Automation Bias

Human-AI Collaboration Needs More Than a Human in the Loop

September 29, 2026 by Basil C. Puglisi 5 Comments

Human oversight of automated systems is older than AI, and current AI practice still falls short of it.

This paper traces human oversight and accountability from the cybernetics literature of 1948 to Article 14 of the EU AI Act, then separates Responsible AI from AI Governance. It sets out Checkpoint-Based Governance, where a named human holds binding authority to accept, modify, or reject an AI output and leaves a record, and it proposes the studies that would test it. It closes on the Growth OS, a future of work in which AI amplifies people instead of replacing them.

What AI Transformation Actually Requires

September 21, 2026 by Basil C. Puglisi 2 Comments

AI products can be purchased. AI transformation has to be built.

This paper reframes fifteen commonly cited requirements for AI transformation as operational management requirements, from data provenance and purpose to decision traceability and value realization. It shows why automation and observation cannot serve as the sole control on consequential work, and how named human checkpoints close that gap.

A Draft Taxonomy of Attacks on AI Agents Leaves Risk Tolerance to Each Organization #AIg

NIST’s November 9 draft taxonomy of attacks on AI agents says guardrails must face adaptive attackers, yet it sets no risk tolerance. Each organization needs a named human to sign the residual risk decision for every relevant attack type.

A UK Bill Clause on AI Agent Loss of Control Assumes Someone Held the Control. Can You Show Who? #AIg

The UK government tabled a clause letting regulations give ministers investigatory and enforcement powers when a regulated system’s compromise traces to loss of control of an AI agent. A loss-of-control rule assumes someone held control, so operators should be able to show who held stop authority before any incident.

Europe Plans to Assess AI Literacy in 15-Year-Olds. Who Holds the Checkpoint at School? #AIg

The European Commission will pilot an AI Literacy Certificate for 15-year-olds in the 2027/2028 school year. The framework behind it teaches pupils to set their own checkpoints; the school still needs a named person who signs which AI tools pupils use.

NIST’s Ten Agent Identity Tenets: What Checkpoint-Based Governance Already Carried

NIST published ten agent identity tenets on October 27, 2026. Nine were already in a public record that began in 2012.

This article lays the ten tenets against Checkpoint-Based Governance, the February 2026 GOPEL specification and code, and the Responsible AI desk that was publishing before the tenets appeared. Five match in full, four in part, and one, the agent’s own credential, is defined in the record and not yet issued. The reader gets the dates, the quotes, and the table to check them.

When an Agent Acts, Who Confirms Before the Kernel Lets It Through? #AIg

Sysdig AI Defense generally available on November 4 puts allow, confirm, or block in front of every agent action, traces each agent to an accountable human, and keeps audit records of who signed off. Conduktor and Google Cloud land the same week on human approval and sandboxed containment.

Bank Examiners Get a New Question: Is the Human Approval More Than a Click? #AIg

The OCC’s new supervisory framework for bank use of super intelligence tells examiners to test whether a human approval is more than a nominal click, flags a named reviewer without practical authority as a risk, and lets the degree of human oversight shape the risk tier.

New Suits Against OpenAI Ask Courts to Put an Outside Reviewer Over ChatGPT Releases #AIg

Two lawsuits filed in late October ask courts to put an outside reviewer between OpenAI and its next release: a court-approved evaluator in Texas and a five-year independent monitor in the Wan case. A third suit asks for automatic escalation. The requests show what injured parties will ask any chatbot company to prove.