Our Standards Development Process
Guidelight's standards describe what responsible frontier AI development looks like in concrete terms, giving companies a target to meet, internal advocates a reference to point to, and external observers a basis for understanding companies' positions.
Our first standards cover Control and Transparency, which are each v1.0 and will continue to be updated. We will publish additional standards over time.
How each standard is organized
Each standard begins with a brief explanation of the risk management concept it covers, followed by three components:
Principles state a high-level goal for adequate risk management in this area, intended to be simple, timeless, and relatively self-evident.
Practices specify concrete actions that are important for developers to take to achieve the principle. We expect practices to evolve over time, generally becoming more ambitious as companies establish what is possible and as new state-of-the-art techniques are established.
Directions-for-development identify where new practices are needed, but where the specifics aren't yet worked out. In these cases, we describe the direction we believe the field needs to move in, and we encourage companies to experiment toward determining what the eventual practice should look like.
How we develop our standards
To develop our standards, we draw on published research on frontier AI safety, frontier developers' own safety frameworks and commitments, laws and guidance from governments and standards bodies, risk management practices from other high-risk industries, conversations with practitioners across the industry, and our team's experience working in frontier AI.
From these sources we build a candidate set of principles and practices, then compare them to what leading developers are already doing and to our analysis of what is most important for avoiding catastrophic risks.
Before publishing a standard, we:
Review the existing literature and draft tentative materials.
Gather feedback from experts across the field (for instance, about practicality, importance, and the current state of implementation).
Iterate on our draft based on that feedback.
Standards are published as a v1.0, after which we solicit continued feedback on a rolling basis.
Prior to conducting an assessment for a standard (e.g., measuring how well companies are adhering to the principles in a category, as gauged by its practices), we notify the frontier AI companies we plan to assess. We then conduct an initial private assessment, share our conclusions with each company for feedback and fact-checking, and publish our assessment findings.
Importantly, our first standards and our processes are living documents, which we expect to revise over time.
Our standards are intended as a North Star, which often go beyond existing frontier AI regulation, such as California's SB 53, New York's RAISE Act, and the EU's AI Act and its General-Purpose AI Code of Practice. Many of our practices reflect behaviors we think are important but that aren't yet required by these laws; other practices represent more precise ways of operationalizing what those laws are aiming at.