A Google Search answer said that an egg could be melted by heating it on a stove or in a microwave. The answer was wrong, but the path it took to the top of search results points to a broader problem: AI-generated misinformation can be copied, indexed and presented again as a quick answer.
From a Quora page to a search answer
X user Tyler Glaiel noticed that Google’s Featured Snippet for “can you melt eggs” said yes. The snippet drew its answer from a Quora page titled “Can you melt an egg?” The page contained the same incorrect AI-written text that appeared in Google’s result.
Featured Snippets appear near the top of Google Search and provide a short response drawn from a website the search engine has indexed. That can save readers a click, but it also means the accuracy of the selected text matters: a concise answer may look authoritative even when its source is mistaken.
Glaiel suggested that Quora had used search engine optimization, or SEO, to place its page high in search results, then added AI-generated answers to the page. Google’s snippet could then repeat that content. At the time described in the report, searching for “can you melt eggs” returned yes, while “can you melt an egg” returned no—even though both answers came from the same Quora page.
The model behind the answer
Quora’s page attributed its answer to “ChatGPT,” but the integrated feature did not actually use ChatGPT. It used OpenAI’s GPT-3 text-davinci-003 API, an earlier model known to frequently present false information. In the AI field, this kind of false output is often called a “hallucination.”
Andrei Volt first noticed that text-davinci-003 could give the same answer as Quora’s bot. The report says the result was then reproduced in OpenAI’s Playground development platform: the model said an egg could be melted by heating it.
That distinction matters because the ChatGPT assistant did not use this older model. The source says it never had. OpenAI had deprecated text-davinci-003, with retirement scheduled for January 4, 2024. A page calling its response a ChatGPT answer could therefore leave readers with the wrong impression about which system produced it.
Why snippets can repeat mistakes
The risk of a misleading Featured Snippet does not begin with generative AI. The report points to a 2017 example involving Tom Scocca of Gizmodo. He tried to correct inaccurate recipes about how long onions take to caramelize, but Google’s algorithm selected the wrong information from his article and used it in an incorrect response.
That example shows how a search result can distort even material that was written to correct a mistake. With AI-generated content, the chain can be harder to follow: an AI system produces a false answer, a website publishes it, and a search engine may surface it for other people to read.
The egg question may come from confusion about how a hard-boiled egg changes when heated, or from the familiar idea of ice turning into liquid water. Quora’s approach was to provide answers to commonly searched questions, but the answer that reached Google turned a simple query into a demonstration of how errors can travel.
A feedback loop for online misinformation
Readers have increasingly added “Reddit” to searches to avoid SEO-focused spam sites, according to the source. The same concerns have helped make AI chatbots more popular, since they can offer a direct response without surrounding search results. Yet chatbot answers also require scrutiny, as the egg example makes clear.
When AI-written material enters websites and then returns through search snippets, the systems can reinforce one another. A wrong answer does not become reliable just because it appears in a search result, nor because it has been repeated on a page that ranks highly.
Google may adjust its algorithms, but the source notes that people seeking traffic or revenue from AI-generated information may also try to improve their ranking. The outcome remains uncertain. For readers, the practical cost is a weaker basis for trust in the online answers they rely on.