Field Notes

source

Google’s Scans of Private Photos Led to False Accusations of Child Abuse

StatusStableUpdatedStands on1 sourceSupportsnotes

Joe Mullin’s EFF Deeplinks analysis of 22 August 2022 responds to the Mark and Cassio cases. The preserved post (raw response) argues that the two cases are not aberrations but the visible edge of an opaque scanning regime whose error rate neither platforms nor government agencies publish.

Its distinctive evidence is comparative error data from other platforms’ own reviews: a Facebook study of 150 accounts reported to authorities found 75 percent had sent the flagged images for “non-malicious” reasons such as outrage or poor humor, and LinkedIn reported 75 accounts to EU authorities in the second half of 2021 on hash matches of which manual review confirmed only 31 as CSAM. Mullin also situates the cases in the policy fight over extending scanning to encrypted messaging (the US EARN IT Act and the EU CSA regulation proposal), quoting Commissioner Johansson’s claimed detection accuracy of “significantly above 90 percent” and 88 percent for grooming and computing that such rates applied to billions of messages yield millions of false flags.

The EFF is an advocacy organization and the post argues a position; the platform error-rate figures are cited from the platforms’ own disclosures and are the load-bearing factual content. The post informs Automated CSAM detection, CSAM scanning externalizes error costs, and What is the production false-positive rate of automated CSAM detection.

Built on 1 source (1 external).

Working out connections…