AIWI - The AI Whistleblower Initiative

We're hiring! AIWI supports frontier AI insiders who want to raise concerns. Funded through 2029, we're growing from 4 to ~15 across SF, DC & London.

5 roles: growth, guidance & preparedness, engineering, research, policy.

Interested to apply? Visit join.com/companies/aiwi…

#hiringnow #AISafety #AIGovernance #whistleblowing #AIPolicy

Most AI lab insiders reach out for support way too late.

Our founder Karl Koch with Kay Kozaronek on WHAI: get advice early, document as things happen, and know that talking to a lawyer is privileged. youtu.be/ZOeoVIwQ51w?si…

Laws only protect AI whistleblowers if they're written well.

We helped shape the world's first AI whistleblowing channel. Now we're hiring a DC AI Whistleblowing Policy Lead to strengthen US protections.

📍 Washington, DC · Apply by 8 Nov

join.com/companies/aiwi…

AIWI is hiring - come work with us!

How can insiders spot motivated reasoning at work? When should they consider leaving? AIWI is hiring a Research Manager to turn questions like these into guidance for AI insiders.

Closing soon - apply by 1 Nov:
join.com/companies/aiwi…

On the day of @Anthropic's S-1, some (or a few more) thoughts on their draft federal AI bill from a whistleblowing policy perspective:

While it has drawn criticism, its whistleblower section reads as one of the better parts, and we'd like to see (substantial parts of it) carried into whichever federal AI bill passes!

Checked against our AI Whistleblowing Law Best Practice Guide and our feedback on @ChuckGrassley's AI Whistleblower Protection Act (AWPA), both in thread, it gets a lot right:

Anthropic sets precedent in U.S. AI whistleblowing policy in two places:

• It protects staff at independent evaluators, not just lab employees. This is significant! So far we have not, to our knowledge, seen any AI company openly call for protections of auditors/ evaluators. There is an asterisk here (see below), but this is a good step and any IVO bill we end up seeing should inherit this.

• Companies that retaliate face fines that can run to a share of their global revenue -- this is also a first and something we called for! (If penalties are going to be capped at the larger of USD 1/5m, as proposed in the draft, or x% then of course the x% are irrelevant though)

It also follows best practice in other places:
• Courts can protect a whistleblower before the case is decided via injunctive relief.
• Companies can't use NDAs or forced arbitration to stop staff from reporting.
• A fair burden of proof and double back pay*.
• Broad channel coverage: Enforcement Agency (Commerce), Congress, "or another
Federal or State regulatory or law enforcement agency" or reporting to a supervisor
• Labs must run an anonymous internal reporting channel, with regular updates to the person who reported and to the board.
• For avoidance of doubt: State whistleblower laws stay in place

Having said that: Looking closer, the bill is weaker than it reads in two ways:
1. Some gains for whistleblowers are weakened by other parts of the larger bill
2. It's lagging current U.S. precedent in multiple places

Starting with (1):
Evaluators' retaliation protections are dramatically weakened as companies can pick, pay, and switch evaluators ~at will. Commerce can't block or assign them. This will enable retaliation.
Security auditors, the other type of outside checks in the bill, get no whistleblower protection at all.
The lab still can't gag evaluator staff or censor their reports (good), but that's still a way from protection.

Next, we also understand the bill as Commerce having to inform a company if it receives incident reports. That means Commerce may have to 'out' whistleblowers that claim non-incident-reporting (not by name, but the fact of a report may be a hint). This would be alleviated if there were confidentiality obligations for the Commerce whistleblowing channel, however...
...Usually, when enforcement agencies are prescribed as whistleblowing channels, rules are set in place for them. This bill does not do that. Follow the SEC: Mandate Commerce to maintain confidentiality on whistleblowing reports, anonymous reporting, skilled and dedicated staff. Ideally follow the EU as well and set response deadlines (as company internal).

Now to (2), where is the bill is behind current U.S. precedent:
Main Gap: Many safety concerns aren't covered and the threshold for risk reporting is unnecessarily high.
Staff are only protected when they report a breach of this bill, or a danger on the bill's own short list (more than 50 people killed or seriously hurt, or over $1bn in damage, from a set list of causes). The AWPA protects reports of any (no minimum number of deaths!) substantial and specific danger to public health and safety (this has precedent in many whistleblower laws), to national security, and reports of security vulnerabilities (what did we see during the past days?).

It also protects reports of AI-related breaches of any federal law.

Insiders shouldn't have to work out whether a concern could cost $1bn before they speak up.

Other gaps:
• No protection for going public, even when the danger is imminent
• No clear protection for stock and equity, which make up much of AI pay, or for staff on visas. Likely implied via "double back pay", but good to make explicit..
•*The burdens of proof currently are cross-referenced to those in AIR 21, 49 USC 4121. It should be 41 USC 4712(c)(6).
Anthropic calls its strategy a "race to the top" and has said it will propose rules to stop labs shopping for easy evaluators. That revision is the moment to fix this.

Thank you to @GovAcctProj's Tom Devine for the contribution to this piece!