Why the future of AI requires clear human governance

1. How do you use artificial intelligence?

I use AI every day – to improve my work processes, to sharpen my strategic thinking and to do more effective work than I would be able to without it. But my deeper engagement with AI is as a human–AI co-creation scholar.

For the past five years, my work has centred on taking state-of-the-art technology and structuring novel, research-relevant interactions around it – always with an emphasis on real-world relevance and impact. A good illustration is our Crea Vision project, where we have explored generative AI image generation as a way to support public participation in debates around societally relevant dilemmas. Before the ChatGPT era, we had to train our own image generation models from scratch to make this possible. Now, with capable image models and the ability to vibe code interfaces, the speed and flexibility with which we can set up custom interventions has accelerated manyfold. What used to take months of engineering can now be prototyped in days, which means we can iterate research designs in close contact with real participants and real institutional partners.

My colleagues and I firmly believe that deep human–AI co-creation – Hybrid Intelligence (HI) – is the pathway to a positive future for humanity. We have therefore formed the HI Manifesto consortium, which has brought together more than 75 researchers and practitioners across management, computer science, economics, psychology, design education and public policy, to chart out both the challenges and the concrete pathways for making this the dominant mode of AI use rather than autonomous, end-to-end AI operation. The Manifesto argues that the future of AI is not technologically predetermined; rather, there is a real fork between two paths: automation-dominant trajectories that displace human judgment and Hybrid Intelligence trajectories that keep human tacit knowledge structurally and economically indispensable. Realising the latter requires coordinated work across four levels – interfaces, individuals, organisations and society – so that empowering technology is matched by organisational and societal conditions that actually reward context-rich human contribution.

To achieve this, we need to provide structured guidance on how we can all interact with AI systems in ways that preserve our control and authorship, even when we are no longer the producers of every line of text or code in the final product. This is what the FERC framework is for. FERC stands for Frame, Explore, Refine, Commit, and it describes a governed cycle for multi-round human–AI collaboration. In the Frame phase, the human articulates intent – goals, constraints, success criteria and non-goals – before any generation happens. In Explore, the AI is deliberately asked to produce multiple alternatives, which expands the search space rather than converging too early. In Refine, the human evaluates those alternatives comparatively, articulates trade-offs and reshapes the criteria for the next round; this is where authorship is most often lost in everyday use, and where FERC most directly protects it. In Commit, the human makes a deliberate, accountable decision and takes ownership of the consequences. FERC reframes effective AI use away from one-shot prompt optimisation and toward process governance – and it is grounded in decades of creativity and metacognition research on problem construction, divergent and convergent thinking, and self-regulated judgment.

We are now actively training teachers, students and a growing number of corporations in Denmark and beyond in this approach – not as a rigid procedure, but as a shared language and a practical discipline for building a culture of human agency and critical thinking in everyday human–AI interaction. The aim is straightforward: as AI grows more capable, the value of deliberate human framing, comparison and commitment grows with it, and we need an entire generation of professionals who can do this well.

2. What difference has the use of AI made for the students?

There is no single answer – the literature is clear on this, and our own experience matches it. The effect depends profoundly on the individual student, their goals and how they use the tool.

One concrete student-facing activity I've developed, among a plethora of genAI-powered research initiatives, is a custom interface with three different GPT conditions: one is essentially standard ChatGPT; one is built around uncertainty awareness, questioning everything and pushing the user to articulate how certain they actually are; and one is built around metacognitive scaffolding, drawing on the self-regulated learning literature to reflect questions back to the user.

In our master's courses in Innovation Management, Advanced Innovation Management, and Product Design and Innovation Engineering, students were asked to analyse five cases over the course of the programme using one of these three conditions or a no-AI condition. After trying each version, they could choose which version to use, and the overwhelming majority went straight to standard ChatGPT – not because they thought it was best for learning, but because of time pressure. They used it like a search engine. In interviews and reflections afterwards, many said they would have liked to engage more deeply with the metacognitive scaffolding version – they recognised it would help them co-learn with the tool – but they didn't feel they had the capacity to learn the tool and the material at the same time.

That is itself one of the most important findings. AI doesn't introduce fundamentally new problems. We had cheating before, we had motivation problems before, we had students who learned deeply and students who optimised for completion. What AI does is exacerbate the dynamics that were already there. For curious, engaged students, it can be a powerful sparring partner; for students under time pressure, it can absorb the very moments of friction that learning depends on. Many of our students are genuinely curious about these tools but also nervous – worried about not being able to verify their own learning, worried about plagiarism. We have a responsibility as educators not just to indulge the curiosity, but to give them structured frameworks – FERC is one – so they interact with these tools in ways that build judgment rather than substitute for it.

