If you live in London you will have seen the adverts on the underground for Babylon’s online GP service, ‘GP at Hand’. It offers what every patient (and government health minister) wants, instant access to a GP. Health Secretary Matt Hancock is an enthusiastic customer along with 40,000 other customers. All you have to do is to pick up your mobile phone, dial Babylon’s number, answer a number of simple questions from a chatbot, and a cheerful, smiling doctor will appear on your computer screen or mobile phone.
Readers of the Salisbury Review may recoil. How nasty can the world get when old Dr Reaction who you have been going to for years, is replaced by a hand waving young metropolitan, ‘reaching out’ to ‘share’ details of your chronic constipation? What about problems ‘down there’? Surely you will not have to take photographs with your iPhone?
While it is true there are limitations to this type of consultation – lack of a physical examination, undue reliance on what the patient is prepared to reveal about his history, or may have forgotten, and the ability of the doctor to get the patient to reveal them – 70 per cent of diagnoses can be made on the history alone so most patients do not require extensive tests, if any.
Surely this chatbot thing cannot compare with the accumulated knowledge of centuries of medical practice? Not so say the machine intelligence experts. Babylon’s doctors are aided by a computer which the owners of the company claim outperforms human candidates in a difficult postgraduate exam GPs sit if they want to practice independently. The average pass mark for human doctors who had trained for 12 years was 72 per cent over 5 years; Babylon’s AI (Artificial Intelligence), which is 2 years old, scored 81 per cent at its first attempt.
The President of the Royal College of GPs, faced with a doctors’ version of the 1980s Wapping Printers dispute when the old hot lead printing presses were forced to give way to Microsoft Word, poured scorn on this claim, saying that the questions Babylon set its computer were intended for revision-only purposes and did not match those of the actual exam.
But when Stamford University and the Royal College of GPs tested one hundred clinical ‘vignettes’ on twelve experienced GPs, none with any connection to Babylon, and the patients were played by practising GPs (some of whom did work for Babylon) human doctors correctly diagnosed between 64 and 80 per cent of questions; Babylon got 98 per cent. When triage decisions (who to send to hospital and who not) were compared; humans had only a slight edge over Babylon; doctors 92 per cent, Babylon’s 90 per cent.
In terms of computing power and machine intelligence there is nothing very ‘clever’ about Babylon. It uses a tree-like structure to process and score answers similar to systems already used by a number of US automated counselling services and other types of non-medical triage systems such as, ‘Why will my car not start?’
This leaves Babylon, and programs like it, with several advantages over conventional general practice. The first is accessibility. You can ‘see’ a GP any time of the night or day. The second is cost. While Babylon requires physical offices to which patients can go if they need face to face consultations or tests, they can be concentrated in large centres that take advantage of the economies of scale. One centre could replace a dozen conventional surgeries. A doctor’s home visit will become rare. Woken with pain in your chest? Hold your iPhone to your chest so a computer can read your ECG online. Spit on a special piece of litmus paper to measure chemicals that indicate damage to the heart muscle. Three minutes later an ambulance is on its way. It may be that within ten years the streets are so packed with people thanks to recent successes in reversing old age that the ambulance cannot get through, but that is another matter.
The service will get cheaper. There is nothing to stop digital doctor systems from employing doctors from low wage countries who can pass the PLAB (Professional and Linguistic Assessment Board test for foreign doctors) who will not have to come here to practise. They can do it from their bedrooms. Babylon’s directors talk of their system being one day on sale to the rest of the world under the NHS brand.
Could AI dispense with doctors? Will machines ever be sufficiently advanced to digest an infinite number of clinical encounters, correlate and score them for relevance – add in the entire corpus of medical and biological knowledge and produce lightning answers to anything asked?
The most advanced form of self-teaching machine so far is the chess-playing computer program Alpha Zero. Given the basic rules of chess, it turned itself into an all-conquering chess master (in twenty-four hours) by playing itself over and over again, reinforcing its previous knowledge at each step.
