Risk should follow context and impact.
NIST treats AI as a socio-technical risk problem involving validity, safety, security, accountability, transparency, explainability, privacy, fairness, and lifecycle management.
What survives strongly
Human agency and oversight are legitimate design goals.
The OECD AI Principles include human-centred values, rights, transparency, robustness, safety, privacy, accountability, and mechanisms for agency and oversight.
Exposure is not job loss.
ILO's 2025 global index finds broad potential GenAI exposure but emphasizes job transformation rather than wholesale replacement as the likelier near-term pattern.
Workers can contribute useful implementation knowledge.
ILO case studies document social dialogue shaping AI and algorithmic-management deployment across multiple settings.
What becomes more careful
Augmentation: not automatically better than automation. Preserve human capability where failure, rights, judgment, relationship, or resilience require it.
Human in the loop: meaningful only with evidence, time, competence, authority, an alternative process, and a real override/appeal path.
Deskilling: a domain-specific hypothesis. Measure what declines, what new capability appears, and whether the lost skill is still operationally necessary.
Open/local/cloud/proprietary: architecture choices with different privacy, safety, cost, capability, resilience, support, and lock-in tradeoffs.
Productivity: technology does not decide who receives the gain. Human time is one political claimant among wages, prices, profit, public benefits, and ownership.
Compare against reality, not perfection
Humans and legacy institutions make errors too. Evaluate an AI-assisted system against the realistic alternative it replaces or augments: outcomes, calibration, subgroup performance, rare failure costs, security, latency, availability, workflow, user comprehension, appeal, monitoring, rollback, and fallback.
An AI system can be more accurate on average and still be unacceptable if its rare failures are catastrophic or impossible to challenge.
AI companionship becomes a research track
Do not assume AI relationships are worthless, and do not assume they substitute for people without cost. Measure loneliness, human-relationship displacement, dependency, crisis behavior, disclosure, intimate-data collection, business incentives, and practical exit.
Recommended correction
Keep risk-specific governance.
Replace augmentation ideology with capability analysis.
Separate exposure, task automation, job transformation, and job loss.
Define effective review operationally.
Treat deskilling as testable, not inevitable.
Demote architecture labels from moral categories.
Require monitoring, rollback, fallback, and incident learning.