If an AI assistant writes false answers itself, can the platform still hide behind 'we just host what others post'?

This explores whether the protections platforms get for hosting other people's content still apply when the platform's own AI generates the words. The corpus has no dedicated analysis of the Digital Services Act. It does have one sharp data point that goes to the heart of the question: a Munich Regional Court ruled that Google is directly liable for false statements in its AI Overviews Can courts hold Google liable for AI Overview false claims?. The court treated those summaries as Google's own content rather than as neutral search results pointing to someone else's page, and said Google has to check what its algorithms produce. That is the key distinction. Host protections exist for intermediaries that pass along what others wrote. Once the system writes the answer itself, it starts to look like a publisher.

Other material in the corpus shows why courts may find that framing persuasive. A randomized field experiment found that AI Overviews cut clicks to publisher sites by 39.8 percent without making users any happier with their results Do AI Overviews actually reduce clicks to publisher sites?. In practice, the AI summary replaces the sources instead of leading people to them. That weakens the claim that the platform is only a conduit. The stakes rise further because people rarely check fluent AI output. One note calls this 'cognitive surrender', with studies finding about 80 percent of AI suggestions adopted without challenge When do users stop checking whether AI output is actually backed?. When nobody checks downstream, responsibility shifts toward whoever produced the claim.

There is a second case the question may be pointing at: AI-generated content that users upload to a platform. That is a different situation. A creator who posts AI-made images is still a third party, so the platform's role as host is unchanged. Even so, platforms are already treating that content differently. One study of 178,854 matched pairs found that a platform's recommendation algorithm gave AI-generated posts less exposure than human-made ones Can algorithmic distribution prevent AI content from overwhelming creator diversity?. Choosing what to amplify is an editorial act, and a platform that does it is harder to describe as passive. It also matters that people can't reliably tell AI content from human content: across 30 studies, detection accuracy hovers around chance Can people reliably spot content made by AI?. Separately, telling audiences that AI was involved leaves 34–62 percent of them still persuaded Does telling people an AI wrote something actually stop them from believing it?. Together, these weaken the idea that labeling AI content is enough to manage it.

One more twist: the line between 'our output' and 'someone else's content' can be blurred on purpose. Advertisement embedding attacks slip promotional or malicious material into a model's answers through hijacked distribution channels or tampered model checkpoints Can language models be hijacked to embed hidden advertisements?. If a third party poisons the model and the platform's AI repeats it, it is unclear whether that counts as hosted content or the platform's own speech. The corpus doesn't settle this, but it is likely where the next round of liability disputes will happen.

The short answer from this collection: at least one European court has said no. When the platform's AI writes the answer, the platform owns it. For the specific wording of the DSA itself, you would need sources outside this library.