Skip to content

Commit 14801e1

Browse files
simabasimaba
committed
Add comprehensive README with pattern catalog and NIST RMF alignment
1 parent 1313d8f commit 14801e1

1 file changed

Lines changed: 53 additions & 70 deletions

File tree

‎README.md‎

Lines changed: 53 additions & 70 deletions
Original file line numberDiff line numberDiff line change
@@ -1,98 +1,81 @@
11
# AI Accountability Design Patterns
22

3-
[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
4-
[![Last Commit](https://img.shields.io/github/last-commit/simaba/ai-accountability-design-patterns)](https://github.com/simaba/ai-accountability-design-patterns/commits/main)
3+
[![NIST AI RMF](https://img.shields.io/badge/NIST%20AI%20RMF-Aligned-0055A4?style=flat-square)](https://airc.nist.gov/home)
4+
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg?style=flat-square)](LICENSE)
5+
[![Discussions](https://img.shields.io/badge/Discussions-Join-7289da?style=flat-square&logo=github)](https://github.com/simaba/ai-accountability-design-patterns/discussions)
56

6-
A practical pattern library for designing human accountability into AI-enabled systems — covering escalation logic, ownership models, and intervention paths.
7+
A catalog of design patterns for building accountable AI systems in regulated industries.
8+
Each pattern provides a problem statement, solution structure, implementation guidance,
9+
and mapping to NIST AI RMF and EU AI Act requirements.
710

811
---
912

10-
## Why this exists
13+
## What Is AI Accountability?
1114

12-
AI systems often fail operationally not only because of model behaviour, but because:
15+
AI accountability means that individuals and organizations can be held responsible
16+
for the outcomes of AI systems — that there are clear lines of ownership, transparent
17+
decision processes, and mechanisms for redress when things go wrong.
1318

14-
- escalation logic is vague or missing
15-
- ownership is fragmented across teams
16-
- humans are nominally "in the loop" but lack real authority
17-
- override paths are under-specified or untested
19+
The NIST AI RMF defines accountability as one of seven characteristics of trustworthy AI:
20+
> *"AI actors should be accountable for the development, deployment, and impacts of AI
21+
> systems, including supporting human oversight."*
1822
1923
---
2024

21-
## Core design principle
22-
23-
```mermaid
24-
flowchart TD
25-
A[AI system output] --> B{Intervention
26-
conditions met?}
27-
B -->|No| C[Output delivered]
28-
B -->|Yes| D[Human notified
29-
with context]
30-
D --> E{Authority to
31-
intervene?}
32-
E -->|Yes| F[Human overrides
33-
or confirms]
34-
E -->|No| G[Escalate to
35-
authorised party]
36-
F & G --> H[Decision logged
37-
and reviewable]
38-
```
39-
40-
> Human oversight is only meaningful when: intervention conditions are explicit, authority is real, context is sufficient, and decisions are logged and reviewable.
25+
## Pattern Catalog
4126

42-
---
43-
44-
## Patterns included
27+
### Governance Patterns
4528

46-
| Pattern | What it addresses |
47-
|---------|-----------------|
48-
| `patterns/human-override.md` | When and how humans can override AI decisions |
49-
| `patterns/escalation-thresholds.md` | Defining triggers for human escalation |
50-
| `patterns/ownership-models.md` | Assigning clear operational ownership |
51-
| `patterns/decision-context.md` | Ensuring humans have sufficient context to act |
52-
| `patterns/incident-accountability.md` | Post-incident ownership and review |
29+
| Pattern | Problem | Solution |
30+
|---|---|---|
31+
| **Model Inventory** | No central registry of AI systems in production | Maintain a versioned, owner-assigned inventory of all deployed models |
32+
| **Ownership Assignment** | Unclear who is responsible when an AI system fails | Assign a named technical owner and business owner to every AI system |
33+
| **AI Policy Cascade** | Governance policies not reaching practitioners | Publish policy as code — embed governance rules in CI/CD pipelines |
34+
| **Governance Gate** | AI systems deployed without appropriate review | Require signed-off checklists at defined lifecycle milestones |
5335

54-
---
36+
### Transparency Patterns
5537

56-
## Worked examples
38+
| Pattern | Problem | Solution |
39+
|---|---|---|
40+
| **Model Card** | No documentation of model capabilities and limitations | Create a structured model card for every production model |
41+
| **Decision Log** | AI decisions not auditable after the fact | Log inputs, outputs, model version, and confidence for every decision |
42+
| **Confidence Surfacing** | Users cannot tell when AI is uncertain | Surface confidence scores and uncertainty estimates in the UI |
43+
| **Explanation on Demand** | Stakeholders cannot understand AI decisions | Implement on-demand SHAP/LIME explanations for high-stakes decisions |
5744

