AI literacy
Probabilistic = the model samples — fluent, varied, surprising. Deterministic (operator sense) = you constrain inputs and rules so the output matches your business, brand, and checklist.
How it shows up
Ask the same question three times in a plain chat window. You often get three plausible answers. That is useful for exploration — and dangerous for assets that must stay on-brand (offers, infographics, client language, Codex reading experience).
In the room 2026-08-19, Martine and Branka named the risk: without a harness, everything online starts to look the same — generic AI sameness. Differentiation lives in what you constrain, not which logo is on the chat box.
Probabilistic (when to welcome it)
- Early ideation — category sentences, metaphor hunts, “what if” passes.
- Interview content, story variance, exploration blocks (Probabilism on the path).
- Learning a new domain before you commit doctrine.
Cost: fluency feels like truth. Pair with Critical thinking on AI — three counter-questions before you ship.
Deterministic (when to require it)
- Visuals that must share one signature (tokens, palette, layout rules).
- Coaching language tied to your Manual and Battle.
- Business brand voice — what the company column owns in Tripartite reins — not a vibe you improvise each Tuesday.
- Publish paths where teachability and readability are non-negotiable (Teachability index).
Mechanism: named files, mega-prompts, Gems, harness steps — so the model reads your context before it writes.
Map to Atlas (deeper)
| AI literacy | Atlas shelf |
|---|---|
| Deterministic output | Determinism |
| Exploration & surprise | Probabilism |
| Inspectable wiring | The Glassbox Mindset |
Deterministic output means controlling critical links; probabilism parks variance on the right link; glassbox keeps wiring inspectable.
Your move this week
Pick one asset (post, slide, client email). Run it once in raw chat — then once with a fixed instruction file + source doc. Diff the two outputs: what changed when the harness appeared?