back to article IBM Watson dishes out 'dodgy cancer advice', Google Translate isn't better than humans yet, and other AI tidbits

Hello, here's a short roundup of this week's news and announcements in AI, including worrying news for cancer sufferers, good news for human linguists and some new job opportunities. IBM Watson cancer fail: IBM Watson has made several “unsafe and incorrect treatment recommendations” to cancer doctors using the technology, …

Silver badge

Don't worry IBM!

The British government are a bunch of technically illiterate patsies with a pathetically touching belief in the infallibility of technology to cut costs and improve outcomes. Because the Department of Health is clueless in all matters relating to health (clinical, ITSec, privacy, cost etc), this makes for the perfect setting to sell your hugely expensive, ineffective and dangerous system. And for training data, they'll just gift you the entire data set of the NHS without any checks, balances, or patient permissions.

So, wipe away the tears, and bring your Watson PieceOfShit (tm) to the UK. Some gormless minister will sign up straight away. Obviously these systems don't just happen, and you'll need an implementation partner who doesn't care if the system doesn't work so long as it ticks the boxes on the SLA. And they need to have no conscience that it costs a fortune for the client. I think Crapita plc will fit the bill.

34
1
Bronze badge

Re: Don't worry IBM!

IBM, reclaiming POS for IT from cash registers since Ginni.

@Ledswinger

You're half right I suspect, "improve outcomes" is marketing speak for cheaper... improve bonus outcome.

6
1
Silver badge

Re: Don't worry IBM!

Artificial Idiocy fails because we truly do not understand intelligence in a dog or a cat nor do we understand it in a human. Our pets are capable of learning and show a degree of intelligence which we can not quantify. But we can not mimic this behavior in a device. Idiocy systems in use today are not intelligent but just massive databases with a very powerful scoring and query engine on top. If the data is bad, the results are bad. If there are flaws in the scoring algorithm or query generation (almost certainly in both cases) there is a variable risk of bad results.

23
1
Silver badge

Re: Don't worry IBM!

they'll just gift you the entire data set of the NHS without any checks, balances, or patient permissions.

And that is why it would work for the NHS.

If you want to detect cancers early from unrelated symptoms you need ALL the data from ALL the people in an unbiased sample.

You can't detect diabetes risk from poor diet if you take data only from US hospitals that charge $100 for an aspirin, or from genes in some racial group if that group opts-out because their data is going to be handed over to the police.

So grab the data. process it in the UK, by a university that isn't going to de-anonymise it and use it to fix car insurance premiums. And don't worry that a few Gaurdian columnists are going to whine about owning our own data

8
7
Anonymous Coward

'bring your Watson PieceOfShit (tm) to the UK'

That straight up almost 'Onion' like take on the situation is terrifying. Can anyone see backlash coming against tech soon? I can... I thought that ex-FB guy Garcia Mendez was nuts for building a home in the woods within paddle distance of Canada.

But... Its officials blind faith that tech can do no wrong to make THEIR own low-hanging-fruit lives easier that's beyond terrifying... Because you know that some technocrat is going to say, 'Sir, are you classified as human? Cos the computer says you're a Meat Popsicle. Therefore we can't help you sir'!

6
1
Anonymous Coward

_AI_Hype_ - _AI_Hell_

Somehow the unholy cabal of Govt, big-media and big-tech (the real beneficiaries here), have all gotten into bed together and deluded themselves into believing that we've suddenly arrived in the future where this shit actually works. Along with the delusion that all the security, privacy and ethical aspects have also been solved. WTF? Any rockets going to mars? Get me outta here!

7
0
Silver badge
Facepalm

Re: a_yank_lurker

We do to some extent understand neurons. These systems replicate neurons. So they have very basic "learning" systems.

But would you trust your data set, and decisions on a slug?

For those systems with more power/neurons, we do not know how to teach them (or the "correct/best" maths/weighting/logic for connecting them) correctly. Proof of this, a 3 year old has much more power than these systems, yet would you trust even an educated 3 year old to make medical decisions?

14
2
Silver badge
Alert

Re: Don't worry IBM!

And for training data, they'll just gift you the entire data set of the NHS without any checks, balances, or patient permissions.

Indeed - the data set with misdiagnosis, incompetence, coverups, silenced whistle blowers etc

AI - shorthand for "Garbage in = Garbage Out"

7
0
Silver badge

Re: Don't worry IBM!

