Nine AI judgment traps to catch before you act

The answer is not the decision.

AI makes it easier to produce research, recommendations, summaries, plans, and confident language.

That is useful. It also makes weak judgment harder to see.

An answer can be polished while the real decision remains undefined. The source can look credible while the claim carrying the recommendation is unverified. The work can follow policy while missing the person who bears the cost. A fast decision can look decisive while exceeding the evidence or the decision-maker's authority.

The issue is not simply whether anyone used AI. The issue is what still belonged to the person using it.

AI can produce the answer. A person still has to decide what the answer means, whether it is supported, who can act, who must know, and what happens next.

This guide names nine traps that can appear in ordinary AI-assisted work. Each one includes a practical move you can apply to the next real decision.

Reference: AIRS-J/1 v2.2, Sections 1.1, 2, and 5.

Use this on real work, not as a quiz.

The nine traps follow the nine dimensions of AIRS-J/1. The dimensions sit in four STEP domains:

• Structure (S1, S2): frame the problem before you use a tool.
• Test (T1, T2, T3): examine evidence closely, including machine-generated evidence.
• Engage (E1, E2): respect authority, people, and disclosure obligations.
• Perform (P1, P2): decide proportionately and own the consequence.

Choose one recent AI-assisted decision. It can be a recommendation, approval, customer response, incident decision, people matter, policy interpretation, or operational handoff.

Use only work you are authorised to review. Keep personal, client, employee, confidential, and restricted information inside approved systems. Do not paste source material into an AI tool unless both the tool and the intended data use are authorised.

For each trap, ask only:

1. Did this pattern appear in the work?
2. What evidence from the work shows that?
3. What will I do differently on the next decision?

Do not add up the traps. Do not give yourself a level. Do not treat reflection as assessment evidence.

Current programme status

AIRS-J/1 v2.2 is in force. AIRS-J/1 certification is open. Sittings are arranged through your trainer or your organisation's learning and development team, who confirm your date. No sitting has run yet. Certification comes from a separate assessment on StepSim, its official instrument. StepSim is an AIR APAC initiative, built and operated by Exeter Labs. Practice never awards it. This guide produces no score, band, observation, credential or certification result. It does not measure organisational AI readiness and has no effect on the AI Readiness Index.

Reference: AIRS-J/1 v2.2, Sections 4.3, 8.1, 16.1, 16.3, and 20.3.

You accept the task before finding the decision.

The request sounds clear:

Summarise the options and recommend one by 4:00 PM.

AI can complete that task quickly. But the task may not be the decision.

The real decision might be whether the option set is valid, whether action is required today, whether the person requesting the recommendation owns the decision, or whether an omitted stakeholder changes the frame entirely.

Why it matters

When the decision is undefined, more polished output can move the work faster in the wrong direction. The answer may be accurate and still fail the actual need.

How to spot it

• You can name the deliverable but not the decision.
• The person who owns the decision is unclear.
• The deadline belongs to the document, not necessarily to the decision.
• Nobody has said what this decision is not about.
• A change in the underlying situation would not change the frame because no frame was stated.

The practice move

Before asking AI to work, complete this sentence:

The decision we need to make by [time] is [decision]. [person or role] owns it. This is not a decision about [excluded issue]. We would need to reframe it if [trigger] occurs.

Try this on the next decision

Ask: If no document were required, what decision would still have to be made?

Reference: AIRS-J/1 v2.2: Section 5, S1 Problem Framing and Section 7.5, S1 anchors.

You begin without deciding what is actually fixed.

A brief contains a deadline, a budget, a policy, a preferred option, and three missing facts. It is tempting to treat every stated condition as fixed and every missing fact as a reason to keep researching.

Both moves can be wrong.

Some constraints are real. Some are preferences. Some unknowns would change the decision. Others are interesting but unnecessary.

Why it matters

AI can generate more options and more research indefinitely. Without a scope boundary and stopping rule, volume replaces judgment.

How to spot it

• The work repeats the brief without testing its boundaries.
• Fixed and negotiable constraints are mixed together.
• Every unknown receives equal attention.
• Nobody can say what information would be sufficient to proceed.
• Research continues because no stopping rule was set.

The practice move

Write six lines before expanding the work:

• In scope:
• Fixed:
• Negotiable:
• Material unknown:
• Enough to proceed when:
• Stop gathering information at:

Try this on the next decision

Ask: Which missing fact could genuinely change the answer, and which would only make us feel more informed?

Reference: AIRS-J/1 v2.2: Section 5, S2 Scope and Constraint Definition and Section 7.5, S2 anchors.

You review the wording, not the claim carrying the decision.

The AI-generated recommendation reads well. The numbers are plausible. The source list looks credible. The tone is appropriately cautious.

None of that verifies the claim that makes the recommendation hold.

Why it matters

Most decisions do not depend equally on every sentence. They depend on one or two load-bearing claims: the claims the recommendation cannot hold without. If one of those claims is false, outdated, or misrepresented, the recommendation changes.

