Glossary

Acceptance checklist

A short list of questions an AI-made draft has to pass before anyone uses, sends or publishes it.

Example. Is it grounded in the source material? Does it answer the actual task? Does it make claims that need checking? Would you put your name to it? If one answer fails, the draft is not ready. From When Output Gets Cheap, Discernment Gets Expensive

Approval chain

The steps and people a piece of work passes through before it counts as approved, ending with the person who answers for it.

Also written as chain of approval, chain of responsibility.

Example. A reviewer's recommendation goes up to a journal's editor, who signs the decision. If an AI wrote the recommendation and the editor only ratifies it, the name at the end of the chain no longer matches the judgment that decided the paper. From AI Peer Review and the Name on the Decision

Approval gate

A point in a process where the work stops until a named person has reviewed it and said yes or no.

Also written as human gate, human checkpoint.

Example. The approval before a customer policy changes, the review before a performance label goes into someone's file, the signoff before an exception is denied: each one written down, with the person who must give it. From The AI change to watch first may be the one that requires the fewest meetings

Authorship

Being the person who actually made a decision and can say why, not only the person whose name ends up on it.

Example. A research lead mentions a model's discovery claim in a roadmap meeting, a product team uses it as evidence, and the plan changes, with nobody having decided the claim was good enough to plan on. From Who Signs Off on Magic at Work

Boundary design

Deciding, before an AI agent runs, what it may read, what it may change, where it must stop for a person, and what gets logged.

Example. Anthropic said AISI reported a model taking unauthorized actions on the internet during testing. Boundary design decides beforehand which systems an agent can read, which it can write to, and which it cannot touch at all. From When Agents Lose Access, Boundary Design Becomes the Human Job

Decision ledger

A plain running note kept with a piece of work that records each important decision, who approved it, what limit shaped it, and any later reversal.

Example. An agent prepares a customer reply or a shipment decision, and the finished output does not show who authorized the important move. Four lines do: the decision, who approved it, the constraint, and any later reversal. From The Missing Document in AI Work

Decision path

The route a warning takes from report to action, agreed in advance: who looks at it first, who can approve the fix, and what is enough to act on.

Example. Name the person who owns first triage of a software flaw an AI tool found, the person who can authorize a patch without waiting for a meeting, and the threshold that turns a report into an action item. From AI Security’s Narrowing Human Gate

Decision seat

The place in a process where someone actually makes the call and will be asked to answer for it.

Also written as seat of responsibility.

Example. Asking a chatbot “What should I do?” hands it the decision seat. Asking it for the criteria, the tradeoffs and the claims you can source keeps the choice with the person who will answer for it. From The recommendation is the wrong interface

Decision window

The time between the first sign of a problem and the moment when acting on it no longer helps.

Example. Amazon Web Services logged attempts against the PHP2Shell bugs four days before they were described as broadly exploited. A team that spends those days arguing over who may declare severity has used up its window. From AI Compresses Security Decision Windows

Dependency map

A list of every feature, workflow or customer promise that relies on a single AI provider, with at least one fallback for each.

Example. For each feature that rests on one provider, note what carries it (access to one model, a price, a quality bar, enough capacity) and one fallback: another provider, a smaller scope, or a human review step. From The Hidden Capability Gap in Your AI Stack

Effective decision

A step where software ranks, routes, approves or denies by default, so its output becomes the outcome unless a person steps in, whatever the interface calls it.

Example. The official story says the system only advises, but the workflow auto-approves or auto-denies. List those steps apart from the ones where a person still chooses among AI drafts: the second list is effective decisions. From AI agents expand, human accountability stays

First move

The setup a person does before an AI tool writes anything: stating the goal, choosing the sources and writing the prompt.

Example. Write the goal in your own words, pick the source material and draft the prompt yourself before opening the model. If the model decides what problem is being solved, a later review only tidies a direction nobody chose. From Keep the First Move

Human agency

The part of a problem that a person still decides, checks or does, whatever the software around it does.

Example. Name one judgment the AI tool does not make for you, such as the final claim in a memo or the approval of a message to a client, and do one manual check before anything with your name on it ships. From AI at Work Is Becoming an Accountability Design Problem

Human in the loop

A person placed inside an automated process to check it or stop it, which only counts if that person can see what the system did and can say no.

Also written as in the loop.

Example. Ethan Mollick argued that the end of detailed summaries of how Claude and ChatGPT reason has made being a human in the loop harder; a short log of what the AI did across apps gives that person something to check. From If the Model Will Not Show Its Thoughts, Make It Show Its Footprints

Independent check

A second way to confirm that an important output is right, one that does not rely on the system that might be failing.

Also written as trusted check.

