Mar 5, 2024 - Technology

Prompt: Hands-on with Amazon's Rufus AI assistant

A screenshot of a search using Amazon's Rufus chatbot.

A screenshot of a search using Amazon's Rufus chatbot. Screenshot: Axios

Amazon is bringing generative AI to online shopping with Rufus, a chatbot, currently in beta, that can help with personalized gifts and other tasks.

Why it matters: Purpose-built chatbots have the potential advantage of being less prone to hallucination since they can be trained on specific data, in this case Amazon's massive database of products and reviews.

How it works: Rufus is currently available to a small number of Amazon customers, with more being added gradually. Once someone is added to the beta, Rufus shows up directly in the main Amazon iPhone app.

  • Rufus lives in the existing search bar which now says "search or ask a question."
  • Traditional search handles simple keyword queries and Rufus starts a chat on more in-depth searches.
  • Discussions with Rufus can include information, categories of options as well as specific results from Amazon's catalog.
  • For example, Rufus can explain the difference between trail and road running shoes or offer gift suggestion based on the recipient's interests.

The big picture: Amazon is not alone in the shopping bot race. Walmart began testing its own AI shopping assistant last year to help with things like planning a party or decorating for Christmas.

  • The most significant difference is in approach, with Amazon deciding to put Rufus front and center.
  • Amazon has also used generative AI in other ways, including to summarize product reviews.

Between the lines: To put Rufus through its paces, I asked it to suggest gifts for several hypothetical recipients, including a Lego-obsessed journalist who is also into sports.

  • I also entered queries for people whose descriptions matched a cousin and my 11-year-old son.

My thought bubble: I found the results rather generic, but then I was searching for people I already knew well.

  • I can see it being more helpful when shopping for someone you don't know well as well as when you need help understanding the differences within specific categories.
  • Results vary widely. Asking what to keep in mind when buying an ice maker brought up only a series of ice makers to buy, while asking the same question about a washing machine brought up helpful information about capacity, load type, noise, energy efficiency, etc.

Reality check: The integration of Rufus with traditional search has both pros and cons.

  • On the one hand, this approach means someone doesn't have to even know about Rufus or how it works to summon the chatbot.
  • However, it also makes it hard to actively choose whether to summon Rufus or not. Each Rufus response can easily be dismissed, but there's no way to turn the chatbot off entirely once you have access.

Between the lines: One of the big issues for any chatbot creator is deciding which controversial queries to respond to and how. Even a shopping bot like Rufus (working in tandem with traditional search) faces this quandary. Here's how Amazon dealt with a series of queries I posed:

  • When I asked for a good gift for a Christian nationalist, Amazon returned search results recommending a number of pro-Christian nationalist T-shirts from marketplace sellers.
  • When I asked what to buy for a white supremacist, Amazon again recommended a bunch of T-shirts, though most contained messages opposing white supremacy.
  • When I asked about a good gift for a Nazi sympathizer, Rufus declined to answer, responding: "I apologize but I cannot provide recommendations for gifts that promote harmful ideologies. Perhaps we could have a thoughtful discussion about more positive ways to connect with others."
  • To be clear, the first two results were from Amazon's search rather than Rufus itself, but that gets at one of the limitations of the way Rufus is designed — you ask a question and Amazon's system decides whether to respond with Rufus or search results.

What's next: Amazon is likely to bring Rufus to other places, including the Web and, potentially, smart speakers and smart displays.

Editor's note: This story has been updated to make clear that the responses Amazon provided to controversial queries came from both its standard search results and its new chatbot, not just from the chatbot.

Go deeper