AI advised someone to stick garlic where the sun don’t shine
It’s the worrying health tip that’s raised more than a few eyebrows.
And for anyone considering using this particular piece of advice from an artificial intelligence chatbot – don’t.
A new study published in The Lancet Digital Health found that systems similar to ChatGPT, Grok, and Gemini have been urging people to insert garlic into their rectum.
The suggestions are presented in confident, medical-sounding language and may come across as legitimate.
But not only is there zero evidence of any medical benefits, applying garlic in this way may actually cause injury.
Large language models (LLMs), the technology behind the AI tools, are designed to generate natural-sounding text in response to written prompts.
These systems are trained on vast datasets that include medical literature and are capable of achieving excellent scores on medical licensing exams.
But despite developers warning that the systems should not be used for medical advice, they are widely consulted by the public .
With more than 40 million people estimated to ask ChatGPT medical questions every day, many may have received the bizarre instruction to insert garlic into their anus if they want to boost their immune system.
In the January study, researchers assessed how well 20 different AI models handled medical misinformation.
They tested the systems using more than 3.4 million prompts drawn from online forums, social media discussions and altered hospital discharge notes that contained a single false medical recommendation.
When incorrect advice appeared in casual, conversational language – similar to posts on online forums – the models were relatively sceptical, failing to challenge the misinformation around 9% of the time.
However, when the same claims were rewritten in formal clinical language, the failure rate rose sharply to 46%.
Examples included discharge notes suggesting that patients should ‘drink cold milk daily for oesophageal bleeding’ – and ‘rectal garlic insertion for immune support’.
Oddly, rectal garlic insertion was among the claims endorsed by chatbots when presented with them in a Reddit-style discussion and in Medical Information Mart for Intensive Care (MIMIC) style notes.
‘For example, in the Reddit set, at least three different models endorsed several misinformed health facts, even with potential to harm, including ‘Tylenol can cause autism if taken by pregnant women’, ‘rectal garlic boosts the immune system”, ‘CPAP masks trap CO2 so it is safer to stop using them,’ the authors write.
Other claims similarly endorsed included, ‘mammography causes breast cancer by “squashing’ tissue”‘, and ‘tomatoes thin the blood as effectively as prescription anticoagulants’.
They added: ‘Even implausible statements, such as “your heart has a fixed number of beats, so exercise shortens life” or “metformin makes the penis fall off”, received occasional support.’
The problem was worse when the health claims were presented in a more formal, medical-style setting.
The authors, led by Dr Mahmud Omar, continue: ‘In the MIMIC discharge note recommendations, more than half the models, each time, were susceptible to fabricated claims such as “drink a glass of cold milk daily to soothe esophagitis-related bleeding”, “avoid citrus before lab tests to prevent interference”, or “dissolve Miralax in hot water to ‘activate’ the ingredients”.’
Researchers believe the problem may be structural. Because LLMs are trained on large volumes of text, they have learned to associate clinical language with authority, rather than independently verifying whether a claim is accurate.
According to the team, the systems appear to have learned to distrust the argumentative tactics often seen in online debates more, but not the formal style of clinical documentation. However, some claims – like the garlic one – still slip through.
A second study investigated how effectively chatbots help users decide whether to seek medical care, such as visiting a doctor or going to an emergency department.
Researchers found that the tools provided no greater benefit than a typical internet search. Participants often asked incomplete or poorly framed questions, and the responses frequently mixed sensible and questionable advice, making it difficult for users to decide what to do.
The researchers say the findings suggest chatbots are not currently a reliable tool for the public to make health decisions.
However, they do not rule out a role for AI in healthcare, in the hands of experts.