3-Shot Learning #3
Lawyers’ Committee legislation, Midjourney v6, and Superalignment grants
Welcome to 3-Shot Learning, a weekly newsletter from the Center for AI Policy. Each issue explores three important developments in AI, curated specifically for AI policy professionals.
Lawyers’ Committee Pens Model Legislation to Protect Civil Rights
A new Online Civil Rights Act was proposed by the Lawyers’ Committee for Civil Rights Under Law, a nonprofit organization dedicated to civil rights and racial justice. The Act aims to prevent a variety of harms from AI systems, especially harms relating to discrimination and bias, which the President of the Lawyers’ Committee described in detail in his 50-page testimony for a recent Senate hearing.
One notable feature of the Act is its sequential procedure for evaluating AI systems before and after they are released on the market. First, before releasing an AI algorithm, a company would need to conduct a preliminary assessment to determine whether that algorithm could plausibly cause harm. If so, the company would then work with an independent auditor to assess how harm might arise, how to reduce this risk, and so on. After releasing the system, the company would follow a similar procedure to conduct annual (post-deployment) assessments. Companies would share the results of all this work with the FTC, and make straightforward summaries available to the public.
To ensure that companies actually comply with these pre- and post-deployment audit processes, the Act contains enforcement measures such as FTC authority and private rights of action. Furthermore, the Act would hold companies liable for failing to take reasonable measures to prevent the harms identified in the audits.
Midjourney Releases Its Latest Image Generation System
AI startup Midjourney has introduced the newest version of its text-to-image AI model, dubbed “Midjourney v6,” which the startup claims is “now MUCH better at understanding you.” Remarkably, Midjourney might exceed $200 million in revenue this year, a significant achievement for an organization with just a few dozen employees, especially when competing against giants like Meta, OpenAI, and Adobe. Midjourney customers have used the Midjourney AI system to create newsworthy images, including not only award-winning artwork, but also (fake) viral photos of Donald Trump’s arrest and the “Balenciaga Pope.” However, the startup has attracted concerns regarding copyright infringement, and is facing an ongoing lawsuit.
OpenAI Launches Grant Program Focused on Aligning Superintelligent AI
With help from a $5M donation from Eric Schmidt, OpenAI is offering $10M in grants to support research on “one of the most important unsolved technical problems in the world”: finding technical guardrails to control AI systems that vastly exceed the intellect of any human. The grants will support research directions such as interpretability and honesty, as well as techniques to measure AI systems’ abilities to autonomously copy themselves or deliberately deceive humans.
News at CAIP
Jason Green-Lowe and Joe Buccino published an op-ed that highlights important gaps in Biden’s AI Executive Order.
Quote of the Week
Autonomous targeting and harming of human beings by machines is one of those “red lines” that should not be crossed.
—the UN’s AI Advisory Body, in its interim report on AI governance