The current state is a human transition
AI’s current state cannot be captured by benchmark scores alone. The International Labour Organization estimates that one in four jobs worldwide is in an occupation with some generative-AI exposure, while concluding that transformation is more likely than wholesale replacement. The lived question is who gets augmentation, who receives surveillance, whose work is deskilled, and who is expected to absorb errors made in the name of efficiency.
The effects are not evenly distributed. Stanford’s 2026 AI Index reports persistent performance gaps across languages and dialects. It also records 362 documented AI incidents in 2025, up from 233 the prior year. A system can look successful in aggregate while placing a recurring burden on people who have less power, less representation in the data, or fewer ways to contest an outcome.
Ask where power moves
Every AI deployment changes who knows, who decides, who must explain, and who can object. In the workplace, algorithmic management may improve consistency, but OECD research also identifies concerns about unclear accountability, opaque logic, and inadequate protection of worker health.
A human-in-the-loop label is not enough. The human must have time, information, authority, and psychological safety to disagree. If the workflow punishes overrides or makes them practically impossible, human review is theater.
The limits people should be able to rely on
The people affected by AI deserve boundaries they can experience, not principles they never see.
- No covert emotion inference, behavioral manipulation, or continuous worker surveillance; do not convert intimacy, stress, or attention into an invisible management score.
- No denial of livelihood, opportunity, essential service, or safety-critical care based solely on an AI output.
- Provide plain-language notice when AI materially shapes an interaction or decision, including what it does, what data it uses, and what its known limits are.
- Give people an accessible route to a qualified human who can review evidence, change the outcome, and provide a remedy within a meaningful timeframe.
- Involve affected workers and communities before deployment, and repeat the review when the use, model, data, population, or consequences change.
- Share productivity gains through better work design, training, mobility, and reduced burden—not only through labor reduction.
Design for dignity under ordinary conditions
Ethical review often focuses on dramatic failure. Daily dignity can be eroded more quietly: a person cannot reach support, a worker must obey an inexplicable score, a customer is forced into an AI channel, or a community bears errors that never appear in the executive dashboard.
Measure burden as deliberately as benefit. Track overrides, appeals, resolution time, language performance, accessibility, complaints, worker experience, and who leaves the process. AI should expand human capability without shrinking human standing. If an organization cannot explain how an affected person can understand, refuse, challenge, and recover from the system, it has not finished the design.
The decisive limit is not simply ‘keep a human involved.’ Preserve human agency: informed participation, real authority to disagree, and an effective route to correction and remedy.