Cost-of-Wrong

Definition

Cost-of-Wrong is the risk axis at the center of Greg Kihlström‘s AI Capability Framework, from the book Stop Saying “AI”. It asks a single blunt question about any AI capability: when this is wrong, what does it cost, and would anyone notice in time?

Cost-of-wrong is deliberately independent of the mechanism used to build a capability and independent of how impressive the technology is. A five-line rule and an agentic loop can both carry a catastrophic cost-of-wrong, or a trivial one. What sets the cost isn’t the sophistication of the system. It’s the consequence of the error and whether it surfaces before damage is done.

The concept does two jobs in the framework. First, it sets the accuracy bar that the lightest-mechanism discipline aims at — a high cost-of-wrong demands a higher bar, which is what justifies reaching for a heavier mechanism or a heavier Trust layer. Second, it’s the hinge of the book’s fade thesis. The Fade is uneven precisely because cost-of-wrong varies: where it’s low, a capability can safely disappear into the background; where it’s high, the capability stays named and watched.

How It Relates to Marketing

Cost-of-wrong is the sorting principle a marketing leader uses to decide where to spend oversight. Marketing runs a wide range of AI, and treating all of it with the same caution is both expensive and, for the high-stakes cases, insufficient. Cost-of-wrong separates the piles.

• Low cost-of-wrong. A weak subject-line variant trims the open rate a little, and someone sees the number the next morning. A slightly-off product recommendation gets ignored. These can fade into the workflow.
• High cost-of-wrong. A mispriced quote, a wrongful discount, an inaccurate claim in an ad, a message that violates a consent rule — each costs money, a customer, or a regulator’s attention, and some of them fail silently. These stay named, logged, and reviewed.

The line between the piles doesn’t track how advanced the AI is. A humble automated email can carry a high cost-of-wrong if it goes to the wrong segment with the wrong promise, while a sophisticated generative tool drafting internal notes carries almost none.

The Two Dimensions of Cost-of-Wrong

Cost-of-wrong has two components, and both matter:

1. Magnitude — how much a wrong output costs. Money, a lost customer, legal exposure, reputational damage, or an account of yourself you owe someone angry. The larger the magnitude, the higher the bar.
2. Visibility — whether anyone notices in time. A loud error that surfaces immediately is far safer than a quiet one that looks like success. Because AI fails silently, a confident wrong answer can look identical to a confident right one, which makes low visibility its own source of risk.

The dangerous quadrant is high magnitude paired with low visibility — expensive errors that don’t announce themselves. That combination is what keeps a capability firmly in the stay-named pile no matter how routine it looks.

How to Assess Cost-of-Wrong

1. Name the failure. Describe concretely what a wrong output looks like for this capability.
2. Price the magnitude. What does that specific error cost in money, customers, compliance, or trust?
3. Judge the visibility. Would the error be caught quickly and loudly, or could it pass silently as success?
4. Place it on the axis. Combine magnitude and visibility into a high or low cost-of-wrong.
5. Route the result. Use the placement to set the accuracy bar for mechanism choice, to decide fade versus stay-named, and to set the supervision dial.

Comparison to Similar Concepts

Concept	Measures	Relationship
Cost-of-Wrong	What an error costs and whether it’s noticed	The framework’s risk axis
Risk = likelihood × impact	Probability-weighted exposure	Cost-of-wrong emphasizes the impact-and-visibility side
Algorithmic Aversion	Human reluctance to trust algorithms	A downstream effect; high cost-of-wrong intensifies aversion
Blast radius	Scope of a failure’s damage	Close kin; cost-of-wrong adds the visibility dimension

Cost-of-wrong differs from a standard risk calculation by foregrounding visibility. Classic risk framing weighs likelihood against impact. Cost-of-wrong insists that a low-visibility error — one that fails silently — is more dangerous than its raw impact suggests, because for a probabilistic technology, silent failure is the characteristic failure.

Best Practices

• Judge the task, not the technology. Cost-of-wrong is a property of the consequence, not of how advanced the AI is.
• Weight silent failure heavily. Low visibility is a risk multiplier, because AI’s errors don’t announce themselves.
• Let it set the bar. A high cost-of-wrong is the justification for a heavier mechanism or a heavier Trust layer, not the technology’s capability.
• Re-assess when scope changes. A capability’s cost-of-wrong rises the moment it starts touching a decision, a price, or a regulated action it didn’t before.

Future Trends

• Cost-of-wrong becomes the unit of governance. As “AI” fades as a category, oversight increasingly attaches to specific capabilities and their cost-of-wrong rather than to the technology label.
• Visibility tooling matters more. Investment shifts toward catching silent failures — monitoring, confidence calibration, provenance — because that’s where the dangerous quadrant lives.
• The line moves as autonomy grows. As capabilities slide toward the unsupervised end of the supervision dial, more of them cross into high cost-of-wrong, raising the premium on getting the assessment right.

FAQs

1. What is cost-of-wrong? The measure of what a wrong AI output actually costs and whether anyone would notice in time. It’s the risk axis of the AI Capability Framework.

2. Does it depend on the mechanism? No. A simple rule and an agentic loop can both carry a high or low cost-of-wrong. The cost comes from the consequence of the error, not the sophistication of the system.

3. Why does visibility matter so much? Because AI fails silently. A confident wrong answer looks like a confident right one, so an expensive error that isn’t noticed in time is the most dangerous case.

4. How does it connect to the fade? Cost-of-wrong is why the fade is uneven. Low cost-of-wrong lets a capability fade into the background; high cost-of-wrong keeps it named and watched.

5. How does it connect to mechanism choice? It sets the accuracy bar. The higher the cost-of-wrong, the higher the bar, which is what justifies a heavier mechanism or a heavier Trust layer.

Related Terms

1. The AI Capability Framework (Master Stack)
2. The Fade / Stay-Named Line
3. The Fade (Uneven Disappearance)
4. The Supervision Dial
5. The Spine
6. The Lightest-Mechanism Discipline
7. Human-in-the-Loop (HITL)
8. Algorithmic Aversion
9. AI Governance Board (AIGB)

Sources

• Kihlström, Greg. Stop Saying “AI”. https://amzn.to/4wilWcA
• Greg Kihlström — official site. https://www.gregkihlstrom.com