The Five Levels of Meaningful Transparency is a framework attributed to Nita Farahany that ranks AI transparency not by how much a system discloses but by what a person can do with the disclosure. It was introduced in Class 10 of her introductory course (September 28, 2025) and draws on Mike Hsu's "AI Actionability Over Interpretability" framing (Substack). The framework reframes the question from "is the AI system transparent?" to what action each level enables, arguing that most AI transparency stops at the lower levels and that accountability requires reaching higher ones.
The five levels
Each level is presented as conferring more capacity to act, not just more information. The framework holds that most companies operate at Levels 1-2, and that meaningful transparency requires Level 4 at minimum.
| Level | Definition | Illustrative example | What the subject can do | Who does this |
|---|---|---|---|---|
| 1. Notification | "AI was used" to make the decision | An insurance denial stating the decision was "processed using automation" | Nothing; the disclosure conveys only that a decision was made | Everyone; treated as bare-minimum legal compliance |
| 2. Explanation | "Here's roughly what happened," given as vague categories ("We considered credit, income, and other factors") | "Your risk score exceeded our threshold based on financial factors" | Guess what went wrong among the listed and unlisted factors | Some lenders, NYC's audit law, H.R. 7532's stated goal |
| 3. Interrogation | "What if I changed X?"; testing counterfactual scenarios | Adding a co-signer with a 720+ credit score yields approval; whether $10K more income would change the decision | Spot errors and identify what would change the outcome | Almost nobody; framed as revealing the system's internal logic |
| 4. Contestation | "I dispute this"; appeal with evidence, reviewable and overridable by a human | "You say I'm unemployed. Here are five years of W-2s." | Correct mistakes, add context the system missed, obtain human judgment | Europe's GDPR Article 22 (in principle), some banks |
| 5. System change | "Your dispute fixed the model"; an appeal changes the system for everyone | A case revealing bias against self-employed applicants leads to retraining and review of past decisions | Produce systemic change beyond the individual case | Nobody; framed as implying admission of systematic discrimination and associated liability |
Constraints the framework identifies
Farahany's framework identifies three constraints on transparency. First, perfect transparency is treated as impossible because systems cannot explain what their developers do not understand; the framework cites Anthropic as having mapped only about 1% of Claude's features as of May 2024. Second, partial transparency may be counterproductive, the example given being NYC's published discrimination ratios, which can signal how discriminatory a system is calibrated to be. Third, zero transparency is presented as unacceptable on accountability grounds. The framework's stated conclusion is to shift the debate from whether AI should be transparent to what level of meaningful transparency a given deployment requires.
Relation to other transparency cases
The framework is applied to several transparency-policy cases.
NYC Local Law 144 is presented as a transparency paradox: posting bias-audit numbers (Level 2 explanation) without a recourse mechanism (Level 4 contestation) is described as transparency theater that may worsen outcomes, including driving gaming toward 79.99% impact ratios.
The Houston teachers case is cited as an example in which plaintiffs sued for transparency over a $2 million algorithm and won, receiving 900 pages of formulas and 10,000 lines of code. The framework characterizes this as technical transparency at scale amounting to practical opacity, because the cost to understand the material (a $2 million expert plus statistical and legal work) far exceeded the plaintiffs' $50K budget.
H.R. 7532 is cited as requiring systems to be "sufficiently explainable" but with the qualifier "to the extent practicable," which the framework argues reduces it to Level 1 notification plus Level 2 vague explanation.
Anthropic's interpretability work is used to bound the technical ceiling: with mechanistic interpretability mapping only about 1% of model features, Levels 1-2 may be the asymptotic limit for technical transparency on frontier models, and reaching Levels 3-5 may require institutional or procedural mechanisms (appeal, contestation, retraining) rather than technical ones.
Relationships
- introduced-by: Inside My AI Law & Policy Class 10: The Glass Box Paradox (Farahany, September 2025)
- draws-on: Mike Hsu, "AI Actionability Over Interpretability" (Substack)
- related: AI Transparency, System Card Due Diligence (Clearwater's parallel framework for technical transparency), Post-Deployment AI System Monitoring
- instance-of: AI Governance (umbrella)