Key takeaways

• Mandatory human-in-the-loop on every AI workflow is not safety. It is liability theatre that caps ROI and trains the organisation to never let go.
• The over-supervision tax is real and documented. HBR's 'workslop' research, the Augmentation Trap studies, and METR Time Horizon all point to the same curve. Past a low threshold, every extra review erodes the productivity it was meant to protect.
• Four anti-patterns dominate today: rubber-stamp review, defensive review, theatre review, and fatigue review. None of them produce the safety they claim. All of them degrade the system over time.
• The grown-up replacement is human-on-the-loop, exception-based escalation, post-hoc audit, and adversarial AI review. The human stops being the gate and becomes the architect of the gate.
• HITL is not always wrong. For genuinely irreversible, regulated, or high-stakes calls it is essential. The mistake is making it the default rather than the deliberate exception.