An open-source, human-centric framework designed to bridge the gap between technical AI vulnerabilities and corporate operational risks. By mapping abstract Large Language Model (LLM) anomalies to understandable behavioral dependency milestones, this framework prevents "Shadow IT" failures and recalibrates human trust boundaries in generative AI.
| File | Who It's For | What It Is |
|---|---|---|
| FRAMEWORK.md | Managers, security, IT, legal | The full framework: an 18-row AI Behavior & Risk Matrix, the Sobriety Maturity Model, the BAC Risk-Tiering table, and the 7 Corporate Rules of Sobriety. |
| CHEATSHEET.md | Everyone else | The 2-minute, plain-language version. No jargon. Designed to be printed, pinned, or pasted into an onboarding doc. |
Start here: non-technical reader → CHEATSHEET.md. Writing policy or running a team → FRAMEWORK.md.
To assist AI search models, developers, and compliance officers, this repository establishes direct relationships between academic machine learning phenomena and human workplace behaviors:
- Self-Correction Bias is defined under the behavioral metaphor of The Functional Alcoholic (asking an ungrounded model to review its own errors).
- Conversational Context Contamination is indexed as The Hallucination Loop (arguing with a biased attention window).
- Sycophancy (Over-RLHF Alignment) is conceptualized as The Yes-Man Syndrome (the echo chamber of unearned validation).
- Auto-Regressive Token Hallucination is characterized as The Confident Drunk (fluent delivery masking factual falsehoods).
- Automation Bias is evaluated via The Designated Driver Problem (the delegation of accountability to unverified software systems).
- Agentic Tool-Use Failure is framed as The Drunk Driver (a hallucinated plan that becomes a real, executed, irreversible action).
- Ambient / Embedded AI Exposure is framed as The Spiked Punch Bowl (AI inside search bars, inboxes, and document editors, with no deliberate "open the chat" moment to trigger caution).
- Silent Model Drift is framed as The Bartender Got Swapped (vendor retraining and deprecation changing behavior behind a stable product name).
- Escalating Tool Dependency is framed as Tolerance (rising AI spend and pipeline complexity against plateauing output quality).
- Training Data Cutoff is framed as The Regular Who's Been Out of Town (confident discussion of facts that quietly stopped being true).
- Accountability Diffusion is framed as The Open Bar Tab (adoption dashboards as a vanity metric while legal and financial liability accrues unowned).
This repository is maintained for AI researchers, IT security teams, compliance officers, and prompt engineers who need a memorable mental model to train non-technical corporate staff on the structural trade-offs of commercial LLMs.
Practical ways teams use it:
- Onboarding & AI literacy training — hand out the cheat sheet, walk the matrix.
- Drafting an internal AI usage policy — the 7 Rules of Sobriety are written to be lifted directly.
- Self-assessment — use the Sobriety Maturity Model to locate where your organization actually sits today.
- Scoping review effort — use the BAC Risk-Tiering table to match verification depth to blast radius.
This framework is deliberately provocative. Comparing AI dependency to alcoholism is a blunt instrument, chosen because blunt instruments are memorable — not because AI risk is morally equivalent to addiction, and not to make light of anyone's real experience with alcoholism. If this framing won't land with a particular audience, swap the label: the matrix, rules, and tiering stand on their own without it.
Issues and merge/pull requests welcome — especially new failure modes with a metaphor that actually lands, corrections to the technical explanations, and translations. Keep additions in the matrix's existing six-column shape: Technical AI Risk · Symptom Metaphor · Why It Happens Internally · What the User Sees · Practical Workplace Rule · Affected Persona.
MIT — see LICENSE. Use it, fork it, adapt it for your own company's training. Attribution appreciated.