Anthropic has filed its IPO prospectus, and MIT Technology Review’s weekly AI newsletter, The Algorithm, uses the moment to ask a pointed question: what exactly is Anthropic, and how much should we trust the story it tells?

What the filing says

The document runs 335 pages, with roughly 100 of them devoted to risks. It warns that the company’s technology could damage trust, harm society and destabilize the geopolitical order, and that in a worst case it could pose catastrophic or existential risks to humanity. It also asks prospective buyers to see a purchase as more than a transaction. The newsletter notes that one risk factor even covers debates about whether AI models could be sentient and what obligations that might create.

A company that tells big stories

Author Will Douglas Heaven argues the filing fits a pattern of narrative-building. He points to two examples:

• Cofounder Jack Clark told an IMF audience in Bangkok this month that the singularity has already begun.
• Cofounder Chris Olah told Vatican officials last year that Anthropic keeps finding unsettling things inside its models, including signs of internal states that resemble human emotions.

Heaven notes that many researchers would push back on suggestions of humanlike inner experience, and that in a world where nobody knows what comes next, a company like Anthropic gets to fill the gap with its own account.

Safety group, megacorp, or both?

The newsletter was prompted by an exchange on X. Nathan Calvin of Encode AI criticized a draft bill Anthropic circulated privately to US lawmakers, which later leaked, and asked what Anthropic really is. The piece offers several answers: a company whose IPO values it above $1.5 trillion, a safety-focused lab, or a group racing to build superhuman technology first. CEO Dario Amodei replied that Anthropic’s strategy has always been a race to the top, showing that building AI responsibly is possible so that the whole industry raises its standards.

Heaven’s view is that Anthropic is all of these at once, and that juggling roles creates real tension between a safety mission and a profit-driven business. A plainer technology company, he suggests, would be easier to read.

Why it matters: how much to trust AI labs to regulate themselves, and how to read their lobbying, depends on which of these identities you think is in charge.

Source: MIT Technology Review, The Algorithm. Image: Stephanie Arnett/MIT Technology Review | Getty Images.