Ever since artificial intelligence became a part of our lives, we have fallen into the habit of quickly firing up that chat box and typing our biggest worries, darkest thoughts among many other things. But it has become important to stop and ponder whether AI is really producing the most useful and viable solutions for us. A recent study published in NEJM AI suggests that artificial intelligence systems may recommend more expensive medical care than doctors. The findings said that AI-generated care plans may cost more than three times as much as doctors’ recommendations.The researchers compared recommendations from 24 artificial intelligence systems with the recommendations by physicians across 200 outpatient and emergency department visits. They found that AI systems suggested more diagnostic tests, specialist referrals and medications than doctors.The average estimated cost of care recommended by 20 general-purpose AI systems was USD 314 per visit compared to USD 97 for physicians’ plans. That was equal to a 3.2-fold difference.How Did Researchers Compare AI’s Recommendations With Doctors’?The researchers used 200 de-identified patient history and notes from outpatient and emergency department visits, sourced from MTSamples, a public collection of clinical documentation. They gave the notes to 24 AI systems, including 20 general-purpose large language models and four specialised medical AI systems. Each system produced an assessment and plan based on that information. Researchers then compared these recommendations with the original physician’s documented plan.The researchers estimated costs using 2026 US Medicare rates. The cost comparison included diagnostic tests and specialist referrals. New medications were counted separately. The results showed that AI-generated plans recommended more investigations and specialist consultations, increasing the estimated cost of care.Also read: Woman With Epilepsy Dies After Following ChatGPT’s Advice To Stop Medication: Why AI Cannot Replace A DoctorAI Recommended More Tests Even When Doctors Ordered NoneOne of the key findings concerned visits in which physicians had not ordered diagnostic tests.Doctors had recommended no diagnostic testing in approximately 66% of the encounters. Across these visits, the AI systems consistently suggested tests, often including standard panels of blood investigations.Overall, diagnostic testing recommended by the general-purpose AI systems cost an estimated USD 176 per visit compared to USD 71 for doctors’ plans. AI systems also recommended specialist referrals and new medications more frequently than doctors. The researchers saw an average of 1.14 new medications per visit across the AI systems, compared to 0.59 in physicians’ plans.These findings suggest that AI-generated care plans may favour additional investigations and interventions.Also read: AI Diet Advice Turns Dangerous: ChatGPT Plan Lands Man In ER With Bromide Poisoning Can Changing Prompts Help Obtain Less Expensive Medical Care From AI? The researchers also examined whether changing the prompts given to AI systems could reduce the cost of their recommendations.Prompts containing savings and budget-friendly suggestions reduced the estimated cost difference, but did not eliminate it. The cost ratio fell from 3.2 times physicians’ plans to 2.5 times when systems were specifically instructed to recommend cost-friendly tests and medicines.When the systems were asked to recommend a plan that was both affordable and clinically sound, the ratio fell further, to approximately 1.4 times the physicians’ cost.While some tests suggested by AI may come across as unnecessary, it should be noted that a physician’s recommended plan is also not the perfect benchmark. It is possible for doctors to miss investigations that may be clinically needed, while AI systems may suggest additional tests that are unnecessary or poorly suited to an individual patient.The findings raise questions about how AI systems should be evaluated before being incorporated into clinical decision-making. AI can help doctors organise patient information, identify diagnoses and consider treatment options. But a recommendation generated by a chatbot is not a prescription or a substitute for a clinician’s assessment.