Claude Cowork Cites Wrong Sources

This is the core failure mode of AI for outward facing marketing, it can produce copy that sounds polished while quietly grounding itself in the wrong source. In this workflow, the model was not making up language from nowhere. It was importing a viewpoint from the open web that conflicted with the company’s product and positioning, which is exactly why human review matters most on anything customer facing.

• The interview shows a pattern, not a one off. The same operator also caught Cowork using competitor landing pages as cited evidence in competitive analysis, then had to force the workflow to check internal product docs first. The issue is source ranking, not just wording.
  1 sacra
• Anthropic’s own guidance says Claude can be incorrect or misleading and should not be treated as a single source of truth, especially when using web results. That fits this example closely, because a plausible social post can still be strategically wrong if the evidence chain is weak.
  2 sacra 3 anthropic
• The product setup makes this risk easy to understand. Claude can search the web and connect to tools like Google Drive and Slack, which is powerful, but also means one workflow can mix canonical company docs with noisy public internet material unless the prompt and source order are tightly controlled.
  1 sacra 3 anthropic 4 claude

The next step in this market is not more autonomous posting, it is better source control and approval workflows. The winning agent products will treat internal docs as the first stop for product claims, flag weak sources before draft generation, and require a human signoff before anything reaches customers.