Monday, November 06, 2017

47% of jobs will be automated... oh yeah...10 reasons why they won’t….

I’ve lost count of the times I’ve seen this mentioned in newspapers, articles and conference slides. It is from a 2013 paper by Frey and Osborne. First, it refers only to the US, and only states that such jobs are under threat. Dig a little deeper and you find that it is a rather speculative piece of work. AI is an ‘idiot savant’, very smart on specific tasks but very stupid and prone to massive error when it goes beyond its narrow domain. This paper errs on the idiot side.
They looked at 702 job types then, interestingly, used AI itself (machine learning) which they trained with 70 jobs, judged by humans as being at risk of automation or not. They then trained a ‘classifier’ or software program with this data, to predict the probability of the other 632 jobs in being automated. You can already see the weaknesses. First the human trained data set – get this wrong and it sweeps through the much larger AI generated conclusions. Second, the classifier, even if it is out by a little can make wildly wrong conclusions. The study itself, largely automated by AI, rather than being a credible forecast, is more useful as a study of what can go wrong in AI. Many other similar reports  company in the market parrot these results. To be fair, some are more fine-grained than the Frey and Osborne paper but most suffer from the same basic flaws.
Flaw 1: Human fears trumps tech
The great flaw is over-egging the headline. The fact that 47% of jobs may be automated makes a great headline but is a lousy piece of analysis. Change does not happen this way. In many jobs the context or culture means that complete automation will not happen quickly. There are human fears and expectations that demand the presence of humans in the workplace. We can automate cars, even airplanes, but it will be a long time before airplanes will fly across the Atlantic with several hundred passengers and no pilot. There are human perceptions that, even if irrational, have to be overcome. We may have automated waiters that trolley food to your table but the expectation that a real person will deliver the food and engage with you is all too real. 

Flaw 2: Institutional inertia trumps tech
Organisations grow around people and are run by people. These people build systems, processes, budget plans and funding processes that do not necessarily quickly lead to productivity gains through automation. They often protect people, products and processes that put a brake on automation. Most organisations have an ecosystem that makes change difficult – poor forecasting, no room for innovation, arcane procurement and sclerotic regulations. This all militates against innovative change. Even when faced with something that saves a huge amount of time and cost, there is a tendency to stick to existing practice. As Upton Sinclair said, “It is difficult to get a man to understand something, when his salary depends on his not understanding it.”
Flaw 3: Low labour costs
What is often forgotten in such analyses is the business case and labour supply context. Automation will not happen where the investment cost is higher than hiring human labour, and is less likely to occur where labour supply is high and wages low. We have seen this recently, in countries such as the UK, where the low-cost labour supply through immigration has been high, making the business case for innovation and automation low. Many jobs could be automates but the lack of investment money, availability of cheap labour and low wages makes the human bar quite low. There are complex economic decision chains at work here that slow down automation.
Flaw 4: Hyperbole around Robots
Another flaw is the hyperbole around ‘robots'. Most AI does not need to be embedded in a humanoid form. Self-driving cars do not need robot drivers, vacuum cleaners do not need humanoid robots pushing them around. Most AI is invisible, online or embedded onine or in the electronics of a device. As Toby Walsh rightly says, when he eviscerates certain parts of the Frey and Osborne report, there’s no way robots will be cutting your hair or serving your food by weaving through busy restaurants with several plates of food, any time soon. The ‘Reductive Robot Fallacy’, is the anthropomorphic tendency to equate AI with robots along with the idea that robot technology has to look like us and do things the way humans do them. The vast majority of robots, AI-driven machines that perform a useful function, do not look like humans, many are online and almost invisible.

