What Must Humans Remain Responsible For?

• Stanford's AI Index documents continuing gains in reasoning, tool use and agentic performance alongside a jagged capability frontier — exceptional on sophisticated tasks, unexpectedly weak on others.
• OpenAI's GDPval evaluates AI on economically valuable professional tasks across dozens of occupations, and Anthropic's Economic Index shows AI used across a growing share of occupational tasks.
• The ILO takes a cautious labour-market view: task transformation is currently more plausible than wholesale job replacement for most exposed occupations.
• Research on worker preferences suggests people do not uniformly want every technically automatable activity automated — work also provides competence, autonomy, connection, identity and meaning.
• NIST, the OECD and the EU AI Act place increasing emphasis on explicitly defined human oversight, human agency and responsibility in consequential AI systems.

Concepts, frameworks, models and decision principles developed by Ravinder — practitioner synthesis, distinct from the external evidence cited above.

• The Human Responsibility Framework — a two-question matrix (How capable is AI of this task? How much human responsibility should remain?) producing four operating modes: Human Performs, Human Owns, AI Performs, Human Governs.
• The Oversight Paradox — human responsibility can remain while human capability quietly disappears; the machine performs the reasoning, the human approves it and remains accountable, and may no longer possess enough independent expertise to challenge it.
• Minimum Viable Human Capability — once responsibility is assigned, define the minimum a person must retain to discharge it safely: know, interpret, verify, decide, or perform. Not every employee needs all five.
• Five levels of human control — Perform, Decide, Approve, Monitor, Audit — chosen by consequence, reversibility, reliability, legal requirements and the capability that must be retained.
• Six responsibility tests — consequence, judgment, values, relationship, accountability and capability preservation — for assessing how much human control a task warrants.
• The Human Responsibility Canvas — a one-page operating tool covering task, AI capability, trajectory, consequence, human responsibility, control level, minimum capability, oversight readiness, atrophy risk, escalation trigger, intervention and performance evidence.
• The operating principle: Automate capability freely. Delegate responsibility deliberately.

• What authority are we prepared to delegate — not merely what activity can we automate?
• If AI removes the developmental work, where will tomorrow's experts come from?
• What capabilities must humans continue practising even if AI can normally perform them?
• If humans remain responsible, have we preserved enough human capability for that responsibility to mean anything?

Composite scenarios for illustration — they describe hypothetical situations, not client results.

• Financial analysis: AI increasingly performs data gathering, calculations, variance analysis, scenario generation and first-draft commentary, while humans remain responsible for validating assumptions, understanding context, recognizing unusual conditions, challenging implausible conclusions and owning the action.
• People managers: AI can summarize performance information, draft reviews and detect patterns, while humans deliberately govern context, mitigating circumstances, consequential conversations, employment decisions, trust and the relationship itself.
• Applied to L&D itself: AI can increasingly own research, first drafts, outlines, assessments and routine production, while the profession retains problem diagnosis, intervention selection, evidence interpretation, capability strategy and human-AI work design.

• L&D begins with the architecture of work, not with learning: business change, work decomposition, AI capability assessment and the human responsibility decision come before any development.
• The function's value moves toward four strategic roles — Capability Architect, Human-AI Work Designer, Capability Risk Manager and Performance Engineer. Training remains one intervention; it is no longer the organizing principle.
• Some capability is an organizational resilience asset and must be deliberately maintained even if rarely used: domain expertise for abnormal cases, diagnostic reasoning, ethical judgment, fallback capability, and the ability to challenge, override and escalate AI.
• Capability requirements can become more selective — for every traditional requirement, ask whether the employee needs to know, understand, interpret, verify, decide, perform, or simply know when to escalate.

• Break strategically important roles into meaningful tasks and assess where AI can already contribute.
• Track where AI capability is advancing fastest.
• Determine which responsibilities should remain human.
• Assign the appropriate level of human control to each consequential task.
• Run the oversight readiness check for every Human Governs task.
• Define the minimum human capability required for the responsibilities people retain.
• Identify capability-atrophy and apprenticeship risk before automating.
• Protect expertise where necessary.
• Build the minimum effective intervention.
• Measure whether people can exercise responsibility — not merely whether they completed learning.
• Treat capability preservation as a resilience strategy.
• Reassess as technology, regulation and work change.

Published research and governance guidance are identified as evidence throughout. The Human Responsibility Framework, the Oversight Paradox, Minimum Viable Human Capability, the Human Responsibility Canvas and the principle “Automate capability freely. Delegate responsibility deliberately.” are synthesis developed for this report — they have not been validated as universal models, and the research base does not prescribe them directly.

• The pace of AI capability improvement is uncertain; the framework is designed to survive capability changes by separating what AI can do from what organizations decide humans should remain responsible for.
• The strongest objection is taken seriously: if AI becomes materially better than humans at production, judgment, verification and monitoring, mandatory human intervention could itself reduce performance.
• Human responsibility should remain where there are compelling reasons — legitimacy, accountability, legal requirements, human rights, agency, relationships, resilience or moral responsibility. The framework is a defence of deliberate responsibility, not of human work for its own sake.