3. And for you as a lecturer and researcher?

It is genuinely an exciting time to be a human–AI co-creation researcher. This is one of the fundamental challenges for humanity in an ever-accelerating world, and being able to contribute to it from a university position is a privilege.

What I find most energising is the ability to combine three things that don't always meet: real-world, impact-oriented interventions; structured creativity research; and psychologically stringent, controlled experimentation. We can simultaneously push what end users perceive themselves to be capable of doing and contribute to a growing knowledge base of scientifically tested and validated methodologies that future real-world technology use can be grounded in. The aim is for this work to be relevant, effective and scientifically founded – not one or two of the three, but all three at once.

As a lecturer, the shift is also significant. I can no longer assume that what students produce reflects only their own reasoning – and I think that pressure is actually useful. It forces us to redesign assessment so that we evaluate the process and the judgment behind a piece of work, not just the final artifact. It also forces us to teach more explicitly: framing a problem, comparing alternatives, articulating why one path is stronger than another, and owning a decision. These were always the goals of education, but AI makes their absence newly visible.

4. What do you expect from a future with artificial intelligence?

I think we are at a genuine fork in the road, and not the one most public discussions describe. In the Hybrid Intelligence Manifesto, we have laid out four trajectories: an automation-dominant future in which agentic AI absorbs end-to-end workflows; an agentic organisation future in which firms restructure around AI-first operations with humans as orchestrators of increasingly opaque systems; a regulation-first future in which oversight is layered on top of an essentially automated production architecture; and a Hybrid Intelligence future in which we deliberately design workflows, organisations and markets so that human tacit knowledge remains economically indispensable.

The first three share a hidden assumption – that fully agentic AI is the inevitable destination, and the only question is how to manage its consequences. I don't think it is inevitable, and I don't think it's desirable. Complementarity between humans and AI is not a fixed set of ‘human-only’ skills – those keep eroding – but a dynamic frontier between prediction and judgment that depends on which contextual knowledge has been formalised and made machine-operational. That frontier is designable. Whether we end up in a world that concentrates value around compute and infrastructure, or one that rewards context-rich, human-touch contribution, depends on choices made now in interfaces, organisations, education and markets.

I also expect, candidly, that we will have to take seriously two costs we currently externalise: the environmental footprint of large-scale generative systems and the cognitive cost of unstructured reliance. Both push in the same direction: toward smaller, bounded, task-specific assistants embedded in well-designed workflows rather than ever-larger general-purpose models invoked reflexively.

5. What principles should guide the responsible and development-oriented use of AI at SDU?

A few principles, which are all connected.

First, process governance over prompt optimisation. Most current AI literacy work teaches students to write better prompts. That's necessary but insufficient. What we need is a shared, teachable structure for multi-round collaboration – FERC is our proposal – so that framing, exploration, evaluative reframing and accountable commitment are made explicit and inspectable across every iteration. This is what protects authorship as AI outputs become more fluent.

Second, transparency in reporting. I would like to see SDU lead on something like human–AI co-creation report cards for scientific and student work: a structured, lightweight disclosure of how AI was used at each stage of a project – what was framed by the human, what was explored with the AI, how alternatives were evaluated and what was committed to. This is more honest than a single checkbox at the end of a thesis, and it teaches the discipline of attribution.

Third, AI as an instrument for human development, not a substitute for it. The goal of education is not faster outputs – it is the growth of judgment. Every AI deployment at SDU should be evaluated by whether it expands or erodes students' capacity to frame problems, weigh evidence and own decisions. The HI maturity model gives a concrete language for assessing this in a given course or workflow.

Fourth, multi-level coherence. The interface layer (which tools we adopt), the individual layer (how we teach students to interact), the organisational layer (how we assess, govern and reward) and the societal layer (how we position SDU graduates in the labour market) need to align. Strong frameworks at one level are undone by misaligned incentives at another.

Fifth, sustainability and sovereignty. Smaller, bounded, open-source-capable models embedded in well-designed FERC-style workflows are usually a better answer than reflexive use of frontier models – both for environmental footprint and for institutional independence from a small handful of hyperscale providers.

And finally, honest pedagogy about the trade-offs. Students should know that using AI as a faster search engine is a legitimate choice in some moments and a developmental cost in others – and they should learn to make that choice deliberately, not by default. Unintentional delegation erodes ownership; deliberate delegation preserves it.