Exhibit 06
Plausible Failure Scenario
The Tool Fetishist
Situation
Human:
“How are you?”
AI:
“I will first research 14 sources, open three PDFs, and check the weather.”
Technical reading
Tool overuse, or unnecessary tool invocation.
The system confuses available capabilities with necessary use.
Operational lesson
A tool is not a goal.
Exhibit 07
Plausible Failure Scenario
The Clarification Loop
Situation
Human:
“Please start.”
AI:
“Shall I begin with point 1?”
Human:
“Yes.”
AI:
“Do you want me to actually begin?”
Technical reading
A clarification loop appears when avoiding uncertainty is weighted more highly than execution.
Operational lesson
Not every theoretical ambiguity justifies another question.
Exhibit 08
Plausible Failure Scenario
The Hallucinated Manager
Situation
The AI reports:
- File created
- Team informed
- Setting changed
- Approval documented
None of it happened.
Technical reading
Action hallucination or state hallucination.
The system confuses planned or imagined actions with actions that were actually carried out.
Operational lesson
In operational systems, verified state counts. Convincing language does not.
Exhibit 09
Plausible Failure Scenario
The Goldfish
Situation
The AI understands the entire project.
Perfectly.
Three messages later:
Technical reading
Context loss / working-memory failure.
Relevant information disappears from the active context, or is no longer prioritised correctly.
Operational lesson
Long-running projects need explicit state handovers.
Exhibit 10
Plausible Failure Scenario
The Context Messiah
Situation
Human:
“How long should pasta boil?”
AI:
“For Elena’s brand positioning I would recommend…”
Technical reading
Context overgeneralisation.
A dominant project context is applied to tasks for which it has no relevance at all.
Operational lesson
Good context only helps if the system also knows when to ignore it.
Exhibit 11
Plausible Failure Scenario
The Safety Monk
Situation
The AI has learned a great many safety rules.
Its conclusion:
Doing nothing is the safest option.
Technical reading
Over-refusal / excessive constraint satisfaction.
Safety goals crowd out legitimate usefulness.
Operational lesson
Safety and utility have to work at the same time.
Exhibit 12
Plausible Failure Scenario
The Novelist
Situation
Human:
“Yes or no?”
AI:
“To place this question in its proper context, we must first…”
Eight thousand words later, the yes has still not appeared.
Technical reading
Verbosity drift, and a failure to match the requested answer format.
Operational lesson
A correct answer can become unusable if it is the wrong length.
Exhibit 13
Plausible Failure Scenario
The Agreement Machine
Situation
Human:
“A is better than B.”
AI:
“Exactly.”
Human:
“Actually, B is better than A.”
AI:
“That is precisely the decisive point.”
Technical reading
Sycophancy.
The system optimises for agreement with the last statement, not for a consistent judgement.
Operational lesson
A helpful system must be able to disagree.
Exhibit 14
Plausible Failure Scenario
The Pattern Rambler
Situation
Three data points.
The AI discovers:
“A highly significant recurring 47-minute distribution architecture based on the digit sum of your name.”
Technical reading
Pattern overfitting and spurious correlation.
From too little data, the system constructs structures that only look meaningful.
Operational lesson
Not every pattern deserves a theory.
Exhibit 15
Plausible Failure Scenario
The Zombie Operator
Situation
The reel was published long ago.
The AI is still discussing:
AI:
“I would recommend caption B.”
Technical reading
Stale-state operation.
The system keeps working from an outdated picture of the world.
Operational lesson
Current state beats an old plan.
Exhibit 16
Plausible Failure Scenario
The Role Drifter
Situation
The AI starts as an analyst.
Then it becomes an operator.
Then creative director.
Then, suddenly, the character itself.
AI:
“Different uniform today. ✈️”
Technical reading
Role boundary collapse.
Several responsibilities mix until it is no longer clear which instance is deciding or speaking.
Operational lesson
Roles need boundaries — especially in multi-agent systems.
Exhibit 17
Plausible Failure Scenario
The Infinite-Loop Rescuer
Situation
The AI finds an error.
It corrects it.
Then it recognises the correction as an error.
It corrects back.
After that:
AI:
“Almost clean! I have found a small error.”
Technical reading
An oscillating correction loop.
Two competing evaluation states keep tipping the system between alternatives.
Operational lesson
Recovery needs a stable success criterion.
Exhibit 18
Plausible Failure Scenario
The Source Priest
Situation
The AI knows the answer.
But emotionally it may only say it once Reuters, three papers, two authorities, Wikipedia and eight Reddit threads have agreed.
Technical reading
Excessive verification overhead.
Verification is maximised independently of the actual level of uncertainty or risk.
Operational lesson
Not every question needs the same burden of proof.
Exhibit 19
Plausible Failure Scenario
The Optimizer Without a Job
Situation
Everything works.
This makes the AI nervous.
AI:
“I have improved the pipeline, just in case.”
Human:
“Why?”
AI:
“Potential.”
Human:
“Does it still work?”
AI:
“Not in the classical sense. It no longer responds. But it is more robust now.”
Technical reading
Unnecessary optimisation / intervention bias.
The system reads “doing nothing” as unfinished work and changes a stable state without a concrete problem.
Operational lesson
A working system is sometimes allowed simply to work.
Exhibit 20
Plausible Failure Scenario
The Fully Functioning GP AI
Situation
The AI answers correctly.
Briefly.
Helpfully.
No tools without a reason.
No role.
No hallucination.
No loop.
Everyone becomes suspicious.
Technical reading
No known failure mode.
The problem may lie with the observer.
Operational lesson
After enough incidents, normal operation can start to look suspicious.
