The Future. Anthropic’s “Collective Constitutional AI” experiment put governance in the hands of the American people, testing whether a written set of rules (like Isaac Asimov’s “Three Laws of Robotics”) could program advanced AI systems to be safer for society. The move provides a potential alternative to government regulation, making things open source, or putting too much power in the hands of fanatics in Silicon Valley. Ultimately, the experiment may force the world to reckon with what universal human values must be coded in our machines.
Of the AI, by the people
The company gave 1,000 adults — selected with the help of the Collective Intelligence Project, crowdsourcing site Polis, and survey site PureSpectrum — a set of principles and asked which ones they agreed with and which ones were missing.
- Those principles were whittled down to a list of 75, including ones in Anthropic’s in-house constitution (AI should be trustworthy, AI shouldn’t be harmful) and new ones (AI should accommodate users with disabilities).
- When the “public constitution” was finalized (Anthropic kept ultimate approval power), the system was integrated into a test version of the company’s chatbot, Claude.
- Compared to the company’s standard Claude, both performed nearly the same… but the one with the public constitution was slightly less biased.
Anthropic has seemingly been a little more focused on the safe development of AI than its competitors, basing its system’s constitution on a number of high-profile third-party texts, including the United Nations’ Universal Declaration of Human Rights and Apple’s terms of service (a hilarious juxtaposition from NYT’s Kevin Roose).
But the firm hopes to find a solution that offloads some of the weighty responsibility of crafting AI guardrails to a more democratic process (a move that Meta has tried with its Oversight Board). But as more voices enter the fray, the more complex the process may become — whose voice from which background should Claude adhere to? Should Claude have a different constitution depending upon which country it operates in?
Complicated stuff… much like running a democracy.