethics of Information (2013, Oxford University Press). On the moral infrastructure that policy choices around AI inevitably carry, even when that is not explicitly named.