Flaw 5: Hyperbole around AI
AI is an idiot savant. It is incredibly smart at specific things in specific domains but profoundly stupid at flexible and general tasks. This is why entire jobs are rarely eliminated through automation, except for very narrow, routine jobs, like warehouse picking and packaging, spray painting a car and so on. Accountants use spreadsheets, restaurants use dishwashers, mixers and microwaves. Most automation is partial, as the general worker still outfoxes AI. There are severe limitation to AI in many fields, not least the sheer amount of processing power needed to fuel the applications as well as limitations in the maths itself. There is also a great deal of hype around the 'cognitive' capabilities of AI, led I suspect by that misleading word 'intelligence'. AI is not conscious and has little in the way of cognitive skills. It may win at GO but it doesn;t know it has won.
Flaw 6: Garbage-in, garbage-out
This common flaw, as Walsh rightly spotted, was that the training data in the Frey and Osborne paper was either a 0 or 1 probability of automation but the outputs were between 0 and 1. This is an example, not so much as garbage in-garbage out, as binary-in range-out. You can see this manifest itself in some absurd predictions around jobs that are unlikely to be automated, as well as underestimates in others, like hairdressing, waiters and cleaners. Beware of AI generated predictions.
Flaw 7: Heuristics
The process of automation in employment is a messy business with many variables. Heuristics can help here. First we can categorise jobs first as: Cognitive v Manual; then… Cognitive routine, Manual routine, Cognitive non-routine, Manual non-routine. But even the distinction between manual and cognitive is not mutually exclusive. Few manual jobs require no knowledge, planning or problem solving. These can be useful rules of thumb but the world rarely falls neatly into these binary of four-way categories. Yet they often lie at the heart of predictive analysis. Beware of simplistic heuristics.
Flaw 8: Human bias
Bias in analysis is all too common. Take just one axemple; the analysis of education. The people doing the analysis are often academics or people who have an academic bent. The Frey and Osborne paper conflates education into one group, as if kindergarden work was the same as academic research. The routine aspects of education, the fact that most teachers, trainers and lecturers do a lot of admin and work that is actually routine and repetitive, is conveniently ignored. Google, Wikipedia and online management and learning has already eaten into the employment of librarians and teachers. It is a displacement industry. Take one service – Google. As the task of finding things became super-fast, the process of learning, research and teaching became quicker. Library footfall falls, as we no longer have to troop off to the library to get the information. Amazon has commoditised the purchase of books. Commoditisation is what technology is good at and what Marx recognised as a driving force in market economies. Educators don’t like to hear this but they have a lot to gain here. Teaching is a means to and end not an end in itself. It has and will continue to be automated, not by robots but by smart, personalised, online learning.
Flaw 9: Activities get automated, not jobs
In truth most jobs will be partially automated. This has been going on for centuries with technological advances. Sure, horse grooms and carriage drivers no longer exist but car mechanics and taxi drivers do. Typesetters have been replaced by web designers. ATMs have simply changed the nature of bank tellers, not completely automated the process. Indeed, in many professions the shift has been towards more customer service and less mechanical service. What matters is not necessary the crude measure of ‘jobs’ being automated but rather activities’ being automated. By activities, we mean specific tasks, competences and skills. 
Flaw 10: New jobs
65% of today’s students will be employed in jobs that don’t exist yet.” This is the sort of exaggeration that feeds bad consultancy. Most will be doing jobs that have existed for some time. Many will simply be doing jobs they didn’t plan on doing (and don’t like) or jobs that have changed somewhat through automation. Predicting which jobs or activities get automated is easy compared to predicting what new jobs will be created. The net total is therefore difficult to establish. Fewer people may be needed in certain areas but new jobs will be created, especially in services.
Conclusion
To be fair, more recent analyses have moved on to more fine grained concepts and data. McKinsey did a detailed analysis of 2,000-plus work activities in 800 occupations, with data from the US Bureau of Labor Statistics and O*Net. They quantified the time spent on these activities and the technical feasibility of automating them. NESTA dis a breakdown of specific skills.
The crude headlines will continue but we’re starting to see more detailed and realistic analysis that will lead to better predictions. This is important, as educational bodies need to be able to adapt to what they will be required to teach as well as what they teach and to whom. As the change accelerates, education and training will need to be more sensitive and adaptive to the changes. This means more accurate predicting of demand and quick adjustments in supply. I’d go for around half of the Oxford figures with the caveat that more service jobs will be created, so that the net total will be 10-20%. There will be no sudden shift in months but a gradual bite by bite into activities within jobs. This is the field I work in, invest in, write and talk about (see WildFire), so I'm not coming at this from the sceptic point of view. AI will change the world and the world of learning but not in the way we think it will.

 Subscribe to RSS

1 Comments:

Anonymous David@knowledgestar.com said...

Great analysis Don. Not sure it accounts for the seismic shift from labor-intensive to mind-intensive work. I'll think about it some more.Thanks!

4:34 PM  

Post a Comment

<< Home