Although stakeholder participation is expected to promote equitable and sustainable natural resource management, lessons from the past tell us that more careful attention needs to be paid to achieving equitable impacts. Now the question is how to address social inequities and power asymmetries. Some authors emphasize the need for more dialogue, while others prefer a critical perspective, arguing that dialogue might not be sufficient to avert the risk of a process deepening existing social inequities. This article aims to enrich this debate and question the practical implications of the critical perspective through an in-depth analysis of power games in a participatory process. A Companion Modelling (ComMod) process was conducted in an Akha community of Northern Thailand with a critical perspective, i.e. with a concern for the less influent stakeholders. Simulation tools such as role-playing games were used to mediate a cross-cultural learning process among researchers, farmers and administrators about a local irrigation water management problem. The detailed analysis of power games in this learning and negotiation process reveals that in spite of initial power asymmetries, the poorest farmers of the community started to voice and assert their interests. This was very much due to the role of a Western researcher who put the equity issue on the public agenda and to the strategic support of a charismatic Christian leader. We identify a set of practical facilitation methods that helped to manage power asymmetries and to level the playing field, but we also discuss the main limits of our cultural-embedded methodological choices. Acknowledging that "the facilitators' neutrality" is an illusion, this study allows us to raise the question of their social legitimacy. We suggest that they should systematically make explicit and reflect on their cultural-ideological background and methodological hypothesis and choices and their effects on the socio-political context. This article is an original attempt to accept this challenge. (Contains 2 tables and 3 figures.)