Do 'participatory design' and 'co-design' actually give people different amounts of say in an AI system, or is that just naming?

This explores whether participatory design and co-design give stakeholders different amounts of real control over AI systems. The corpus has a frank answer that moves the question somewhere more useful: the label matters less than which decisions stakeholders are actually allowed to touch. One source maps eight participatory traditions shaping AI work, including user-centered design, participatory design, co-design and value-sensitive design. It treats them as having distinct goals and power dynamics, and offers them as a shared vocabulary for asking how much agency stakeholders really gain What participatory traditions shape AI design practice?. The retrieved material names the traditions as different. It doesn't spell out, side by side, how participatory design and co-design divide power. So the corpus can't settle the head-to-head comparison directly.

What it does show is how agency gets lost in practice, whatever the method is called. An analysis of 80 participatory-AI papers found that stakeholders shaped the user interface in every project but influenced core model decisions in only 10% of them Does participatory AI actually give stakeholders real power?. The mechanism is quiet. Project leads usually picked the participants and used methods that ask people what they prefer. Those methods leave the final say with the designers. Participation that collects opinions is not the same as participation that shares decisions, even when the word on the label promises the second.

The more surprising finding is that co-design can make people feel represented without actually representing them. In a 12-person study, participants co-designed AI agents meant to act on their preferences and mostly felt the agents captured them well. Independent checks found mixed alignment, and the agents came out more generic and abstract than the people's own answers Does co-design participation hide misalignment in preference agents?. The researchers point to the co-design process itself as the cause: its transparency, limited testing and ordinary cognitive biases built trust faster than they built accuracy. Being in the room can feel like agency while hiding the gap.

A lateral thread in the corpus reframes agency as a question of who starts and who answers for a decision, not who was consulted. Moran separates AI-initiated interface generation from user-initiated 'vibe coding' by who decides to build. That origin point changes what the system is accountable for: design judgment in one case, faithful execution in the other Does who initiates a UI element matter for accountability?. Holldack, Banh and Strobel go further and argue that agentic systems can hand tasks and responsibilities back to humans, so delegation stops running in one direction only Can AI systems delegate tasks back to humans?. Applied to participation, these suggest a sharper test than 'participatory design or co-design?' Ask who chose the participants, which layer they could change (interface or model), and whether anyone checked that the result reflects them.