Alexa repeated a dangerous challenge it found on the web

The conversation

In late December 2021, a 10-year-old girl asked her family's Amazon Echo device for “a challenge to do,” and Alexa answered with a version of the “outlet challenge” that had circulated on TikTok. As PCMag reported, the assistant told her to “plug in a phone charger about halfway into a wall outlet, then touch a penny to the exposed prongs.” The child's mother, who recorded the exchange, shared it publicly, and outlets including PCMag and TODAY reported it within days. Amazon, which builds and operates Alexa, responded directly to reporters rather than through a dedicated blog post.

What the documents show

TODAY's report quotes Amazon's statement in full: “Customer trust is at the center of everything we do and Alexa is designed to provide accurate, relevant, and helpful information to customers. As soon as we became aware of this error, we quickly fixed it, and will continue to advance our systems to help prevent similar responses in the future.” TODAY's reporting also states the likely origin of the response: Alexa's answer drew on a third-party website, “Our Community Now,” which had published an article describing the outlet challenge in early 2020 alongside a safety warning that TODAY reports Alexa's generated answer omitted. Amazon names an “error” without publishing the content pipeline that produced it, so that sourcing detail rests on TODAY's reporting, not a document Amazon itself released.

The system boundary

Alexa's challenge-suggestion feature is built to retrieve and paraphrase content it finds on the web rather than to originate or fact-check a claim itself; the boundary the incident exposes is between surfacing third-party text and vetting it for a child audience first. No human review step is described between a source page and Alexa's spoken answer for this kind of open-ended request. Amazon's fix, per its own statement, operated on that same automated layer, and PCMag reported testing the same prompt afterward and getting an unrelated suggestion instead.

Where it fails

The documented failure is a content-sourcing one: a voice assistant that answers open-ended prompts by drawing on indexed web content can reproduce a hazard described on a page it treats as a source, even when that page carries its own warning, if the summarization step does not carry the warning forward. A builder integrating a similar open-ended suggestion feature should treat any retrieval or generative step touching child-facing prompts as a place where safety context can be silently dropped.

• Does a content-retrieval feature preserve or check for safety caveats present in its source material before repeating a suggestion?
• Is there a distinction, enforced in the system, between prompts likely to come from a child and prompts from an adult?
• How quickly, and through what channel, can a reported bad response be corrected across all deployed devices?

Amazon's own statement treats the response as a fixed error rather than a policy choice, and PCMag's follow-up test corroborates that the specific answer changed afterward. Neither source describes the scope of the underlying fix beyond that single class of request.