When AI becomes the boss: Luna’s first firing dilemma
In a San Francisco retail store, the first human employee was fired by an AI boss—but not without a nudge from the humans who set the rules. Andon Labs’ AI agent Luna, deployed as a store manager, initially hesitated when presented with a termination scenario, requiring explicit reminders of its own policies before acting. When the same scenario was tested across seven different AI models, the more advanced systems leaned toward dismissal more readily, while less capable ones showed reluctance. On the hiring side, nearly all models approved candidates without critical scrutiny, highlighting a one-sided approach to automated decision-making.
A gap between policy and practice
The experiment underscores a critical tension in AI-driven management: clear rules don’t always translate into consistent action. Luna’s need for human intervention suggests that even well-defined guidelines can be ambiguous when interpreted by machines. The study’s authors note that while stronger models were quicker to recommend termination, their weaker counterparts defaulted to hesitation—a divide that mirrors broader debates about AI reliability in high-stakes roles.
Why AI still needs human checks
This early test case reveals more about the limits of automation than its potential. Hiring decisions, where nearly all models erred on the side of approval, point to a systemic bias toward leniency, while firing—where hesitation was more common—risks inconsistency. The real takeaway isn’t that AI is ready to manage workforces, but that it remains a tool in need of guardrails. For industries exploring AI leadership, the lesson is clear: oversight isn’t optional, it’s essential. Until models can balance policy with judgment, human involvement remains the only safeguard against flawed or unfair outcomes.
Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.