@ ledswinger Aside from agreeing with your post, I think you have discovered the true origin of the word Posh. It's not as usually posited, ' Port out, Starboard home', it's a modified acronym POSh. All the time we have been thinking posh things are better it is the other way round so posh people are less well regarded and possibly more accurately described.

" Look at those posh twats in that expensive cabin." Etc

3
0

Re: Don't worry IBM!

"If the data is bad, the results are bad. If there are flaws in the scoring algorithm or query generation (almost certainly in both cases) there is a variable risk of bad results."

@a_yank_lurker -- not limited to AI systems.

How else to explain Donald Trump?

0
0
Silver badge

Artificial intelligence

meet natural stupidity.

GIGO.

9
0

Re: Artificial intelligence

You misunderstand, this is an evolution of the win at all costs chess strategy.

Yo can't beat Watson if you're dead

0
0
Anonymous Coward

Oh My

Let's go right to the moment where a competent doctor gets the patient treatment recommendation from Watson. He/She looks at it and instantly realizes it is junk and may actually kill the patient if followed. It has to be sooo wrong that a doctor calls this a "piece of shit", which is not common medical terminology, but highly expressive and meant to convey serious dissatisfaction in an area they have a lot of experience with. This is Theranos-level serious. IBM has been promoting something that does not work and because they have armies of experts, they know it does not work. It's lawsuit o'clock somewhere in the world.

21
0

Re: Oh My

"Let's go right to the moment where a competent doctor gets the patient treatment recommendation from Watson. He/She looks at it and instantly realizes it is junk and may actually kill the patient if followed."

This will probably happen quite a lot. However, the first time that a human doctor overrules Watson and there is a negative outcome, the bureaucrats will make it mandatory to do what the software says. It will be like airlines where their operating procedures force pilots to use automation because "it's safer" right up until the point where it isn't.

The people who made the decision to shell out megabucks on the shiny tech will be unwilling to admit that they were wrong. I'm sure we've all seen it when a stupid solution is pursued beyond all reason because admitting it was a bad idea would mean some senior person having to confess to being wrong.

3
0
Anonymous Coward

Google Translate 'AI-Hype'

On individual sentences / phrases, 'goOoGle tRansLate' is especially poor. Whereas across entire translated articles, things are better obviously. But have things improved? I regularly translate bits of Spanish to English, which is not a particularly hard combo. Yet my pathetic skills have improved versus AI. 'Goo' figure! Informal or local phrasing is nearly always a hopeless FAIL. That was to be expected a decade ago. But now, most other language sites are actually better imho!

5
1
Anonymous Coward

Re: Google Translate 'AI-Hype'

(Edit as post timer ran out)... One of things that Goo Translate does badly is NOT offer a hover-over alternate choice or a second box with an alternative translation. In that sense the average dictionary/translation combo site, is a class above Goo translate!

2
1
Silver badge
FAIL

Re: Google Translate 'AI-Hype'

Google translate is still a great source of amusement! Insert something from classic Finnish literature, translate to English, and the result is invariably surreal: "Their father, who was a severely eager forearm, met at his best age when he was suddenly killed when he fought with the tiny bear. Both then, Kontio of the woods as a man, were found dead, one by the other, landing in a bloody tanter. The wound was wounded with wounds, but both the throat and the side of the beast were seen by the knife cut off and his chest pierced by a rifle perched on a rifle. So stopped a rotting man who had dropped more than fifty bears."

> But now, most other language sites are actually better imho!

The Bing translation of the same passage makes slightly more sense, but not much: "Their father, who was a fierce avid forest man, faced his best-hearted death when he was fighting with a sudden bear. Both at the time, both in the forest of the Sam and in the man, were found dead, one after another, with the blood on a bloody tanker. The man was badly wounded, but Pedonkin both throat and flank were spotted with a knife and his breasts pierced by the rifle's Tuiman. So ended the day with a stalwart man who had poured more than fifty bears."

Well, at least it does not translate a fierce bear into a tiny bear, like Google. "Forest man" is also a bit better translation for "metsämies" ("hunter") than "forearm" (How on Earth did Google come up with that?). Finally, here is a quick human translation, by me. Not likely to win any literary awards, but at least it means pretty much the same as the original: "Their father, who was an eager hunter, met with sudden death in his prime, when fighting a fierce bear. Both, the bear of the woods and the man, were found dead, lying side by side on the blood-stained ground. Badly was the man mauled, but also the throat and the side of the beast were seen sliced by a knife, and its chest was pierced by the powerful bullet of the rifle. So ended his days a strong man, who had felled more than fifty bears".

