Google’s AI search feature is being tested in a very public way: by ordinary users turning its mistakes into memes. The most visible failures are funny at first glance, but the pattern behind them is more serious.
As AI search becomes more common, screenshots of bad answers are becoming a feedback loop for the companies building these systems. Some examples involve absurd queries. Others touch on health and safety, where a wrong answer can carry real consequences.
How users became public red teams
In cybersecurity, a red team tries to break into a product before attackers do. The goal is to find weak points while there is still time to fix them. Tech companies apply versions of that idea before launching new products, especially when the product will reach a large audience.
Google Search is estimated to process trillions of queries per day, so any AI feature added to it needs to handle an enormous range of questions. Yet social media has shown that even heavily tested AI systems can still fail in obvious ways once they meet the open web.
One viral example involved Google’s AI feature describing “Running with scissors” as a cardio exercise. The answer said it could increase heart rate and require concentration and focus, and added that “Some say it can also improve your pores and give you strength.”
The problem was not only that the answer was wrong. It came from Little Old Lady Comedy, a comedy blog. The AI treated a joke as useful information, and users quickly spread the result as proof that the product could miss basic context.
Why the mistakes spread so fast
AI failures have become a recurring internet spectacle because they are easy to understand. A search product gives a bizarre answer, someone posts a screenshot, and the mistake becomes a public demonstration of what the system failed to grasp.
The source article points to several examples beyond Google Search: bad spelling on ChatGPT, video generators failing to understand how humans eat spaghetti, and Grok AI news summaries on X that did not understand satire. The common thread is not one company or one feature. It is the difficulty of making AI systems reliably interpret messy human content.
Google told TechCrunch in an emailed statement: “The examples we’ve seen are generally very uncommon queries, and aren’t representative of most people’s experiences.” The company also said it had conducted extensive testing before launch and would use isolated examples to refine its systems overall.
That response reflects a tension around AI products. A mistake may be uncommon across all searches, but once it is visible, searchable and shareable, it can shape public trust far beyond the original query.
Satire, Reddit and the data problem
Another widely shared example involved pizza. Google suggested that if cheese would not stick, a user could add about an eighth of a cup of glue to the sauce to “give it more tackiness.” The source of that answer was an eleven-year-old Reddit comment from a user named “f––smith.”
That mistake matters because tech companies are paying for the kinds of data that can contain jokes, sarcasm and bad advice. Google has a $60 million contract with Reddit to license its content for AI model training. Reddit also signed a similar deal with OpenAI, and Automattic properties WordPress.org and Tumblr are rumored to be in talks to sell data to Midjourney and OpenAI.
None of that means every piece of licensed data is low quality. But the examples show the core risk: AI systems trained on or grounded in internet content must distinguish between helpful information, satire, old jokes and deliberate nonsense.
When they do not, the result can be ridiculous. It can also become part of the next wave of web content. A bad AI answer generates articles and posts about the mistake, and those new pages can later appear in related search results.
The safety stakes are not always funny
Some viral examples are easy to laugh at because the query itself appears designed to provoke a bad answer. The source article notes that it is hard to imagine someone seriously searching for “health benefits of running with scissors.”
Other cases are different. Science journalist Erin Ross posted on X that Google returned incorrect guidance about what to do after a rattlesnake bite. Her post, which got over 13,000 likes, showed AI recommending applying a tourniquet, cutting the wound and sucking out the venom.
According to the U.S. Forest Service, those are all things a person should not do after being bitten. In another case, author T Kingfisher amplified a Bluesky post showing Google’s Gemini misidentifying a poisonous mushroom as a common white button mushroom. Screenshots of that post then spread to other platforms as a warning.
These examples change the conversation. A strange sports answer or a glue-on-pizza suggestion exposes a reliability problem. A wrong answer about snakebites or mushrooms raises a safety problem.
A feedback loop AI companies cannot ignore
The source article also describes a strange loop around AI mistakes. New York Times reporter Aric Toler posted a screenshot on X of a query asking whether a dog had ever played in the NHL. The AI answered yes and, for some reason, called Calgary Flames player Martin Pospisil a dog.
After that mistake spread, the same query began pulling up an article from the Daily Dot about Google’s AI thinking dogs are playing sports. In other words, public coverage of an AI error can become part of the information environment the AI later surfaces.
That is the deeper issue behind the memes. Large-scale AI models trained on the internet must work with a web full of jokes, errors and misleading material. When companies ship systems that summarize that material too confidently, users will keep finding the gaps.
The old phrase still applies: garbage in, garbage out. What is different now is that the garbage can become viral, searchable and visible to everyone.