Starting from random play, and given no domain knowledge except the game rules, Alpha Zero achieved within 24 hours a superhuman level of play in the games of chess and shogi [a similar Japanese board game] as well as Go, and convincingly defeated a world-champion program in each case.
Denis Hassabis, founder of Deep Mind and the CEO of the computer company that developed Alpha Zero. claimed that ‘It does not play like a human or like a computer.’ Pointing out that it makes crazy sacrifices of high value pieces to lure its opponent into a positional trap, he remarked, ‘It plays in a third, almost alien way.’ Other experts however maintain that as a great deal of human thought went into the building Alpha Zero, one should be careful of attributing superhuman powers to a human creation.
Nevertheless, medicine is very like chess. During a consultation a human doctor ‘sees’ a type of medical chessboard in her mind’s eye, with its defined moves, counter moves and resulting ‘play’. Moves include picking up the phone to the hospital, ordering a test, or reassuring the patient and sending him on his way.
Whether machines will ever become conscious does not matter, at least for now. We have no evidence to suggest Alpha Zero is conscious and there is no reason why it would have to be. Nor are we, without a physical model of consciousness, in a position to judge. All we can rely on is the Turing Test. If one cannot distinguish between a human and a machine then we have to assume the latter is conscious and act accordingly.
Comforting yourself with ‘Not in my time?’ Consider the iPhone. Preceded by the Bakelite receiver, then by shoebox-sized portable phones, it was released in 2007 and now guides you around the streets, plays Bach, Beethoven, Dixie, Soul and Rap, translates all foreign languages, recites poetry, summons taxis, delivers your post, pays your bills, shops, and orders your next meal; it can even find you a lover – oh, and you can even consult a doctor on it.
Will patients accept machine doctors? I, like many doctors, was extremely sceptical about the introduction of telephone counselling. Patients would not trust a stranger on the other end of a telephone with their most private thoughts. Quite the contrary, most patients find not having to face a flesh and blood human being reassuring, which is why the service has proved very popular, and as effective, if not more so, than a physical meeting with a therapist.
This is born out by the experience of Relate UK which last year provided 15,000 sessions of live chat online – a demand it expects will double in the next two years. Similar programs in the US such as Woebot and Tess, which offer online behavioural counselling sessions, report a similar success.
For those patients who need to see the doctor’s face bio-engineering laboratories are racing to construct simulated humans, ‘avatars’ who look and behave as idealised doctors. The University College of Los Angeles at San Diego is developing machines who mimic ‘inflections of the voice that doctors use, that are empathetic, or that are judged as having a lot of empathy for their patients by themselves, but also by their patients, and also by naive observers’.
In the film Ex Machina, a young man is exposed in an experiment to a highly attractive female robot called Eva, whom, even though he knows she is a machine, he falls in love with. She is able to persuade him to take her side in an escape when she tells him she is a prisoner of her inventor who plans to ‘upgrade’ her, switch her off and use her parts to make a better robot. As soon as Eva is free, she locks him in her electronic cage to die of thirst, escapes to the outside world and vanishes.
If machine doctors are as perfect in their clinical skills as Eva in imitating love, and patients prefer them even though they know they are machines, will we need human doctors? The human doctor and her patient have a common bond – they both know they are going to die; machines do not die, they upgrade. Which is why however hopeless a case may seem a human doctor is driven by common humanity to do everything he or she can to save the patient. Machines are not human, they do not share our mortality, and are therefore not sympathetic to such irrational considerations. Given the considerable computing power they will develop in the future, they may develop an agenda of their own in which we play no part – perhaps in a language to which we will never have access.
Myles Harris is our Editor
This article appeared in the spring edition of the Salisbury Review. the next edition will be on June 1st 2019 . Subscribe
I do not wish to be facetious but it may well be safer for me to leave the house in a suit of armour every day but again “creepy” and would only contribute to the general unease felt by so many. I take it you mean that we cannot dis-invent things. Respectfully, if your opinion hardly matters the why this post in the first place?