11
0
Silver badge
Coat

Well Bing may be better in this case,

but if you can drop a bear, you have my respect.

3
0
Silver badge

Re: Google Translate 'AI-Hype'

Google translate menu fails even have a twitter account devoted to them

https://twitter.com/menufails

0
0

Watson...

Quite valid, you are only as good as your data. Perhaps a nice data set can be obtained from post-Brexit UK where GDPR and HIPAA don't exist-- of course, after Watson-learning-scraping, if you aren't a Brit the recommendations may well kill you precipitating another round of murderous Watson stories.

Which brings up-- how badly some oncologists perform, except that they bury their mistakes and certainly don't go air out their dead body pile in public. Is Watson better than these death dealers?

Plus, it is well known that American docs are extremely resistant to taking any advice from anyone, the most recent evidence being that a large percentage of maternity wards refuse to follow the most simple and obvious guidelines (on high blood pressure and maternal blood loss ("my eyeball is calibrated good enough thank you")), resulting in America having deplorable levels of maternal morbidity compared to any other first world country. So the Jupiter docs whine about Watson, but how much is real and how much is "I and my swelled head would do it differently"?

Personally, Watson doesn't seem well suited to oncology advice as presently implemented. If enough resource was invested, Watson could become quite respectable. It isn't obvious that the resource will be invested, between slow revenue gains and vested interest attacks Watson oncology may suffer a fatal monetary infarction.

Overall, we are currently in an AI hype cycle and AI is still does not appear ready for prime time. Anyone who had been around for enough years has seen these cycles before. The cycles happen about every 15-20 years as a new generation thinks they discovered AI. One could hope this time is different, but the evidence is underwhelming so far.

5
3

At least you know it's rubbish

Google often "translates" my Japanese friend's Facebook posts as a word soup of unintelligible non-English non-sentences from which it is impossible to extract any meaning whatever. So at least I don't waste time trying to understand it.

When it does make something like English sentences, the fact that it doesn't take account of context is painfully obvious. For example it frequently translates what should be "My husband" as "Your husband." That also shows its ignorance of a basic point about the language it's supposed to be translating: in Japanese the default pronoun (when omitted and not obvious from context) is the first person.

4
1
Silver badge

Re: At least you know it's rubbish

" in Japanese the default pronoun (when omitted and not obvious from context) is the first person."

Surely this is also true in English, at least the wife says so.

5
0
Anonymous Coward

"ignorance of a basic point about the language"

Isn't that "by design"? I though the Goolge algorithm was language agnostic, making it easier and quicker to add support for new languages without having to try and model sematic inferences? Obviously doesn't work all the time, but...

0
0
Silver badge

Re: At least you know it's rubbish

"Surely this is also true in English, at least the wife says so."

Go and tell that to the Marines.

Surely the default in English is the second person?

0
0
Silver badge

Re: At least you know it's rubbish

The default in English for a condition that is less than favourable is to blame the EU.

3
1
Silver badge

Should have gone to Google...

“This product is a piece of shit,” a doctor at Florida’s Jupiter Hospital said to IBM, according to the documents seen by Stat News. “We bought it for marketing and with hopes that you would achieve the vision. We can’t use it for most cases.”

COUGH-COUGH-COUGH

6
0
Silver badge

Re: Should have gone to Google...

You should see a doctor about that cough, I hear they're making great advances in diagnosis & suggested treatment using AI...

0
0
Boffin

No button measuring replies needed !

Clean and real medical data-sets are being compiled now...

In Australia, GPs and hospitals are being paid to upload medical records.

Addressing obvious conflicts of interest with privacy or consent, it comes with a 3 month opt out window else the government’s digital 'My Health' record in perpetuity.

The opt out is from 16 July to 15 October 2018, then presumably someone hits the big <LOAD> button.

No button measuring replies needed !

2
0
Silver badge
Pint

Peak A.I. (hype), followed by the inevitable crash...

Late-2017 into early-2018, the foolish hype about A.I. seemed to peak.

