SUPERCOMPUTING LEGEND IBM Watson, fresh from analysing Wimbledon, appears to have slipped up on the treatment of cancer, giving treatment recommendations described as “dangerous”.
Internal IBM documents obtained by medical journal Stat suggest that bad training by IBM and doctors from Memorial Sloan Kettering Cancer Centre led to a series of incorrect recommendations during tests.
Thankfully, the people and their cancers were all fiction, but they did serve to highlight a problem which, if it had been allowed to go into live usage, would have had the potential to be life-threatening.
IBM’s former deputy chief of health, Andrew Norden is said to have included the information in two presentations during Summer 2017.
It said that the diagnoses were “often inaccurate” and raise “serious questions about the process for building content and the underlying technology.”
A 65-year-old man was diagnosed with lung cancer and heavy bleeding but was prescribed a drug called Bevacizumab which can cause further haemorrhages.
Are we calm about this? Well, no.
“This product is a piece of shit,” said a doctor at Florida’s Jupiter Hospital in a complaint to IBM which was included in the reports. “We bought it for marketing and with hopes that you would achieve the vision. We can’t use it for most cases.”
Ouch. Now, nobody is saying that this is proof that AI doesn’t work in medicine. But it’s early days and this incident reflects one of the biggest concerns in the field, and it’s one that has affected every part of computing since Babbage and Lovelace.
The involved parties are keen to point out that this is exactly why AI needs so much research before it becomes stable enough. Doctors testing are only using it as a second pair of eyes, not an alternative. It’s just too soon for that – the buck stops with the doctor.
At least that gives some leeway if it tries to AI you into a bad assessment at work μ
Source : Inquirer