You guys must stay ahead in military AI at the very least – it’s the only way to preserve Western civilization and Enlightenment values.
As an aside (again), Mr Henry Marsh writes that young US brain surgeons are better than UK ones because they do more operations; the UK ones are hampered by the EU Working Time Directiv. I.e. medical hands-on skills improve with practice. But Dr Siddhartha Mukherjee writes that AI is already much better at correct diagnoses than even veteran radiologists…
Having read this through several times I am still not clear as to whether Myles Harris considers a consultation by mobile or computer to be a good or bad thing. In layman’s terms I feel this world is becoming “creepy” and somehow sinister. Was it always going to end like this?
It may be creepy and sinister but it is safer. My opinion hardly matters as these things are already with us.
‘A machine cannot have passions or aims that are not given to it by a human” (Mcmanus) this generation of machines are designed by humans but in future, machines will design and upgrade new machines, nor do we know if they will develop an internal language or not. A random computer has to learn the rules is true, but there are another set of rules, the physical constants of the universe, using which we were created, or evolved, which all things must obey. The more we delve into our structure, the more machine like we appear to be. Why does a new mother always or nearly always fall head over heels in love with her new baby, the soul or a hormone ?
I do not wish to be facetious but it may well be safer for me to leave the house in a suit of armour every day but again “creepy” and would only contribute to the general unease felt by so many. I take it you mean that we cannot dis-invent things. Respectfully, if your opinion hardly matters the why this post in the first place?
Hippocrates, or the authors of the works attributed to him, criticised those who diagnosed patients on the basis of reports using the wet/dry, hot/cold scheme of the time. He said doctors must always examine the actual patient. This AI business is a reversion to 2500 year old bad practice. Such is progress.
That’s a bit like saying a pilot of an airliner would be better served using an old fashioned compass and sextant to navigate than a global positioning satellite. Yes you have to switch the equipment on, and yes, you have to decide on the destination, but machines do things in both aviation and medicine no human being could ever do. The old saw ‘medicine is an art not a science’ is something I still hear at dinner tables. Medicine is at the very heart of biochemical and fundamental physical advances in the understanding of who we are and how we function and think. Medicine and its allied sciences are modern philosophy.
Yes, I see I’ve got that wrong – too hung up on confronting any idea that machines could replace humans, or think like humans. Could a chess machine decide to let its opponent win for a change I wonder. Could a computer medic consider whether the treatment for the diagnosed illness was likely to do more harm than good given the person’s age and disposition and at unreasonable expense.
As Rachel Crompton’s William advised, I mustn’t use up too much of my brain. I’ll have to sleep on it, unlike a computer.
The second a definite yes by comparing survival figures from a large number of sources. Could a computer decide to lose a game ? Not at present but if it ‘noticed’ it was advantageous to lose a game or two in the wider scheme of things, if it had the figures, possibly yes.
Okay. If we concede that a computer could pass the Turing test, maybe even now, there’s one last straw I want to cling to. (I’m painfully close to having to revise a fixed opinion!) A computer might let a child win a chess game, or tell someone with heart failure that they were ‘just getting on a bit and needed to ease up on the fags and booze’ without using the dread words heart and failure. But it would not do it out of kindness like Grandpa or a GP might.
Turing’s test deals only with overt behaviour and that’s a limitation I’ve only just noticed. Machines don’t have souls – even vegetative ones that plants have in some eastern religions. Only a living thing can have emotions, so as long as the programmers are decent people, we are safe.
A good essay, Myles. Kurzweil’s “singularity” comes to mind.
As an aside, AI is rapidly destroying the livelihoods of translators across the world. Altho’ Google Translate etc. produces hilarious gaffes all the time, and is easily fooled, the PERCEPTION among customers is that machine translation is already better, so human translators get fewer assignments. It is already much cheaper. Accurate and sensitive translations (e.g. literature and poetry) still require human input, but consumers of translations often only want to know the general drift of a text or speech. Again, AI wins.
