CNN filed suit against Perplexity in Manhattan federal court on May 28, alleging the AI search startup scraped over 17,000 pieces of content—articles, photos, and video—without license. The complaint centers on how Perplexity’s retrieval-augmented generation (RAG) pipeline ingests publisher content to produce answers that, CNN argues, substitute for visiting its site.

Perplexity’s architecture is worth examining. The system runs real-time web searches, extracts relevant passages, and feeds them into a large language model to generate responses. This RAG pattern keeps outputs current and source-grounded, but creates a technical dependency on third-party content. CNN claims Perplexity’s crawlers bypassed blocked endpoints after licensing talks collapsed last year, copying material behind paywalls and reproducing verbatim passages.

The engineering trade-off here is familiar: RAG improves answer freshness and reduces hallucination risk, but it also makes the system a direct consumer of publisher infrastructure. Perplexity’s “skip the links” feature—which minimizes click-throughs to original sources—amplifies the tension. From a system design perspective, the question is whether caching or summarizing constitutes fair use or infringement. The suit alleges Perplexity stored and reproduced protected expression, not just facts.

CNN had pursued a licensing deal before blocking Perplexity’s crawler. When negotiations stalled over display limits, the network implemented robots.txt restrictions and technical blocks. The complaint states Perplexity continued scraping anyway, raising operational questions about crawler compliance and rate limiting.

This case joins over 100 similar actions since 2023. No appellate court has ruled definitively on whether training models on copyrighted articles constitutes infringement. Lower courts have split: some allow claims over outputs that closely mimic protected works; others are skeptical of claims based solely on data ingestion. For engineers, the outcome will shape how we design crawlers, cache layers, and attribution systems.

Perplexity’s defense—“you can’t copyright facts”—holds for factual data but not for original expression. The technical challenge for courts will be distinguishing between summarizing facts and reproducing protected structure. Discovery will likely expose Perplexity’s internal data flows, caching policies, and model training pipelines. Until then, the engineering community watches a case that could redefine how RAG systems interact with publisher content.

Source: Webpronews