Customers cannot solely add paperwork to Claude, however can even drop hyperlinks into the chat and get it to summarize the contents. Once you add hyperlinks, Anthropic warns that Claude would possibly hallucinate content material when hyperlinks are enter into the chat, which I discovered to be true a couple of third of the time.
The instance beneath reveals how Claude “summarized” a Wall Road Journal article after I gave it simply the hyperlink on the left, and after I gave it the textual content I accessed with my subscription on the appropriate. This activity resulted in two putting points: main hallucinations, and the potential for accessing paywalled content material.
The article discusses controversies involving restaurant tipping in Chicago, the place new laws are being thought-about. Claude’s main hallucination concerned saying that the difficulty was being debated in New York Metropolis as an alternative of Chicago, which adjustments the primary topic of the article.
Additionally: How does ChatGPT truly work?
The truth that it looks like it might probably entry content material by hyperlinks brings up the query of whether or not or not Claude can bypass paywalls, a attribute that might infringe copyright legal guidelines. I did give Claude paywalled hyperlinks to check its capabilities. Claude gave a abstract that was just like the reply it gave utilizing the copied textual content, albeit with hallucinations.
The underside line is that I would not belief the summarization Claude makes by hyperlinks, because the outcomes might be wildly inaccurate. Nevertheless, summarization is a serious function that Anthropic might enhance and construct upon, so long as restrictions to respect paywalled content material are in place.