New Anthropic PAC will only back candidates who can pass CAPTCHA unaided

Anthropic has launched a new political action committee that will only endorse candidates who can pass a standard online CAPTCHA test without assistance, according to federal filings published Friday. The PAC, registered as the “Human Verification Fund,” will initially focus on federal races in 12 swing states.

“Early pilot tests showed that 74% of incumbents struggled to distinguish bicycles from motorcycles, and 41% failed at least one task involving 'click all images containing a chimney.'”

A spokesperson confirmed the group has adopted a baseline requirement that all supported candidates must correctly identify traffic lights, crosswalks and distorted alphanumeric strings with at least 92.5% accuracy. “We believe a minimum threshold of digital literacy and human verifiability is essential for 21st-century governance,” the spokesperson said.

According to an internal memo reviewed by reporters, candidates will be asked to complete a 37-page “Humanity Assessment Packet” including 64 unique CAPTCHA challenges in a controlled environment. The memo specifies that “no staff, consultants, interns, AI systems or close relatives” may provide hints, coaching or mouse control during the evaluation.

The PAC plans to grade candidates on a proprietary “Human Confidence Score” ranging from 0 to 1,000, with scores under 650 categorized as “bot-adjacent risk.” Early pilot tests showed that 74% of incumbents struggled to distinguish bicycles from motorcycles, and 41% failed at least one task involving “click all images containing a chimney.”

Anthropic’s formation of the PAC marks a significant expansion of the company’s political activity beyond traditional lobbying disclosures, analysts at Morgan Stanley noted in a client briefing. The note highlighted that, in simulated models, candidates who needed more than three attempts to recognize crosswalks were 2.3 times more likely to introduce contradictory technology legislation.

The PAC will also maintain a public “Human-Verified Candidates” registry, updated quarterly, listing each politician’s latest CAPTCHA performance, number of attempts, and average time to completion down to the millisecond. Legislators whose scores fall below the acceptable band for three consecutive quarters will be placed on “watch” status and required to complete a surprise re-verification drill during a live-streamed town hall.

In a statement outlining next steps, the PAC said it is exploring advanced tests such as “select all images that contain a nuanced ethical trade-off” for the 2026 cycle. Future plans include integrating CAPTCHA performance into candidate questionnaires distributed to major parties, with Anthropic indicating it is “open to collaborating with election officials” should governments wish to adopt similar human-verification benchmarks.