← All articles

The Responsibility Gap in AI Safety

Reflections from the AI Safety Conference in Stockholm on democratised AI capabilities, technical competence, and how responsibility is distributed across the AI ecosystem.

Yesterday, I attended the AI Safety Conference in Stockholm. I found the discussions on AI alignment particularly interesting, but I noticed that questions of responsibility often seemed to return to AI companies and the necessities of governmental regulation.

However, I see a problem. We already have powerful models and agents that allow people with little or no formal technical training to build and deploy AI-powered applications. The ability to create technological systems is becoming increasingly accessible, but does our ability to understand and evaluate their risks develop at the same pace?

There is a difference between being able to build a functioning system and being able to determine whether that system is safe.

Of course, this is not an entirely new problem. People have long been able to develop and distribute software without fully understanding its security implications. But generative AI may be changing the scale and nature of this challenge. Lower barriers to development, faster deployment, increasingly autonomous systems, and the unpredictability of AI behaviour introduce new dimensions to existing safety concerns.

Legislation such as the EU AI Act establishes responsibilities for different actors, but regulation alone cannot ensure that everyone developing or using AI understands the consequences of what they create.

At the same time, expecting every individual to become an AI safety expert is neither realistic nor necessarily sufficient.

I think this is an interesting gap of responsibility. We are democratising access to powerful technological capabilities without necessarily democratising the knowledge and competence needed to evaluate their risks.

How much responsibility can we reasonably place on individuals when they may not fully understand the systems they are building? And how do we ensure safety without restricting technological development to a small number of companies and experts?

Perhaps the challenge is not simply about assigning more responsibility to individuals, but about reconsidering how responsibility, competence, and control are distributed across the AI ecosystem.

I think AI safety needs to be discussed not only in terms of model alignment and corporate accountability, but also in terms of human judgement, technical competence, and the broader sociotechnical systems in which AI is developed and used.

If we democratise the ability to build powerful AI systems, how can we ensure that the capacity to understand, evaluate, and govern their risks develops alongside it?