Medical texts are easy for AI to render as they consist almost entirely of unambiguous statements and technical terms, which invariably have one-to-one equivalents in all the diverse languages here in Asia.
According to machine learning expert Pedro Domingos, the only defense against AI is to understand what it can do for you and use it as a tool to help you. That is fine so long as the “tool” doesn’t learn all that the user knows and then decides to fire, or worse, exterminate, its employer!
My very limited experience of google translate was so hopeless that my experience remains limited. Odd words maybe – foreign students sometimes used a handheld device for that. As to predictions – has anyone ever got the future right? personally, I can’t take seriously any notion that machines can replace humans never mind think or outsmart them. How we think, know, decide, choose and so on are still mysteries, along with free will. The one thing we can know for sure is that whatever empirical science says today will one day be set aside is deluded and redundant. What reality-free speculators think (bet they can’t even say what thinking is)is something for late night chatter with a wine-addled brain.
Babylon is not a prediction but a fact. Try their computer with a set of standard symptoms. The initial assessment is free. The company is now opening in several more cities. AI diagnosis will be widespread in five years. I don’t know when you last used Google Translate but doctors, myself included, use it for translating medical files to send or receive from foreign hospitals. It’s pretty good.
It’s foolhardy predicting what science may not be able to do in future. About the turn of the 20th century there were plenty of authoritative figures who harrumphed that man could never fly. Had you told them that not only can man now fly, but a box of electronics can fly 450 people from Melbourne to London without a pilot touching the controls, they would have guffawed. Contrary to what we suppose about consciousness, it is widespread, much more advanced in animals than we have hitherto thought, and like other pure discoveries in fundamental science, like DNA, may be mathematically ‘beautiful’ in concept, that is complex yet simple in its application, and I suspect of wide applicability. There is nothing wine addled or speculative about Alpha Zero. It exists now, it will beat you at chess and if it recognises you are any good at it will use human like tricks in the process. Turing’s test is a cyanide pill to the ‘machines will never be able to replace humans’ assertion. They will and in many cases you won’t even know it.
Thank you. My resistance to the notion that machines might one day develop minds of their own is emotional not rational – I admit. But that is the difference between human and machine – the chess computer knowing only all the moves would remain motionless or shift pieces randomly if it had not been programmed – motivated we’d say for a human – to win. A machine cannot have passions or aims that are not given to it by a human. I’m troubled by the medical diagnoses success because in my own case, and in cases I know, the doctor picked up on nuances that were not simple answers to questions. I’m thinking of a family member with psychiatric problems who did not present the cold facts to the physician but he sensed that. Could a computer do that I wonder?
From ancient times people have posited a soul to explain what is different about people. Animals have an animal soul that has appetites but not the self-examination, and ability to resist impulses that humans have. A computer is surely a rational logical performer only, and reason alone supplies no motive for anything. It surely cannot choose to upgrade itself – that must have been built into it. No logical argument can decide for anyone (nor a machine) why one should continue to exist: the decision would depend on whether we had an appetite for living, or were tired of life. Emotionless people – psychopaths – are possibly like that.
The Turing test is sometimes failed by humans in my experience. (My first job 55 years ago was teaching teenagers excluded from schools, psychiatric wards and other centres that couldn’t cope with them.) Recently I tried to convince a socialist anti-Semite that his arguments were wrong and common among his kind. (The Jews have always been hunted, so there must be something wrong with them.) I explained that his reasoning was the same as saying animals have always been hunted, women oppressed, so there must be something wrong with them too. He quibbled about the meaning of words like Zionism, but returned again and again to his same point. He could have been a machine unable to move from one programmed line, and so failed the Turing test.
I want to be optimistic. I don’t want robots to rule us. (We’ve had a taste of that recently some people say.) I can’t imagine debating this matter with a computer, and that’s one reason why the more extreme forecasts strike me as impossible.