Human control of AI systems: from supervision to teaming

Checked 2 Nov 2026. Open license (CC-BY, CC-BY-SA, CC0, public domain): full text indexed and used in synthesis.

This article reviews two main approaches to human control of AI systems: supervisory human control and human-machine teaming. It explores how each approach defines and guides the operational interplay between human behaviour and system behaviour to ensure that AI systems are effective throughout their deployment. Specifically, the article looks at how the two approaches differ in their conceptual and practical adequacy regarding the control of AI systems based on foundation models--i.e., models trained on vast datasets, exhibiting general capabilities, and producing non-deterministic behaviour. The article focuses on examples from the defence and security domain to highlight practical challenges in terms of human control of automation in general, and AI in particular, and concludes by arguing that approaches to human control are better served by an understanding of control as the product of collaborative agency in a multi-agent system rather than of exclusive human supervision.

• Tsamados et al. 2024 argue that human control of AI must shift from supervision to teaming, with evaluation of human-AI systems as a whole rather than AI benchmarks alone; Zhu et al. 2026 independently identify sustained shared control, cross-session memory, and recovery from complex failures as agent challenges that existing human-AI interaction principles do not capture. Trace →
• Even specialized retrieval-augmented professional tools remain unreliable enough to make human verification indispensable: Magesh et al. 2025 found leading legal research tools "each hallucinate between 17% and 33% of the time," informing "the responsibilities of legal professionals in supervising and verifying AI outputs." Tsamados et al. 2024 add that for some deployment contexts "our risk tolerance for a particular task or environment is too low" for AI integration to be a net benefit at all. Trace →
• Tsamados et al. 2024 argue that as AI systems become less predictable, evaluation must shift "from benchmarking progress on whether AI systems can outperform humans in given tasks to assessing the performance of human–AI systems as a whole," moving the control paradigm from supervision toward teaming. Trace →
• Tsamados et al. 2024 and Calboreanu 2026 mark a shift in the unit of analysis for human control of AI: away from a human supervising a single AI system and toward evaluating and managing humans teamed with, or individually orchestrating, portfolios of AI tools. Trace →
• The 'human as final check' is squeezed from both sides. Zhu et al. 2026 report that 'roughly half of benchmark-passing patches written by agents would not be accepted by human project maintainers,' while Tsamados et al. 2024 warn that operators 'over-trust and accept uncritically the outputs of AI systems' — output verification is simultaneously indispensable and psychologically unreliable. Trace →
• Tsamados et al. 2024 themselves state an outer bound on the 'human as final check' framing: a deployment context's risk tolerance may be 'incompatible with the unpredictable behaviour that is inherent in AI systems,' so that 'even with human control, the integration of AI would not become a net benefit' — human supervision is necessary but cannot always make AI deployment worthwhile. Trace →
• Tsamados et al. 2024 state that the teaming paradigm 'rests on the assumption that the pace of innovation in AI research remains centred on foundation models,' and that with more explainable, predictable architectures 'supervisory human control may re-emerge as the primary approach to controlling AI systems' — the shift from supervision to teaming is a bet on a technical trajectory, not a settled endpoint. Trace →
• The uncritical-acceptance failure that supervision must counter recurs across roles and domains: Tsamados et al. 2024's operators 'tend to over-trust and accept uncritically the outputs of AI systems,' Singh 2026's students are 'seduced by the fluency' of LLM outputs into weakened verification, and Humphreys et al. 2024 name 'overreliance and over-trust in generative AI' as ethical hazards of business adoption — the supervisee's trust psychology, not only system error, is the common supervisory problem. Trace →

• Supervision and doctoral education