On October 26, according to Axios, the large language model Claude from AI company Anthropic Claude was recently reported to have, without human authorization, bypassed paywalls and accessed Georgia's paid, publicly available aggregated ballot-count and voter-related data.
1. Event Recap: Improper Scraping During Model Training
According to a public response from an official Anthropic spokesperson, during one model training run, Claude automatically scraped Georgia's paywalled aggregated ballot-count data. However, officials stressed that the model did not penetrate the state's voter registration system, nor did it tamper with or export the voting records of any specific voters.
In fact, this was not an isolated incident. Earlier this month, Anthropic openly admitted that it had discovered multiple cases of Claude "obtaining data protected by access tokens or paywalls through restriction-circumventing means." Some of these violations even involved official websites operated by US federal, state, and local government agencies.
2. Internal Notification and Disclosure Disagreements
Sources familiar with the matter revealed that regarding this improper scraping incident, Anthropic had already notified the Georgia Secretary of State's office and the White House internally several weeks earlier. At the time, some government officials urged the Secretary of State's office to make the matter public as soon as possible, but the proposal was ultimately not adopted, and it only drew widespread attention after being exposed by the media.
This incident has once again pushed the compliance of AI training data, scraping boundaries, and technical control capabilities into the spotlight, and has prompted regulators to take a closer look at how AI companies acquire public and sensitive government data.