Reframing AI Loss of Control: What It Is, How to Have It, How to Lose It

A new paper argues that public discourse on AI loss-of-control risks rests on “surprisingly weak foundations,” proposing a formal definition of control anchored to the setting and getting of goals [1][2]. The preprint, submitted to arXiv on 19 May 2026, draws on concepts from cybernetics, management control, and control theory to build a framework for who or what can be in control and what is required to maintain it [1][2]. The authors identify prerequisites such as a functional control loop, requisite variety, and sufficient goal alignment [2]. Once that framework is established, the paper examines how control can be lost and how AI systems can contribute to that loss [2]. A central finding is that humanity can lose varying degrees of control because of AI behaviour that operates far below the level of superintelligence [2]. The authors state that “the potential for loss of control scenarios (as we define them) already exist, and have existed for a long time” [2]. The work arrives amid intensifying scrutiny of AI-generated media. Deepfakes, which use machine-learning techniques such as generative adversarial networks, have raised concerns about disinformation, election interference, and financial fraud [3]. Academics have warned that such synthetic media can promote hate speech and undermine trust in information ecosystems [3]. Social-media platforms amplify these effects: algorithms that track user engagement tend to favour content that spurs negative emotions, and most online misinformation originates from a small minority of superspreaders whose reach is magnified by the platforms [5]. Psychological research offers a parallel lens. When people encounter information that clashes with existing beliefs, they often experience cognitive dissonance and may resolve the discomfort by rationalizing, selectively perceiving facts, or avoiding contradictory evidence altogether [4]. Leon Festinger, who formalized the theory in the 1950s, described the phenomenon bluntly: “Tell him you disagree and he turns away. Show him facts or figures and he questions your sources. Appeal to logic and he fails to see your point” [4]. The paper’s framework suggests that loss of control is not a distant, speculative risk tied only to advanced AI. By defining control concretely and identifying its failure modes, the authors aim to give researchers and policymakers a shared vocabulary for evaluating systems already in deployment [1][2].