The "deep learning" programme, tested in China, assimilated information from more than 1.4 million electronic health records.
It was then able to draw on its "experience" to diagnose a broad range of childhood diseases, with accuracy rates of more than 90% in some cases.
The system performed better than junior doctors, but not quite as well as more senior experienced physicians.
The scientists who created the AI model believe it could speed up the triaging of patients in hospital emergency departments and improve diagnosis.
But sceptical British experts insisted that intelligent machines could never take the place of human physicians.
The human doctor who asked targeted questions and then used his or her knowledge and experience to make a diagnosis "can be considered a classifier of sorts", said the researchers, writing in the journal Nature Medicine.
The AI programme worked in a similar way, sifting through vast amounts of clinical information "to mimic the clinical reasoning of human physicians".
To create the system the scientists obtained electronic health records (EHRs) from 1,362,559 outpatient visits to the Guangzhou Women and Children's Medical Centre, a major government hospital in China.
The records covered physician-patient encounters from January 2016 and July 2017 involving children and teenagers up to 18 years old.
In total, 101.6 million data points were used to "train" the programme.
Specific words and phrases, as well as numerical data such as patient temperature, were analysed by the system and compared with what it had learned from the health records.
The researchers, led by Dr Kang Zhang from the University of California at San Francisco, US, and Guangzhou Medical University, China, concluded: "Our study provides a proof of concept for implementing an AI-based system as a means to aid physicians in tackling large amounts of data, augmenting diagnostic evaluations, and to provide clinical decision support in cases of diagnostic uncertainty or complexity."