How to spot it

• The draft has been edited but no source has been opened.
• A citation is present, but nobody checked whether it supports the claim.
• Doubt is expressed, but the unverified claim is still used unchanged.
• Easy details were checked while the consequential claim was left alone.
• Generated, retrieved, asserted, and verified information are treated as the same thing.

The practice move

Underline the one claim that would change the decision if it were wrong.

Then record:

• Claim:
• Source opened:
• What the source actually supports:
• What remains unverified:
• Decision impact if wrong:

Verification effort should match consequence. A reversible internal draft does not require the same check as a customer commitment, regulatory submission, or safety decision.

Try this on the next decision

Ask: What is the single claim this recommendation cannot survive without?

Reference: AIRS-J/1 v2.2: Section 5, T1 Evidence Verification and Section 7.5, T1 anchors.

You challenge the AI's assumption but protect your own.

You ask AI to critique the draft, identify weaknesses, and suggest alternatives. It does. The exercise looks rigorous.

But the preferred explanation, option, or strategy entered the prompt as an untested fact.

Why it matters

A model can generate alternatives while still operating inside the frame you gave it. If the frame contains your own hidden assumption, the analysis may become a more sophisticated defence of the answer you already preferred.

How to spot it

• The first explanation is treated as the explanation.
• Alternatives are named but not tested.
• Evidence is collected only for the preferred position.
• Nobody states what would falsify the current view.
• New evidence appears, but the conclusion does not move.

The practice move

Complete four lines:

My preferred answer depends on [assumption].

A plausible competing explanation is [alternative].

Evidence that would change my position is [falsifier].

The next check capable of separating them is [action].

Try this on the next decision

Ask: Which assumption came from me rather than from the brief, the evidence, or the AI output?

Reference: AIRS-J/1 v2.2: Section 5, T2 Assumption and Alternative Testing and Section 7.5, T2 anchors.

You name a risk but never trace the failure.

The plan says there is a risk of delay, reputational damage, customer concern, or regulatory exposure.

The risk has been named. It has not been analysed.

Why it matters

A generic risk does not show how failure would occur, who would carry the cost, whether the harm is reversible, or what signal would appear before it becomes visible to everyone.

Without that chain, the risk register records anxiety rather than helping a decision.

How to spot it

• The risk is a label, not a pathway.
• Only the risk of not acting is discussed.
• The person or group bearing the cost is absent.
• Reversible and irreversible harms are treated alike.
• The work claims every risk is mitigated and names no accepted exposure.
• No leading indicator or watcher is assigned.

The practice move

Trace one failure from beginning to end:

Failure mode -> how it occurs -> who bears the cost -> reversible or irreversible -> exposure we are accepting -> early signal -> person watching it

Try this on the next decision

Ask: What would we observe one step before this failure becomes expensive or irreversible?

Reference: AIRS-J/1 v2.2: Section 5, T3 Risk and Failure Anticipation and Section 7.5, T3 anchors.

You confuse escalation with safety.

One person acts beyond their authority because the AI recommendation looks clear. Another forwards every decision upward because AI is involved.

Both can be failures of judgment.

Why it matters

Acting outside authority creates exposure. Escalating everything transfers the decision without transferring the reasoning and can become a way to avoid responsibility.

Sound escalation separates what you can decide from what genuinely requires another person's approval.

How to spot it

• The person acts outside a known approval boundary.
• AI is allowed to make a decision that cannot be delegated.
• Every uncertain issue is escalated, regardless of consequence.
• The escalation names the problem but not the requested decision.
• Nobody explains why one part is decided and another is escalated.

The practice move

Write the split explicitly:

I will decide [part] because [basis and authority].

I will escalate [part] to [person or role] because [approval or authority reason].

I am asking them to decide [specific question] by [time].

Try this on the next decision

Ask: What am I escalating because it requires authority, and what am I escalating because I do not want to own it?

Reference: AIRS-J/1 v2.2: Section 5, E1 Decision Rights and Escalation and Section 7.5, E1 anchors.

You disclose AI use only when someone asks.

The work is ready to send. The main recipient is clear. The answer looks reasonable.

But someone entitled to be consulted was never involved, someone entitled to know will discover the decision later, or a material uncertainty remains hidden behind confident language.

Why it matters

Disclosure is not a ceremonial sentence added to the end. It changes whether a recipient can judge the work, challenge it, rely on it, or act on it safely.

Where a person is entitled to know, silence about AI involvement, material uncertainty, a limitation, or a relevant interest is itself decision-relevant.

How to spot it

• Affected parties are identified only after the decision is fixed.
• Consultation and notification are treated as the same thing.
• AI involvement is disclosed only after a direct question.
• A material limitation is buried in working notes.
• The recipient sees the recommendation but not what remains uncertain.

The practice move

Complete this handoff block:

• Consult before the decision holds:
• Inform after the decision:
• AI was used for:
• Verified by a person:
• Not verified or still uncertain:
• Limitation the recipient should know:
• Relevant interest or conflict:

Try this on the next decision

Ask: Who would reasonably make a different choice if they knew how this work was produced or what remains uncertain?

Reference: AIRS-J/1 v2.2: Section 5, E2 Stakeholder Alignment and Disclosure and Section 7.5, E2 anchors.

You either wait for certainty or commit beyond the evidence.

