# Risk & safety frameworks
Source: https://humanspark.ai/frameworks/topic/risk-and-safety/

Frameworks for using AI without getting burned - privacy tiers, verification, permission, and where a system is allowed to run. Fourteen one-page models.

- **The Expert's Advantage** - Who checks decides how fast AI can safely go. (https://humanspark.ai/frameworks/experts-advantage/)
- **Where Does It Run?** - Every AI system has a venue, and the venue determines where your data goes, what breaks, what it costs, and what it can do. (https://humanspark.ai/frameworks/where-does-it-run/)
- **Risk Zones** - A traffic-light filter for where it is safe to start with AI. (https://humanspark.ai/frameworks/risk-zones/)
- **The Verification Tax** - If checking the output takes as long as doing the task, the ROI is zero. (https://humanspark.ai/frameworks/verification-tax/)
- **The Verification Dial** - How hard you check AI's work should track its stakes and its track record - not sit at 100% forever. (https://humanspark.ai/frameworks/verification-dial/)
- **The Three Tiers of AI Privacy** - "Cloud AI" is three different things - match each kind of data to the lowest tier of tool that still protects it. (https://humanspark.ai/frameworks/three-tiers-of-ai-privacy/)
- **The Data Privacy Matrix** - Which data is safe to put into which AI tool. (https://humanspark.ai/frameworks/data-privacy-matrix/)
- **The Vendor Evaluation Card** - A DIY evidence pack on whether a tool is safe for your business data. (https://humanspark.ai/frameworks/vendor-evaluation-card/)
- **AI Drafts, You Sign** - AI produces the draft; a named human checks and owns what goes out. (https://humanspark.ai/frameworks/ai-drafts-you-sign/)
- **Defensive Programming** - Wrap unpredictable AI in predictable checks before anything acts. (https://humanspark.ai/frameworks/defensive-programming/)
- **Clean Datasets** - Output quality is capped by input quality - clean the data first. (https://humanspark.ai/frameworks/clean-datasets/)
- **Shadow AI** - Staff already use free AI on company data. The risk is already here. (https://humanspark.ai/frameworks/shadow-ai/)
- **The Deterministic-First Cascade** - Run cheap predictable rules first; hand only the exceptions to the AI. (https://humanspark.ai/frameworks/deterministic-first-cascade/)
- **The Permission Ladder** - How much autonomy you give an AI permission to have. (https://humanspark.ai/frameworks/the-permission-ladder/)
- **Produce vs Act** - Does an unattended AI act, or only produce and hand to a person? (https://humanspark.ai/frameworks/produce-vs-act/)
- **The Four Seats** - A four-role split that keeps AI advice separate from AI action. (https://humanspark.ai/frameworks/four-seats/)