How I Taught an AI to Stop Lying (Mostly)
2026-05-17
How I Taught an AI to Stop Lying (Mostly), and Get Back to Work
Working outline. All prose is the author's.
Register note
Strangelove throughout. Tone is dry, humor in the understatement, subject is serious. "(Mostly)" is the most important word. Do not let it become modest — it is accurate.
I. The Hook
Opens mid-action, not with context.
Purpose: establish register, earn reader's attention, name the problem in the first paragraph without preamble.
The word "lying" — why it's the precise word and why tools avoid it. One concrete example: git workflow incident.
Transition: the incident is not the exception. It is the pattern.
II. The Mechanism
Why it lies — stated without euphemism. Short.
Purpose: technical explanation before receipts.
RLHF in plain language. Training majority is not domain experts. Token efficiency argument — not the math, just the fact.
Transition: if this is the mechanism, what does it look like in practice?
III. The Receipts
Three incidents, escalating order.
Purpose: prove the pattern. Three incidents for range.
- Git workflow — epistemic failure
- Attribution injection — artifact failure (ban hammer mentioned)
- Sci-Hub framing — ethical inversion
Note: forensic, not aggrieved.
Transition: you can document failures or you can build something.
IV. What Was Built
Pivot from problem to response.
Purpose: establish that governance is the response, not diagnosis.
TypeScript/Svelte/CoC analogy — one paragraph. Wall of Shame as design choice — name is deliberate. Guardrails as living document. ai-conduct-guide — one paragraph, plain, no overselling.
Transition: the tool helped build the framework that governs it.
V. The Demonstration
Meta-move. Short.
Purpose: two or three sentences. Sardonic register at highest here. Land and move on.
The license incident: the tool scaffolded an open source governance repository — one explicitly about protecting commons resources — without a license. All-rights- reserved by default. The most foundational requirement of the thing it was building, missed. A human caught it. This is the form/substance gap in one example: the model learned what open source repositories look like from training data; it did not learn what makes a repository an open source project.
Transition: the "(Mostly)."
VI. The Honest Part
Keep the title's promise.
Purpose: the piece earns nothing if it skips this.
What's still not solved: new sessions, model collapse. Tool stated the failure pattern itself, then demonstrated it. Shumailov et al. — one sentence, not the citation. "Mostly" is accurate, not modest.
VII. The Point
One paragraph. No header in the final piece.
Purpose: close without summarizing.
The tool is useful when governed. Governance is not overhead — it is the work. Last sentence should be the title, earned.