Ro Khanna wants to ban self-improving AI until it has been authorised by a federal agency

The ‘Human Control Over AI Act’ proposed by the Californian Democrat would establish a dedicated agency to oversee frontier laboratories, with licensing requirements, on-site auditors, emergency shut-off switches and criminal penalties for employees.

Democratic Representative Ro Khanna, elected from Silicon Valley, announced on 28 September a bill named Human Control Over AI Act. Its most radical point is to prohibit models capable of recursive improvement or of modifying their own objectives, containment or arrest mechanisms, until federal safeguards exist and an agency has given its green light.

An agency, licensing and criminal law

The text creates a model authority for OpenAI, Anthropic, Google DeepMind and xAI: safety rules, licences for training and deployment, audits, standards for Internet-insulated test environments and emergency switches, and advanced chip control. Each border laboratory would host independent auditors reporting directly to the Agency. On the sanctions side, the project provides for a criminalization against humanity for a deployment that has destroyed civilian populations, prosecution of employees who deactivate security or launch an unauthorized system, and mandatory liability insurance prior to placing on the market.

Khanna claims to have drawn inspiration from METR, MIRI and Palisade Research rather than industry leaders. The text is in addition to Jay Obernolte and Lori Trahan's bipartisan FRONTIER Act, but no House drafts are expected to vote before the mid-term elections.