So what jobs are AI proof then?

[StrongJamie]

Based on recent media report we are facing an imminent (?) AI revolution.

I imagine that doctors, teachers, lecturers, accountants and civil servants can easily be replaced by AI and tailored professional AI software. I am guessing that jobs that require intricate physical handling are less at risk as it would be expense to mass produce the Hardware. Hospital doctors are more at risk than ward nurses but less than GPs who soon will be obsolete.

It will be a good while until they can mass produce robots that do humans jobs, which require a lot of running around and haptic skill but some jobs don't need a person, they just need the right software (e.g. GPs).

I imagine it like this, you log onto your GP AI service, they know all your medical history and also all the up to date epidemiological data of your neighbourhood as well as your biomarkers, pulse, heart rate etc uploaded via your smart watch continuously. The system knows about all possible diseases and conditions and based on your biomarkers and symptoms knows how to signpost you for further tests or what to prescribe. Job done, no more GP.

Teachers? No need. Ai robots, virtual or physical deliver synchronous teaching the rest is done online.

Please pick holes in my assumptions or add to the list of soon to be obsolete professions.

Which ones are ai bullet proof?

QED, but this applies to all humans.

Also, LLM based transformers like OpenAI's GPT series all can suffer hallucinations too, so, yeah.

Unconscious bias of the coders will impact output.
see self driving cars or facial recognition not recognising black and brown faces as human.

I agree though that practical jobs will be last to go.

How much if the nuanced approach will we miss from eg GPs

Social care

Primary teaching is primarily about relationships, whether that's encouraging children to work or to behave. I just don't know how you replicate that.

AI can be massively biased, it all depends on whether there are biases in the training data. Generally AI takes human bias and amplifies it.

Mind you I’m still waiting for the paperless office!

It depends on the kind of model and what has been implemented in the training process e.g. adversarial training and fairness aware ML protocols. If a dataset is generally good, but the model hasn't been checked to see how it reasons, then you still get garbage out like a bad dataset.

Just being able to assess what caused a particular output can improve results measurably as opposed to black boxing the whole thing and assuming it's doing what you intended.

I’m an accountant and we are slowly moving towards more and more ‘through processing’ with the result that our finance team will soon be smaller but more highly skilled. AI can’t yet join the dots to do the analysis for all it can guess the answers it’s often plain wrong. I don’t think that all accounting jobs will go at all but I do think there will be fewer. (Or new ones…there’s an ever increasing appetite that everything public should be audited for example).

I agree with some of the other posters' points about art. Art is generally a luxury item (if buying an original)) and part of the allure is that a human's thoughts and originality have gone in to it. And in the case of paintings it is a physical object with traces of the person in the form of brush strokes, not just a flat printout of an AI image.

I've used a visual AI fairly extensively, as a basis for images for me to paint. The images are impressive but they always need extensive tweaking when I paint them - changing the design structure, resolving parts of the image that make no sense (they often look coherent at first glance but fall apart when you look closer), changing colours, adding elements that weren't originally there. Essentially putting your own mark on it. So it's great as a prompt, but needs a human with skill to raise it beyond something soulless.

It’s on channel 4 news now about racial bias in AI. Even to soap dispensers!

I reckon a virtual AI councillor delivering CBT to individuals could work very well and doesn't seem far fetched. As humans we respond to speech, voice, sound, and it should be straightforward to design a virtual persona, which can of course be easily personalised to suit different clients. It's very scalable. The 'human touch' can be virtually replicated and our human brains, nervous system etc will respond in similar ways as to real human, we evolved to do so.

I also thought chefs are safe as are hairdressers, beauticians, nursery nurses, infant school teachers, vets (tricky for a robot to deal with a wriggly pet to give it a vaccine), carers, nurses, brick layers, plasterers, sports coaches.

How about apothecaries?

What about the step beyond helpful AI to full on AGI that has alignment problems and decides to turn us all into paper clips?

Civil servants? You do realise they're not one homogeneous lump that all do the same job?

I was a bit shocked at how well ChatGPT wrote. But I felt less unsettled the more I talked to it. Its understanding is really superficial. Better than a 14yr old writing an essay, but not good for the executive functions that adults need to deploy in how we get things done. I just know in my area that the expertise required is far beyond simple essay writing.

Imagine ... you log onto your GP AI service, they know all your medical history

The GP doesn't really know it now, the records are patchy & not updated consistently & get coded in diverse ways, would be even more diverse with different algorithms applied at different times. Patients still supply a lot of info, at an appt, and often it's what they didn't say rather than what they did say that matters. Also, people don't understand their conditions: "I had a little heart failure but that's sorted now." Humans can guide a patient back to better health management. AI can't do it yet.

and also all the up to date epidemiological data of your neighbourhood

it won't be up to date, trust me.

as well as your biomarkers, pulse, heart rate etc uploaded via your smart watch continuously.

if you kept watch on, kept it charged, didn't break it, didn't lose it, was compatible with your phone, sync'd it, kept the App updated... And supposing the cheap wrist devices become as accurate as full chest monitors.

The system knows about all possible diseases and conditions and based on your biomarkers and symptoms

The system won't know, not least because the algorithms will only be Bayesian not definitive.

knows how to signpost you for further tests or what to prescribe.

How will the system physical examine someone? Draw blood, photograph a funny spot, take a cervical sample? Understand unusual accents, or realise the patient is illiterate? Helping with polypharma would be good, though we're back to Bayesian probabilities to try to find best mix of suggested drugs.

I advise on welfare benefits. There are already tools to check entitlement online and explain processes. Quite possibly a lot of my job could go. The bit that may be missed is the person that is very distressed and splitting up or being made redundant. AI needs to recognise the distress. No doubt improved AI counselling and my job would go.

This one will also be subject to input-
what disability benefit can I claim when I get my state pension?
both the following are accurate
“Attendance allowance - here is how to claim” or
“attendance allowance but if you have problems walking claim PIP before that - you may get more money”.

I am pretty poor with technology but very keen on improving efficiency in my school... How would you use it for lesson plans and other teacher paperwork?

Hahahahahaha

If you think AI will replace teachers then clearly you have never tried to trach a child anything. I’m not a teacher but wow trying to get my kid to sit at a screen during lockdowns was no fun and also very unhealthy for all learning to be sitting. Chilsren learn best when moving around for much of the lesson.

Anyway. Ignoring your ignorance about teaching, if you want a 100% AI proof job I suggest plumbing, building etc.

I guess professional sports people are pretty safe because when people watch sport the whole point is that they want to see physical achievements by humans, not (say) one robot programmed to score goals and another robot programmed to save them.

People who write insurance proposals

Too many variables

I'd think embalmer (and the related roles) is pretty safe for a while yet.

But then you have instances like this. A lot of the problems people have with LLMs so far seem to be poor prompting. Yes, they’re not trained n everything or always up-to-date, however, GPT-4 is already running rings around 3.5 on a lot of test parameters, and I don’t see any slowdown happening any time soon. The largest and latest build of GPT-4 has nearly ten times the token number of prior iterations and enables multi-modality to allow it to look at and assess imagery. We already have GP services like Babylon basically use video calls and text chat for triage of treatment, might as well utilise more advanced agents for this and leave the limited human consultants to deal with the meatspace stuff requiring biopsies or personal care.

Honestly, I was predicting an AI winter this time last year. Companies were shutting down driverless car labs, even Tesla, and various other developments were quite niche such as AlphaFold. There didn’t seem to be any budging on a true AGI date happening within decades. Now all that needs to be reassessed, especially as GPT-4 has been showing signs of being able to reformulate and improve responses and planning in a recursive learning way.

"as well as your biomarkers, pulse, heart rate etc uploaded via your smart watch continuously"

This totally misunderstands the section of the population that need most healthcare - the elderly. The technology would need to improve massively. Smartwatches are fine for younger people. As you get older people lose fine motor control, and with illness get confused more easily and start to struggle with the skills needed to update smartwatches.

The heart monitors patients who have had heart problems wear are paid for by a private company (although patients get it free on the NHS) because they sign to say the private company can use their data collected through the monitor for commercial purposes. They are expensive. They are also updated by the hospital. All the patient has to do is wear it. If it gets broken the hospital replaces or gets it fixed. That is what would be needed, not a smartwatch. And young healthy patients expecting the AI GP to look at the results of their smartwatch data is just a total waste of everyone's time.

You’d have to make sure it strictly adhered to ethical standards. Eg if leaking info to the papers would help its client or fabricating evidence, or destroying/ lying about evidence … ai without very clever boundaries would just do that. The legal system can’t take that risk.

AI will advance. But it is massively oversold at the moment. We still don't have a tool that can do an accurate transcription of speeches being given. A relatively simple task compared to many jobs and yet the tools still get a lot wrong.