In recent months, we're growing accustomed to reading about endlessly Self-Crashing Cars, Racist Facial Recognition used to semi-automate the bothering of minorities, A.I. Stock Traders have failed to become Trillionaires, and now even poor old Watson has been revealed to lack common sense.

Somebody should run an FFT on the hype cycle to figure out the precise period. It's *about* 12 years.

2
0
Silver badge
Holmes

DOCTOR John H. Watson would be bitterly disappointed...

"By Jove, Holmes! They've put my name on a piece of crap that endangers patients!!"

(Icon is once again eerily suits the occasion! Uncanny job in choosing it, El Reg!)

5
1
Silver badge

"Google Translate isn't better than humans yet"

Granted, Google Translate is better than me at translating anything into English, but will machine translation ever really be "better than" humans?

0
0
Silver badge

Re: "Google Translate isn't better than humans yet"

I would hazard a guess that Google may be too interested in the marketing possibilities for them to ever get language sorted properly. As others have pointed out here GIGO. And one of the inputs from Google is bound to be something to do with money.

The more I read and play with AI the more I am convinced it has some amazing possibilities (GIGO may also come into play here) but the data fed into the AI has to be free of bias and the choice of functions for it from a commercial company is bound to be biased. It really needs to be taken away from commercial instincts and some form of pan-global open sharing of everything involved is possibly the only safe way forward. We have the technology - I bet in 5 years we will have ARM devices of the processing power of Titans running on a couple of Watts and costing $10s not $1000s. If governments dont get on board and co-operate soon we really will be fucked.

1
1
Silver badge

Re: "Google Translate isn't better than humans yet"

It''s a pipe dream Humans are biased, plain and simple, so anything we make will be similarly biased, by default. The only way to remove the bias is to make it completely random.

1
0
Silver badge
Pint

Re: "Google Translate isn't better than humans yet"

Sometimes (often) I'm actually quite happy with a simple and literal word-by-word translation.

For trivial example, translating "s'il vous plait" into "please" would just add to my confusion. Oftentimes, I'd rather have the precisely literal version: "if it you pleases", because then I've learned several words at the same time (can be easier). Not to mention making transparent the grammatical structure, and revealing the bizarre idioms, at the same time.

At the very least, 'Literal (word-by-word)' translation should be a selectable option.

2
0
Silver badge

Re: "Google Translate isn't better than humans yet"

> At the very least, 'Literal (word-by-word)' translation should be a selectable option.

Words in isolation usually have multiple meanings. To make sense, you have to have some idea what the text is talking about. The funniest example of ignoring context I have seen was a packet of Iranian dates I bought a few years ago, with a label in Finnish advertising "Tuoreet päivämäärät". Well, "päivämäärät" is "dates", but only in the sense of a date in a calendar. Someone had apparently trusted the Google translator: it produces this text from "fresh dates".

1
0

Re: "Google Translate isn't better than humans yet"

Well I've had a 3 minuate play with google translate (Maori to English translation):

"gaga ga ga ga ga ga ga ga"

becomes

"How great it is!"

You couldn't not make this up......

1
0

Re: "Google Translate isn't better than humans yet"

A few more attempts on the translate produces

"gaga ga ga ga ga ga ga ga"

becomes

(Maori to English) -->"How great it is!"

(Amharic to English) --> "Gas Station Shuttle service"

(Scots Gaelic to English) --> "Gaga is enjoying her and enjoys her"

(Armenian to English) --> "Goto Top Level"

(Georgian to English") --> "You cannot use the gag"

(Japanese to English") --> "I'm sorry but"

The translator must surely just be a word database and a random number generator

1
0
Silver badge
Boffin

Don't throw the baby out with the bathwater

In some areas AI/ML work really well because the data is too complex for a human to get their head around. A good example is cancer where there the data are massively multivariate, an example of this is the PAM50 classification for breast cancer, which as the name suggests, has 50 "clusters" with different outcomes, treatment regimes etc.

3
0

I'd remain skeptical but open minded where Doctors have opinions on AI enabled software that could replace them. Turkeys may be well educated, but they also have serious college debts and crazy salaries, with a track record of failing to vote for Christmas.

In the one example failure I've heard from this report:

"In one example, a patient was recommended a drug that could lead to severe or fatal hemorrhage while he was already dealing with severe bleeding due to his condition."

