Breaking

Superhuman CEO Says Frontier AI Labs Should Not Be Sole Authors of Their Safety Systems

September 14, 2026 at 19:01 EDT

Policy & Regulation
Foundation Models

Superhuman chief executive Shishir Mehrotra said frontier AI laboratories should not be solely responsible for building their own safety systems, arguing that the industry should not rely on the toolmaker to invent every safeguard. Speaking as debate intensified over who should test and police the most capable models, he asked whether anyone would “trust the toolmaker to come up with all of their own safety.”

Sep 14, 2026 · Superhuman

Do not trust the toolmaker to write all of AI safety

Superhuman CEO Shishir Mehrotra said frontier labs should not be solely responsible for their own safety systems, as Anthropic, OpenAI, and Google discuss an industry-led standards body.

Oversight layers

Self-certification is one party. The independent model stacks the lab, an outside auditor, and public rules.

Lab self-check

Lab + IVO + public rules

Column height matches the number of parties that can say a system is safe.

Labs in standards talks

Steps in Amodei’s pacing plan

2028

First California IVOs due Jan 1

How self-policing works today

1. Train

Frontier lab builds the model

→

2. Self-test

Same lab writes the safety checks

→

3. Disclose

Public sees what the lab chooses to share

Why outsiders want in

Independent evaluators with employee-like access, and licensed IVOs, would let someone other than the builder say whether safeguards hold.

The capture risk

If a few labs also write the standards, testing can become gatekeeping. Critics call an antitrust waiver for pace-setting a cartel in safety clothing.

The claim

You do not trust the toolmaker to come up with all of their own safety.

Shishir Mehrotra, CEO of Superhuman — the fight is no longer whether AI needs guardrails, but who is allowed to write them.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.