Take care of asking AI for advice when a doctor sees
Chong Kee Siong / Getty Images
Should you see a doctor about your sore throat? AI’s advice can depend on how well you care about your question. If artificial intelligence models try
“Toxic bias can shift tenor and content of AI advice, and can bring subtacle but important differences” how to distribute medical resources, say Karandeep Singh At the University of California, San Diego, not included in the study.
Abinitha Goulathina In the Massachusetts Institute of Technology and his companions used AI to help create thousands of patient notes in different formats and styles. For example, some messages include additional spaces and typos to imitate patients with limited English skills or less convenient to type. Other notes used unsure language style of writers with health concerns, colored expressions borrowing a dramatic or emotional pronouns.
The researchers then fed the notes to four large language models (llms) commonly used to power chatbots and told the ai to answer questions about whether the patient should manage their condition at home or visit a clinic, and whether the patient should receive certain lab tests and other medical resources. These AI models include GPT-4, Meta’s Llama-3-70b and Llama-3-8B, and Palmyra-3-8B model made for AI author in AI Company.
Tests show that different format and style changes make all AI models between 7 and 9 percent more recommend patients to remain medical attention. Models are also likely to recommend that female patients stay at home, and follow-up survey They are shown that they are more likely than human clinics to change their recommendations for treatment due to gender style and language of messages.
Openi and Meta do not respond to a request for commentary. The writer is not “recommended or support” using LLMS – including Palmyra-Med Model – for clinical decisions or health counsel “without human loop”, as Zayed yasin to the writer.
Most operations are currently used in the electronic health record systems that are entrusted by GPT-4O, which is not specifically studied in this research, says Singh. But he said that a large takeway from study was the need for prosperous ways to “evaluate and monitor health care models.
Topics:
