Anthropic trust nominates four of the company’s seven directors

The $1.4 trillion blockbuster listing of Anthropic, the Claude AI developer, happens this week, and the governance setup is, to put it mildly, different, Nils Pratley wrote in The Guardian. In essence there are two centres of power and neither is the ordinary shareholders, Pratley wrote: the seven founders on one side, and a financially disinterested trust on the other. Boardroom governance matters, even in the world of big tech, he added — investors are discovering that the business of how companies are run can be more complicated than handing the founder or superstar chief executive as much control as he wants and waiting for brilliant returns to appear, and that sometimes a few checks on power are useful, especially when a crisis strikes.

Pratley placed the setup against the governance consensus it departs from. For about 30 years until the first generation of big US tech companies came along, he wrote, the thrust of corporate governance thinking was to insist on the basic fairness of the principle of “one share, one vote” — the owners should be able to remove misbehaving or underperforming managers by a simple majority.

Dual-class structures and unequal voting rights got a fresh lease of life when Google, now Alphabet, and Facebook, now Meta, appeared with their “mission” statements, Pratley wrote. The corporate missions, the argument went, were so important that the founding geniuses needed to be able to pursue them in free-wheeling, risk-taking style without distraction from the short-term demands of ordinary shareholders. In an extreme example, Snap, the Snapchat company, was able to pitch its ambition of “reinventing the camera” as so vital that shares could be issued with zero voting rights, as opposed to merely weakened rights.

Anthropic’s structure has two centres of power, neither of them the ordinary shareholders. One is the seven founders, who collectively have one vote that commands 50.1% of the ultimate voting power — the collective aspect is designed to force them to come to a consensus. The other is the Long-Term Benefit Trust, a financially disinterested body that nominates four of the seven directors and looks out for the balance between making money and benefiting humanity. Pratley described the trustees as “a cast of the great and the good”; they include Ben Bernanke, former chair of the US Federal Reserve. The whole structure is given an overlay as a “public benefit corporation,” a designation reflecting Anthropic’s stated commitment to develop AI responsibly and safely. Dario Amodei leads the founders.

Pratley identified two positive features. First, he wrote, the motivation is plainly more complicated than a simple power grab by the founders led by Dario Amodei: Anthropic’s backing for the Long-Term Benefit Trust seems to be designed to create some form of constitutional balance of power. Second, he wrote, Anthropic probably has a stronger justification than most to depart from the norm of equal rights for equal economic risk. If the development of AI could pose “catastrophic or existential risks to humanity”, as the listing document puts it, a few unconventional governance structures are probably in order. “We’re not talking about Snap’s fluffy videos for social media,” Pratley wrote.

All the same, Anthropic is offering an untested experiment, Pratley wrote. For all the fine words, the company may simply have recreated an inward-looking “trust us” setup, just with more elaborate clothing.

Stuart Riddick, sustainable investment manager at fund manager Aberdeen, supplied a more pointed formulation. “Agreement with a mission in principle does not equate to agreement over how it is interpreted and implemented by the directors,” Riddick said. Riddick noted that Mustafa Suleyman, the chief executive of Microsoft’s AI division, even as he acknowledged Anthropic’s seriousness about AI safety, has made strong and detailed criticisms of the company’s thinking on so-called “model welfare”, arguing that AI models must not be trained to act as if they have rights or consciousness.

How, in practice, such challenges are handled within Anthropic’s supposedly responsible governance arrangements is hard to tell, Pratley wrote, unless the Long-Term Benefit Trust starts to throw its independence around.

Pratley set out the unanswered questions the structure leaves: what happens if the trustees are at loggerheads with the founders on a basic safety question, and will they tell the outside world, or reflect that ultimate voting still sits with the founders so best not to rock the boat? If the shareholders are denied any role in holding the executives to account, he asked, what is the true level of challenge? It is surprising after events at Nvidia that Anthropic’s governance has received so little attention, Pratley wrote. The setup is supposed to be high-minded and responsible, he concluded, but it still represents an extraordinary concentration of power in the hands of a few individuals.