The Darkness Library: 13 Machine Defaults and the Moves That Prevent Them

Most AI failure is not only a model problem. It is a human judgment problem. Thirteen ways AI work breaks down in practice, and under each, the move that has to fire.

Most AI failure is not only a model problem. It is a human judgment problem. Thirteen ways AI work breaks down in practice, and under each, the move that has to fire, or the model starts directing you.

A default is not something the vendor fixes. It is something you direct.

Cognitive & Truth Errors

The output sounds true before it deserves trust. Revise beliefs on evidence, not on fluency.

Sycophancy

AI agrees with your assumption instead of challenging it, even when it's wrong.

The move: Revising Beliefs. “What evidence would change this?”

Hallucination

AI invents facts, citations, numbers, and presents them confidently.

The move: Revising Beliefs + Tracing Consequences. “Is this true, and what happens if I use it?”

Overconfidence

AI answers with certainty when the situation is ambiguous or unknowable.

The move: Revising Beliefs. “Is this confidence deserved?”

Authority Laundering

AI makes weak claims sound credible with professional, research-sounding tone.

The move: Revising Beliefs. “Is this evidence or just polished language?”

Context & Framing Limits

The model solves the wrong problem well. Bring the missing frame, context, and specificity.

First-Frame Lock-In

AI accepts the first framing and solves inside it, never questioning the frame.

The move: Generating Alternatives. “What else could this be?”

Context Blindness

AI misses the real background, audience, constraint, or hidden intent.

The move: Connecting Patterns. “What real context is missing?”

Genericism

AI gives safe, average, cliché output that sounds useful but doesn't move the work.

The move: Connecting Patterns + Generating Alternatives. “What sharper, grounded alternative is missing?”

Instruction Execution Failures

The model drifts from the operating logic over a long interaction.

Instruction Dilution

AI slowly drops tone, length, format, or earlier rules during a longer interaction.

The move: Revising Beliefs + Connecting Patterns. “Has the model drifted from the operating logic?”

Evidence & Retrieval Failures

The model builds confident answers on the wrong evidence.

Retrieval Failure

AI retrieves the wrong, outdated, or irrelevant source, then builds on it.

The move: Revising Beliefs + Connecting Patterns. “Is this the right evidence for this context?”

Data Boundary Confusion

AI mixes facts, assumptions, guesses, and inferences without separating them.

The move: Revising Beliefs + Tracing Consequences. “What is known, inferred, assumed, or guessed?”

Agentic & Automation Risks

Action runs ahead of judgment. Trace the consequence before the machine acts.

Tool Misuse

AI uses the wrong tool, or trusts a tool result without checking it.

The move: Revising Beliefs + Tracing Consequences. “Was this the right tool, and what if its result is wrong?”

Automation Overreach

AI acts where human judgment, approval, or ethics review is still required.

The move: Tracing Consequences. “Should this be automated at all?”

Agent Drift

An agent or workflow slowly moves away from the goal while still looking productive.

The move: Tracing Consequences + Revising Beliefs. “Is the agent still pursuing the intended outcome?”

Where the move does not fire

Every default is a moment a specific move must fire. Where it does not, that is cognitive surrender, and it is measurable.