58-
| Example | Industry context |
59-
|---------|----------------|
60-
| `examples/customer-support-agent.md` | AI-assisted customer service with override path |
61-
| `examples/ivi-assistant.md` | In-vehicle AI assistant with safety escalation |
45+
### Human Oversight Patterns
6246

63-
---
47+
| Pattern | Problem | Solution |
48+
|---|---|---|
49+
| **Human-in-the-Loop Gate** | High-stakes decisions made autonomously | Require human review before action for decisions above a risk threshold |
50+
| **Override Mechanism** | Operators cannot override erroneous AI decisions | Implement a documented, audited override pathway with reason capture |
51+
| **Escalation Ladder** | Edge cases fall through without review | Define a tiered escalation path for low-confidence or novel inputs |
52+
| **Sunset Clause** | Models remain in production past their useful life | Set mandatory model review dates; require affirmative renewal to continue |
6453

65-
## Templates
54+
### Redress Patterns
6655

67-
- `templates/accountability-review-checklist.md` — review checklist for new AI deployments
56+
| Pattern | Problem | Solution |
57+
|---|---|---|
58+
| **Adverse Action Explanation** | Affected individuals cannot understand why they were denied | Generate plain-language explanations with specific contributing factors |
59+
| **Appeal Pathway** | No mechanism for contesting AI decisions | Implement a formal appeal process with human review and documented outcomes |
60+
| **Impact Audit** | Unknown whether AI system is causing disproportionate harm | Conduct regular disparate impact audits by protected characteristics |
6861

6962
---
7063

71-
## Who this is for
64+
## NIST AI RMF Mapping
7265

73-
- AI product managers designing human-in-the-loop systems
74-
- Platform and systems engineers implementing escalation logic
75-
- Governance and risk leaders in regulated industries
76-
- Operations and quality teams accountable for AI outcomes
66+
See [docs/nist-rmf-mapping.md](docs/nist-rmf-mapping.md) for a full mapping of each
67+
pattern to NIST AI RMF functions and subcategories.
7768

7869
---
7970

80-
## Related repositories
81-
82-
This repository is part of a connected toolkit for responsible AI operations:
71+
## Ecosystem
8372

8473
| Repository | Purpose |
85-
|-----------|---------|
86-
| [Enterprise AI Governance Playbook](https://github.com/simaba/enterprise-ai-governance-playbook) | End-to-end AI operating model from intake to improvement |
87-
| [AI Release Governance Framework](https://github.com/simaba/ai-release-governance-framework) | Risk-based release gates for AI systems |
88-
| [AI Release Readiness Checklist](https://github.com/simaba/ai-release-readiness-checklist) | Risk-tiered pre-release checklists with CLI tool |
89-
| [AI Accountability Design Patterns](https://github.com/simaba/ai-accountability-design-patterns) | Patterns for human oversight and escalation |
90-
| [Multi-Agent Governance Framework](https://github.com/simaba/multi-agent-governance-framework) | Roles, authority, and escalation for agent systems |
91-
| [Multi-Agent Orchestration Patterns](https://github.com/simaba/multi-agent-orchestration-patterns) | Sequential, parallel, and feedback-loop patterns |
92-
| [AI Agent Evaluation Framework](https://github.com/simaba/ai-agent-evaluation-framework) | System-level evaluation across 5 dimensions |
93-
| [Agent System Simulator](https://github.com/simaba/agent-system-simulator) | Runnable multi-agent simulator with governance controls |
94-
| [LLM-powered Lean Six Sigma](https://github.com/simaba/LLM-powered-Lean-Six-Sigma) | AI copilot for structured process improvement |
95-
96-
---
97-
98-
*Shared in a personal capacity. Open to collaborations and feedback — connect on [LinkedIn](https://linkedin.com/in/simaba) or [Medium](https://medium.com/@bagheri.sima).*
74+
|---|---|
75+
| [enterprise-ai-governance-playbook](https://github.com/simaba/enterprise-ai-governance-playbook) | End-to-end governance playbook |
76+
| [ai-release-readiness-checklist](https://github.com/simaba/ai-release-readiness-checklist) | Release gate framework + CLI |
77+
| [ai-risk-taxonomy](https://github.com/simaba/ai-risk-taxonomy) | Structured AI risk taxonomy |
78+
| [nist-ai-rmf-implementation-guide](https://github.com/simaba/nist-ai-rmf-implementation-guide) | NIST AI RMF practitioner guide |
79+
| [awesome-ai-governance](https://github.com/simaba/awesome-ai-governance) | Curated governance resources |
80+
81+
*Maintained by [Sima Bagheri](https://github.com/simaba) · Connect on [LinkedIn](https://www.linkedin.com/in/simabagheri)*

0 commit comments

Comments
 (0)