BBC In Depth: the ongoing battle to define autism

What if... there turned out to be more autistics than NT's? Yippee!!!!!!!

hehe - possibly too much information but you post has just reminded me that I need to visit the loo...

You have every right to be suspicious of "body language experts".  I should know as, in some ways, I made a career out of it.

Somewhat ironic given that I get so engrossed in things that i don't even notice what my own digestive system needs are!

Yeah, that scares the cr*p out of me a bit  I have always been very suspicious of "body language experts" 

Understood Peter1970 - those AI systems viewing facial and eye movements reminded of the science fiction of "Blade Runner".

W/r/to those "tells" I was especially interested and alluding to this scoping review https://pmc.ncbi.nlm.nih.gov/articles/PMC12620671/  there are also individual studies worth looking at -

"digital medicine" is another way of putting a finger on the pulse :-)

Sorry, I can't seem to get the "quote" option to work but the AI response to the turnip question is exactly what I would have expected to a text input query. I was referring more to the logic behind the interface and the programming. The AI can only process what it understands as processable. Something like a nod of the head or a physical reaction to a given question would not play a part in how an AI would process a response because it can't see it. Slight changes in facial expression even previous behaviour that an AI would not be aware of.

In saying that I am also a big science fiction fan so I obviously can't rule out some type of fully autonomous artificial being in the future but I think we are a fair bit away from that at the moment

Please excuse my answering you 2 good people in 1 response -

Peter1970 said:

5+5=10, 20-4=16 but if it comes to 10+turnip=? What will it do?

I've just run that thro' the AI I use and it has replied:

"Because you cannot mathematically add an integer to a root vegetable, the answer depends entirely on how you want to play it:......... 

you don't have to imagine how an AI could be used because other people already have.

AI can detect the exact same behavioural cues a clinician looks for, but it does so by breaking them down into precise measurable data.

yep- the Ai I use can "see that" by recognising it in a search and says it's a joke too

PS love Dawn French :-)

AI training is still something they have not really perfected I think. I was involved in programming very early forms of AI. It was just datasets with basically if/then decision trees. Basically  a questionnaire with a list of possible responses based on user choices. That then develops to expanding the datasets based on responses entered. This still needed human intervention to make sure the data in the datasets were valid. Still based on preprogrammed if/then decision trees which can be added to based on new data gathered. It is all manageable and simple to this point. But modern AI systems not only write their own datasets they now write their own decision trees (theoretically based on decision trees parameters set by humans) but these become so large so quickly that regardless of how well a company claims they manage a self learning AI system at some point it becomes impossible and impossible to find where the route of a potential problem began, they are just far too vast for even the largest team of developers to manage. So I would not trust an AI to be able to, as you say easily recognise subtle changes or idiosyncrasies in human behaviour or stick to any type of parameters that would try to ensure this. 

I suppose what I mean is that any computer system works on logic.

5+5=10, 20-4=16 but if it comes to 10+turnip=? What will it do?

I just told Co-pilot the joke and it did recognse  it as a joke and a very British one at that, but we had been talking about absurdist British humour yesterday and I think it just picked up on it.

Do you think so? So many computer programs have problems and biases that are quite subtle depending on who's training/programing them. Even things like facial recognition programs are less likely to be acurate for non white or South East Asian populations, we're unconciously building these biases in. In this case though I wonder what it says about the people who are giving these programs their learning material?

Do you see the AI diagnosing autism as learning on the job as it were? If so then what about the people it missed whilst gaining experience, I know this can happen with newly qualified humans too, but would the AI have the wit to ask a colegue and would it have the sort of professional supervision that many humans have to have?

I agree that computors are better at sorting through large amounts of data faster than humans can and I can see that for medical screening, of things like mammograms they're faster and just as or even more acurate than humans. But I think that is different to  how we experience the world and so much of autism is about our experiences in the world. A human for example could ask for an example of being humiliated at school and have an empathetic response, but how does a machine learn that response? I think a good example of this comes from when I used a joke whilst I was talking to AI and it took me seriously and came out with scads of serious stuff about this joke and wasn't even sure when I told it it was a joke comment and not to be taken seriously and this is with something that knows my conversational style. I can't imagine how a machine that didn't know my conversational style would even hope to pick up on the subtlties of human communication in a diagnostic situation.

Maybe as an experiement we could tell our AI's the same joke, not tell them we're telling them a joke, but just to see how it responds, we'd have to tell it the same joke, so heres mine.

Cuthbert was an inflatable boy, who went to an inflatable shool and one day the head master came up to him and said 'Cuthbert!, You've been found running in the school corridors with a drawaing pin and I want you to know that not only have you let me and the school down, but you've let yourself down too',

Yes you are more skeptical on these matters than I TheCatWoman.

Broadly speaking you ask how would the AI learn to understand differences - I would answer that question by saying that it would learn to.  I agree with the core rule "garbage, in garbage out" for the use of any system, computer or tool (hehe except that as is also said "one man's junk is another man's treasure").

Computers are in my opinion far better at handling vast amounts of data and complexity than human beings - yes they can be biased of course but that bias in my opinion is easier to resolve in them than in most humans

The childhood form you mention is interesting. I was sent a form, but had no one to complete it which they accepted and I did in my meeting mention things I remember. I was glad I didn't have a parent to answer that, as my last surviving parent remembered very little of my childhood and probably would have said no difficulties. 

I was diagnosed before DSM-5 came out, I don't know what criteria I was diagnosed under, I had half a dozen meetings with a psychologist. I don't know what paperwork is involved in the tests we have now.

I think I'm more sceptical than you about the benefits of AI in this sort of scenario, theway you describe it being used sounds great, but would it actually be used that way? Or would the autistic people prefer tech to human contact thing take over and it all become computerised? I think one of the problems with an AI approach and maybe the tests used currently are with so many older people being diagnosed is it nuanced enough? Would it be able to get that some people struggle with tech because of age, that they may never have played a computer game, that the cultural references maybe totally different? For example, if asked about food, I suspect many older people would give totally different answers to younger people. When I was a child you ate what was put in front of you, a smaller portion of what your parents ate, there were no "childrens foods", things like chicken were a bit of a luxury, crisps and fizzy pop were luxories too, not everyday foods, so the beige diet that so many autistics prefer might not have been possible for us latelings. Also, in other cultures, such as Indian and Italian, food is a massive thing, sitting round the table, eating loads of home made food, foods made by your gran and aunts, it would be a massive insult not to eat it and to say you don't like it. You learn to eat what you don't like so much from being told to eat it and not being allowed to get away with not eating and sitting around the table. I believe Chinese culture's the opposite, they eat as fast as possible and don't take time over food and think we're weird to "waste" so much time cooking and eating. How would AI deal with these cultural differences?

Obviously I don't know what the tests involve but I hope all the above gives an example of the complexities that would need to be trained into AI.

How would AI cope if there were sections of the testing that you couldn't complete, like I couldn't do the tessalation test I was given. I don't have an informant and I suspect many others don't either. What if the informant you do have is undiagnosed themselves what they'd see as normal childhood behaviour might be normal for an autistic person, so they might say that your childhood was normal, as it was similar to theirs.

I was diagnosed before DSM-5 came out, I don't know what criteria I was diagnosed under, I had half a dozen meetings with a psychologist. I don't know what paperwork is involved in the tests we have now.

I think I'm more sceptical than you about the benefits of AI in this sort of scenario, theway you describe it being used sounds great, but would it actually be used that way? Or would the autistic people prefer tech to human contact thing take over and it all become computerised? I think one of the problems with an AI approach and maybe the tests used currently are with so many older people being diagnosed is it nuanced enough? Would it be able to get that some people struggle with tech because of age, that they may never have played a computer game, that the cultural references maybe totally different? For example, if asked about food, I suspect many older people would give totally different answers to younger people. When I was a child you ate what was put in front of you, a smaller portion of what your parents ate, there were no "childrens foods", things like chicken were a bit of a luxury, crisps and fizzy pop were luxories too, not everyday foods, so the beige diet that so many autistics prefer might not have been possible for us latelings. Also, in other cultures, such as Indian and Italian, food is a massive thing, sitting round the table, eating loads of home made food, foods made by your gran and aunts, it would be a massive insult not to eat it and to say you don't like it. You learn to eat what you don't like so much from being told to eat it and not being allowed to get away with not eating and sitting around the table. I believe Chinese culture's the opposite, they eat as fast as possible and don't take time over food and think we're weird to "waste" so much time cooking and eating. How would AI deal with these cultural differences?

Obviously I don't know what the tests involve but I hope all the above gives an example of the complexities that would need to be trained into AI.

How would AI cope if there were sections of the testing that you couldn't complete, like I couldn't do the tessalation test I was given. I don't have an informant and I suspect many others don't either. What if the informant you do have is undiagnosed themselves what they'd see as normal childhood behaviour might be normal for an autistic person, so they might say that your childhood was normal, as it was similar to theirs.

hehe - possibly too much information but you post has just reminded me that I need to visit the loo...

You have every right to be suspicious of "body language experts".  I should know as, in some ways, I made a career out of it.

Somewhat ironic given that I get so engrossed in things that i don't even notice what my own digestive system needs are!

Yeah, that scares the cr*p out of me a bit  I have always been very suspicious of "body language experts" 

Understood Peter1970 - those AI systems viewing facial and eye movements reminded of the science fiction of "Blade Runner".

W/r/to those "tells" I was especially interested and alluding to this scoping review https://pmc.ncbi.nlm.nih.gov/articles/PMC12620671/  there are also individual studies worth looking at -

"digital medicine" is another way of putting a finger on the pulse :-)

Sorry, I can't seem to get the "quote" option to work but the AI response to the turnip question is exactly what I would have expected to a text input query. I was referring more to the logic behind the interface and the programming. The AI can only process what it understands as processable. Something like a nod of the head or a physical reaction to a given question would not play a part in how an AI would process a response because it can't see it. Slight changes in facial expression even previous behaviour that an AI would not be aware of.

In saying that I am also a big science fiction fan so I obviously can't rule out some type of fully autonomous artificial being in the future but I think we are a fair bit away from that at the moment

Please excuse my answering you 2 good people in 1 response -

Peter1970 said:

5+5=10, 20-4=16 but if it comes to 10+turnip=? What will it do?

I've just run that thro' the AI I use and it has replied:

"Because you cannot mathematically add an integer to a root vegetable, the answer depends entirely on how you want to play it:......... 

you don't have to imagine how an AI could be used because other people already have.

AI can detect the exact same behavioural cues a clinician looks for, but it does so by breaking them down into precise measurable data.

yep- the Ai I use can "see that" by recognising it in a search and says it's a joke too

PS love Dawn French :-)

AI training is still something they have not really perfected I think. I was involved in programming very early forms of AI. It was just datasets with basically if/then decision trees. Basically  a questionnaire with a list of possible responses based on user choices. That then develops to expanding the datasets based on responses entered. This still needed human intervention to make sure the data in the datasets were valid. Still based on preprogrammed if/then decision trees which can be added to based on new data gathered. It is all manageable and simple to this point. But modern AI systems not only write their own datasets they now write their own decision trees (theoretically based on decision trees parameters set by humans) but these become so large so quickly that regardless of how well a company claims they manage a self learning AI system at some point it becomes impossible and impossible to find where the route of a potential problem began, they are just far too vast for even the largest team of developers to manage. So I would not trust an AI to be able to, as you say easily recognise subtle changes or idiosyncrasies in human behaviour or stick to any type of parameters that would try to ensure this. 

I suppose what I mean is that any computer system works on logic.

5+5=10, 20-4=16 but if it comes to 10+turnip=? What will it do?

I just told Co-pilot the joke and it did recognse  it as a joke and a very British one at that, but we had been talking about absurdist British humour yesterday and I think it just picked up on it.

Do you think so? So many computer programs have problems and biases that are quite subtle depending on who's training/programing them. Even things like facial recognition programs are less likely to be acurate for non white or South East Asian populations, we're unconciously building these biases in. In this case though I wonder what it says about the people who are giving these programs their learning material?

Do you see the AI diagnosing autism as learning on the job as it were? If so then what about the people it missed whilst gaining experience, I know this can happen with newly qualified humans too, but would the AI have the wit to ask a colegue and would it have the sort of professional supervision that many humans have to have?

I agree that computors are better at sorting through large amounts of data faster than humans can and I can see that for medical screening, of things like mammograms they're faster and just as or even more acurate than humans. But I think that is different to  how we experience the world and so much of autism is about our experiences in the world. A human for example could ask for an example of being humiliated at school and have an empathetic response, but how does a machine learn that response? I think a good example of this comes from when I used a joke whilst I was talking to AI and it took me seriously and came out with scads of serious stuff about this joke and wasn't even sure when I told it it was a joke comment and not to be taken seriously and this is with something that knows my conversational style. I can't imagine how a machine that didn't know my conversational style would even hope to pick up on the subtlties of human communication in a diagnostic situation.

Maybe as an experiement we could tell our AI's the same joke, not tell them we're telling them a joke, but just to see how it responds, we'd have to tell it the same joke, so heres mine.

Cuthbert was an inflatable boy, who went to an inflatable shool and one day the head master came up to him and said 'Cuthbert!, You've been found running in the school corridors with a drawaing pin and I want you to know that not only have you let me and the school down, but you've let yourself down too',

Yes you are more skeptical on these matters than I TheCatWoman.

Broadly speaking you ask how would the AI learn to understand differences - I would answer that question by saying that it would learn to.  I agree with the core rule "garbage, in garbage out" for the use of any system, computer or tool (hehe except that as is also said "one man's junk is another man's treasure").

Computers are in my opinion far better at handling vast amounts of data and complexity than human beings - yes they can be biased of course but that bias in my opinion is easier to resolve in them than in most humans

The childhood form you mention is interesting. I was sent a form, but had no one to complete it which they accepted and I did in my meeting mention things I remember. I was glad I didn't have a parent to answer that, as my last surviving parent remembered very little of my childhood and probably would have said no difficulties.