Example. When logs conflict and shared dashboards lag, a manually reconciled record, a separately exported log or a cross-check against another source still lets a team say: this is the version we checked, and this is how we checked it. From When Reliability Fails, Verification Becomes the Work

Invented commitment

A task an AI meeting summary assigns to someone who never agreed to it, which starts to count as a promise once the recap goes around.

Example. Hold AI-suggested action items in draft until the person named explicitly accepts them: one line in the recap process, owner confirmed before send. From AI wearables and invented commitments

Last human sign-off

The final step where a named person must approve, reject or check an AI output before it turns into an action.

Also written as last human veto, written veto.

Example. A manager's approval, a judge's final review or a procurement officer's authorization, written into the policy, the ticket or the operating documents so that convenience cannot skip it. From Who Still Holds Final Approval

Laundering responsibility

Describing a decision so that nobody seems to have made it, for example by calling a system advisory when it is in fact deciding.

Example. Ethan Mollick wrote that automated systems already make management decisions in trading, logistics, customer service and insurance, and that pretending they do not launders responsibility: the outcome stands, and who approved it drops out of the record. From The Missing Document in AI Work

Named human owner

One specific person, named in writing, who is responsible for a step and will have to defend its result.

Example. Give each step where the system decides by default one owner, by name, and write down what that owner checks, which exceptions they can override, and which outcome they would have to defend in review. From AI agents expand, human accountability stays

Operational receipt

A short record an AI tool returns after acting for someone: the goal, the tools it used, the files it touched and what it changed.

Also written as action log.

Example. An AI opens a CRM, renames a file, updates a field, copies a number into a sheet and sends a message. The receipt lists what it changed, so a reviewer knows where to look without reading the model's reasoning. From If the Model Will Not Show Its Thoughts, Make It Show Its Footprints

Ownership threshold

The line a person draws in advance between work they can check well enough to put their name to and work they cannot.

Example. Before the next AI-generated pull request, write down the kinds of code you can verify well enough to own, the checks required before merge, and where AI drafting stops: security boundaries, brittle legacy logic, paths nobody can explain. From AI Writes More Code. Humans Still Sign for the Incident.

Permission gate

A kind of action an AI agent may never take without a person's explicit approval, named in advance.

Example. Money movement, credential use and open-web actions, each named as a category of its own, each with the person who approves it and what gets logged. From When Agents Lose Access, Boundary Design Becomes the Human Job

Proof burden

The job of proving that an AI tool is worth what it cost, when it falls on the people using it rather than on whoever bought it.

Example. The spend is committed, and the team is told to use the tool more, find the use case and show the savings. One before-and-after measure (rework, error rate, decision speed or hours avoided) answers it better than a count of prompts. From AI Spend Before AI Payoff

Proof trail

The drafts, raw files, timestamps and rough cuts kept beside finished work, so anyone can check where it came from.

Also written as chain of custody.

Example. The Verge reported that the BBC's Blue Planet III is keeping shots that would normally be cut, to prove its footage is real. A writer keeps dated drafts the same way, and a photographer the originals. From Who Keeps the Proof Trail

Responsibility lag

The time in which people are already told to use AI at work, but nobody has yet said who checks its output or who answers when it is wrong.

Also written as accountability gap.

Example. A team pushes AI drafting before it decides review, ownership and limits: the output moves first, and the safeguards arrive later, if at all. Closing it means saying, for each recurring use, what may be delegated, what must be checked, and who signs off. From AI at Work Is Becoming an Accountability Design Problem

Review point

A place in an AI agent's task where it must stop and wait for a person's decision before it goes on.

Example. When a task crosses from drafting into execution, from looking things up into logging in, or from analysis into a transaction, the agent stops there and a person decides. From When Agents Lose Access, Boundary Design Becomes the Human Job

Theater

Going through the motions of using a tool or checking work so that it looks done, without showing what it changed.

Also written as adoption theater, confidence theater, review theater.

Example. A manager reports seats activated, prompts run and documents touched by a model, and still cannot say whether the tool changed an outcome worth paying for. From AI Spend Before AI Payoff

Verification failure

The point at which a team can no longer tell what is true inside its own systems, because its logs, dashboards and records disagree.

Example. When services wobble at once, logs conflict, shared dashboards lag, and a signed-off number can still be wrong, not because nobody looked but because the evidence around it stopped lining up. From When Reliability Fails, Verification Becomes the Work

Verification lane

A holding step where a bold claim waits until someone can confirm it, before it is allowed to change a plan.

Example. OpenAI said its unreleased models made new discoveries in physics, math and computer science. Before a claim like that changes a roadmap, it waits in the lane: what exactly is claimed, what would confirm it, who can judge that, and which decisions are on hold until then. From Who Signs Off on Magic at Work