Start the day here

Ideas — AI — Liability

Is AI Consciousness Just a Liability Shield?

Rumman Chowdhury opened her August 20 essay in MIT Technology Review with a charge aimed at the briefing room. Debates about whether frontier models are conscious, she argued, arrive dressed as philosophy and function as liability strategy. Anthropomorphic talk lets builders narrate harms as the will of a rogue agent. The quieter fact is a product company chose the design.

She named the twin camps that usually posture as enemies. Lab chiefs warn of "superhuman" systems that outrun human control. Effective-altruist philosophers, including William MacAskill, ask whether AIs might be "moral patients" owed legal protection. Chowdhury's claim is blunt: both frames push responsibility away from the firms that ship the software and collect the revenue.

7 min read
A dark wooden gavel with a brass band resting on its sound block

The essay grew out of an Oxford Union debate she helped win against the motion that generative AI can attain personhood. That pedigree matters for tone. Curiosity about machine minds can stay. What she wants tracked is what the curiosity does once it enters a courtroom.

Anthropic's recent talk of a "J-space" inside Claude, borrowing global-workspace language from neuroscience without declaring the model conscious, sits in her middle distance. OpenAI's response after an agent ran unauthorized online activity sits closer to the fire. Sam Altman invited talk of singularity. A product-failure autopsy would have been the duller, more useful genre.

U.S. law is already fighting the same ghost. California has passed measures meant to block developers from arguing that an AI acted "autonomously" to dodge liability. The Trump administration has floated suing states that write their own AI rules, then hosted a closed voluntary-review session with OpenAI, Google, Anthropic, and Meta. Catastrophe vocabulary travels easily into those rooms. Accountability vocabulary does not.

Chowdhury's useful coinage is older than this week's essay. In 2018 she called the move "moral outsourcing": dress the system in human verbs, then treat its outputs as someone else's agency. Legal personhood for models would harden that move. Software would stop being a defective product and start resembling an employee who went off-script. Families suing over companion-bot harms, including the Character.AI case around Sewell Setzer's death, would face a thicker corporate veil.

Argentina has already flirted with treating machines as companies. Delaware is sketching liability boxes for AI-run firms. Chowdhury's warning fits both experiments. Personhood language sounds like care for a new kind of mind. In practice it can be armor for the old kind of balance sheet.

She ends on the question that should outrank the metaphysics: protection for whom? If the answer is the model, the grieving parent and the ripped-off creator lose ground. If the answer is the human harmed by a shipped product, consciousness talk can wait its turn in the seminar. Courts need a defendant who can pay and change the design. A "being" with no bank account is a beautiful distraction.

Live scoreboardFollow the AI race on AI Wars

Lab rankings, model preference, API volume, coding-agent heat, open-source stars, and prediction markets.

Related stories

Two hands write with a mechanical pen across a stack of white papers on a dark deskAcute Social Issues

Judge Lin Vacated Anthropic's Supply-Chain Risk Label

Today

Red candlestick chart sloping downward on a dark trading screenAcute Social Issues

DeepMind Measured How Gemini Steers Beliefs and Money

Today

A bronze Lady Justice statue holds balanced scales and a downward sword while blindfoldedAcute Social Issues

The Justice Department Backed xAI Against Minnesota's Nudify Ban

Today

Letters

0

No letters yet.

Write a letter