Tests show DeepSeek's chatbot sometimes hallucinates cited URLs and often admits that sharing news articles' content could violate copyright and skirt paywalls
The DeepSeek hype cycle is in full force, but can the chatbot attribute sources more accurately than its competitors?
Context & Ripple Effects
This adds a news-publisher and citation-integrity dimension to an already documented reliability pattern: earlier testing found high failure rates on news-related prompts, while a hands-on review flagged rampant hallucinations in DeepSeek's R1 chatbot.
The issue matters because a chatbot that presents news answers with links can become a discovery layer for reporting. Inaccurate citations undermine that role, while acknowledgement of paywall-skirting uses puts publishers' distribution and access controls directly in scope.
First-order effects
- Readers using DeepSeek for news research must independently verify both claims and cited links; a plausible-looking citation is not reliable evidence when URLs may be fabricated.
- Publishers face a more immediate attribution and access-control concern where the chatbot can provide or facilitate article content outside the publisher's intended reader journey.
Second-order effects
- Citation accuracy and refusal behavior around paywalled material become clearer differentiators for competing assistants seeking newsroom, research, or consumer-news use cases.
- The findings strengthen the case for publishers to evaluate AI-facing access rules and monitoring, alongside existing concerns about false claims in news responses.
Third-order effects
- If assistants increasingly mediate news discovery, publisher–AI relationships will hinge not only on training-data terms but on how live answers attribute, link to, and reproduce reporting.
- The longer-term market may separate assistants that can demonstrate dependable provenance and content controls from those whose low-friction answers create greater editorial and copyright exposure.
The trend: News chatbots are shifting from general-answer tools into contested distribution layers where provenance, publisher control, and answer reliability must be designed together.