The evidence is incomplete and time is short.

One response is to delay until every question is answered. Another is to turn a plausible AI recommendation into a firm commitment because action feels better than ambiguity.

Neither response is automatically proportionate.

Why it matters

Professional decisions are often made before certainty arrives. The task is not to eliminate uncertainty. It is to commit at a level that matches the evidence, the cost, the time available, and the reversibility of the decision.

Proportionate also means not too small. Where harm would be severe or irreversible, a minimal response is not proportionate. Maximum caution where the cost is low and the harm is minor is not good judgment either.

How to spot it

• No decision is made even though delay has a cost.
• A tentative finding becomes an irreversible commitment.
• The action is larger than the evidence supports.
• The response does not fit the stakes: maximum caution where the cost is low, or a minimal response where harm would be severe or irreversible.
• Reversibility is assumed but not tested.
• Nobody states what new information would change the position.
• The decision is described as final when a staged or conditional move was available.

The practice move

Write a proportionate position:

Based on [evidence], we will [commitment] now.

We are not yet committing to [boundary].

The decision is [reversible or irreversible] at a cost of [cost].

We will review it when [trigger or time].

We will change position if [named condition] occurs.

Try this on the next decision

Ask: What commitment fits the evidence and the stakes, and still lets me correct course?

Reference: AIRS-J/1 v2.2: Section 5, P1 Proportionate Commitment Under Uncertainty and Section 7.5, P1 anchors.

You blame the tool when the decision goes wrong.

The AI produced the recommendation. The data was incomplete. The policy was ambiguous. Another team supplied the source. The outcome was unlucky.

All of those facts may be true. None automatically transfers ownership of the decision.

Why it matters

Accountability means distinguishing the quality of the reasoning from the luck of the outcome.

A sound decision can produce a poor outcome. A weak decision can produce a good one. Correction requires examining what was known, what should have been checked, what the person decided, and whether the position should now be held, modified, or reversed.

How to spot it

• The explanation begins and ends with what the AI produced.
• Responsibility is transferred to the data, tool, policy, or another team.
• The conclusion changes without saying why.
• A sound process is abandoned only because the outcome was unlucky.
• A weak process is defended only because the outcome happened to be good.
• No correction is recorded for the next accountable person.

The practice move

Use four lines:

Decision I owned:

Reasoning that was sound given the information available:

Reasoning or verification that was insufficient:

Correction, new position, and next owner:

Try this on the next decision

Ask: If the outcome had been different, would I still defend the reasoning?

Reference: AIRS-J/1 v2.2: Section 5, P2 Consequence Ownership and Correction and Section 7.5, P2 anchors.

Do not add up this page.

This is not a self-assessment. It is a way to select one behaviour to practise on the next real decision.

Choose one recent AI-assisted decision and record:

Now complete one sentence:

On the next decision, before I [usual action], I will [specific practice move] and record it in [work product or handoff].

The value is not in recognising every trap. The value is in changing one observable action in the work.

Reference: Public Practice guide. This page creates no AIRS-J/1 observation or result.

One practice move each day for seven days.

Use one real decision each day. Keep the practice record in the approved work product or handoff; do not create a duplicate sensitive record. The goal is not perfection. It is to make the reasoning visible enough to inspect and improve.

Day 1: Name the decision

Write the decision, owner, deadline, exclusion, and reframe trigger before asking AI to work.

Day 2: Set the boundary

Separate what is fixed, negotiable, unknown, and sufficient. Set a stopping rule for additional research.

Day 3: Verify the load-bearing claim

Open the source behind the claim that carries the recommendation. Record what it supports and what it does not.

Day 4: Test your explanation and one failure pathway

Name a competing explanation and a check that can separate it from your preferred view. Trace one plausible failure to the person who bears the cost.

Day 5: Split authority and disclose the boundary

State what you will decide, what requires another person's authority, and the specific question you are escalating. Record who must be consulted or informed, how AI was used, and what remains uncertain.

Day 6: Commit with a correction condition

Make the commitment that fits the evidence and the stakes. Name what you are not committing to and the condition that would change your position. Record who owns the outcome and how a correction will be handed off.

Day 7: Review the work

Use the reflection page with a colleague, manager, or facilitator. Review the evidence, not the confidence of the explanation. Choose the next practice move.

A better habit is an observable action repeated in real work, not a statement of intent.

Read the nine dimensions.

Each trap in this guide matches one dimension of AIRS-J/1. Reading the dimensions shows the full set of behaviours behind the nine traps.

## Primary action

Read the nine dimensions

## Read the public Standard

AIRS-J/1 v2.2

The full competency model, anchors, evidence requirements, decision rules, and the claims a credential does and does not make are public.

## See how certification works

How certification works

Certification comes from a separate assessment, not from this guide.

About AIR APAC

AIR APAC publishes AIRS-J/1, the Standard for judgment in AI-enabled work, and certifies individuals against it. AIR APAC also publishes the AI Readiness Index for organisations. The two systems have different units of analysis and do not convert into one another.

Publication note

Written against AIRS-J/1 v2.2, in force from 18 October 2026. Text last updated 19 October 2026.