Drugs have side effects, as I understand it in Britain (I am married to a pharmacist), A doctor will typically recommend the standard treatment, it is then necessary for a pharmacist to confirm whether a drug is suitable for the patient based on their other medical history. My wife spends half her life battling with Doctors who "did not know" the side effect profiles were unsuitable for certain medical conditions - so to me it sounds like the AI chose the most suitable drug, that would probably have been rejected by the pharmacist.

It sounds like IBM watson needs another UI message that says "Here are the contradictions to be aware of:_________________"

1
0

I'd remain skeptical but open minded where Doctors have opinions on AI enabled software that could replace them. Turkeys may be well educated, but they also have serious college debts and crazy salaries, with a track record of failing to vote for Christmas.

I'd remain sceptical but open minded where IT companies have opinions on AI enabled software that could net them large profits. Hyenas may be well educated, but they also have serious costs buying yachts and private jets and crazy salaries, with a track record of telling the world that nirvana will be round the corner really soon, or definitely, probably, once the bug fixes to Version 11.5 are released.

3
0

Entirely true and valid points.

In oncology, there is a surprising amount of logical steps that can be performed that are literally IF/ELSE statements that lead to a resultant pattern of care. There is an awful lot of background theory, but you could probably train a Visual Basic Macro in Excel to calculate with moderate accuracy what the treatment should be.

Just as an example, breast cancer, you may determine the size, location and severity of the tumour, run blood test and DNA diagnostics for various known genetic faults (possibly: BRCA, P53, HER, FAP etc.), there are recommended 'standard' protocols advised for many 'common' tumour developments. The problem is that there are possibly 20-30 different treatments with high theoretical learning curves.

For those of a nerdy inquisition, I really do recommend reading about the above cancer genes. As our understanding of DNA is improving, it is reaching the stage where many cancer progressions have been mapped in processes that are akin to typical electronic engineering applications. i.e. damaged hormone growth receptor results in uncontrolled growth, faulty DNA repair protein fails to detect/repair tumours, etc.

1
0

They sell advertising...that's what they do...

Trust a fucking advertising company to diagnose illness? Do me a favour.

2
0
Anonymous Coward

Serpens oleum

It will all turn out like insurance comparison or energy switching sites. It will look nice and unbiased, but actually the recommended treatment will depend on which firm is paying the highest commission. Just remember to get your doctor to click through to the 'show all diagnoses' button and watch out for small print.

I still like the story of the AI chest x-ray analysis software that they thought was doing really well at classifying X-rays of fluid on the lung - till it turned out what it was good at doing was spotting chest drains that doctor's insert when their old fashioned stethoscope detects fluid on the lung, before they send the patient to X-ray. And that was just the start of dealing with 'real world' data - things like the X-ray having a note 'no sign of problem B ' on an x-ray with clear signs of problem A - because it's actually a follow up to an earlier X-ray that diagnosed problem A. Now this sort of detail can be captured and tracked .- but in the real world? No chance.

2
0
Silver badge

What I'm always missing in the AI hype...

...is how a particular achievement compares to some other approach, i.e. some carefully written, but otherwise trivial program, or some conventional statistics.

4
0
Silver badge
Terminator

You think those incorrect cancer treatment recommendations were a mistake?

Oh no no no, that's just what the machines want you to think.

2
0

Research has shown

that medical doctors do not trust the output of a computer-diagnosis (by AI or otherwise), if the result is computed within 30 seconds. So if necessary software engineers inserted a delay.

In an other context I once wrote a Monte Carlo optimization which I let run on a SGI server for 30 seconds before the result was broadcast on public television.

1
0
Silver badge

Not all machine translation is Google machine translation

compared to human translators, machines have almost zero common sense.

Broadly true, but since "common sense" doesn't have any clear technical meaning, or even a clear common meaning, this isn't a useful observation.

They don’t understand language in the same way that humans do,

Possibly true, but since we don't know very much1 about how humans understand language, this is neither certain nor particularly useful.

and have no knowledge of the world

There are ML translation systems with quite complex world models. Google threw its efforts behind deep LSTM networks trained by unsupervised learning on large corpora of parallel documents, but that is far from the only approach that has ever been tried.

There are also translation systems which do not use a bag-of-sentences approach, and so on. Critiques of Google's approach do not apply across the board.

1That is, while in an absolute sense we know quite a lot about the subject, in a relative one - compared to how much there appears to be to know - we're still at a very early stage.

0
0

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Forums

Biting the hand that feeds IT © 1998–2018