If an AI writes its own summary instead of linking to the source, does the legal reasoning that protects search engines still apply?

This explores whether the German Federal Court of Justice's search-engine rulings, which let search engines show snippets and thumbnails of other people's content, carry over to AI-generated summaries. The corpus has no legal analysis, so it can't settle the doctrinal question. It contains no case law, copyright scholarship or commentary on the BGH. What it can do is test the factual assumption those rulings depend on. In the thumbnail cases (Vorschaubilder), the court reasoned that someone who puts content online without blocking crawlers implicitly agrees to it being indexed and previewed. That consent makes sense because a preview works as a signpost back to the original. The useful question is whether a generative summary still works as a signpost or has become a substitute.

The evidence here leans toward substitute. In seven randomized experiments with more than 10,000 participants, people who learned a topic through ChatGPT summaries felt less ownership of what they learned and gave advice that independent raters judged sparser than advice from people who used web search Does learning from AI summaries produce shallower knowledge than web search?. A snippet makes you click through to get the substance. A summary hands you a condensed version of the substance and removes the reason to go further. That loss of depth is evidence that the summary is doing the source's job rather than pointing to it, and the implied-consent logic only works if the preview points somewhere.

The picture isn't one-sided. Nielsen Norman Group found that people still ran traditional search alongside AI chat and seldom treated chat as a full replacement Does generative AI chat actually replace traditional search?. That supports a narrower claim: some generative answers work as a first pass that still sends people on to sources, especially when users are uncertain. The legal answer may therefore depend on how a given system is built, not on the technology as a category.

The retrieval research shows which design choices would make a summary behave more like a search result. Some systems refuse to answer unless the claim is grounded in retrieved evidence Can RAG systems refuse to answer without reliable evidence?. Others enforce source-attribution checks before generated text is reused Can RAG systems safely learn from their own generated answers?. Both tie the output back to identifiable originals, which is the property the search-engine rulings relied on. A summary that cites and links prominently is closer to a thumbnail than a fluent answer with buried or missing sources. If you want to take this further, the question to bring to legal sources is less "is it a search engine?" and more "does the output still send readers to the source, or keep them away from it?"