Time and The (AI) Machine

The trial of 'Humphrey', a tailored version of ChatGPT for the Civil Service, and the trumpeting of its results in the media, have led me to think a lot about time and efficiency. Now I am a philosopher, not an economist, so there is a risk that what follows is completely misguided. In particular, at least until the very end, I have avoided talking about efficiency in terms of cost, where that is measured by market price. My reason for this is that market prices often mask true costs which have been externalised - we have seen this with environmental harms and forms of financial risk (‘too big to fail’) where the market price has factored in the fact that governments will pick up these costs. Instead I have followed the initiative of Ivan Illich in Tools for Conviviality¹ and considered energy usage is the best measure of true cost when considering efficiency. It is after all intuitive that it is not more efficient to do something slightly more quickly with many times more effort. This may not ultimately be a good approach, but it seems to me to be a really important conversation to have. This blog doesn't have comments, so if you would like to enter into a conversation, either in public or in private, please let me know.

Time, Energy, Productivity

Some machines genuinely save time. The bicycle is a good example: if you expend the same amount of energy cycling from A to B as you would have done walking, you will get there quicker. The washing machine is another: it uses similar amounts of water and energy for heating that water as handwashing clothes used to, but uses a lot less energy for the mechanical action (hand-washing clothes is hard physical labour). So even if a machine-wash cycle isn't a lot quicker than a hand-wash, we can use the difference in energy to do something else while it is running.

In general, then, if a machine does the job of a person more quickly for the same energy expenditure, or in about the same time for less energy expenditure, it overall saves time. If it takes the same time and uses the same energy, then it allows more to be done in a specified time period, but not without added energy usage. That would be an improvement if time - as in working time or labour - is scarce but energy is relatively plentiful. Hence cheap fossil fuels powered the industrial revolution by increasing the productivity of a fixed number of workers. In contrast, if work time is plentiful and energy scarce, you can make the reverse trade: the pyramids were built with human labour with no additional energy costs than the food they ate.

So we can formulate a general thesis about productivity and marginal energy use:

If, but only if, a machine uses less energy in the same time or less time for the same energy to complete a task as a human would, then it allows greater productivity.

Machines that do not allow greater productivity may still be valuable. They could allow tasks to be completed when there is not enough human labour. But with 8 billion humans on the planet, there is plenty of labour, if only we set our minds to co-ordinating the tasks and the people. Alternatively if the energy is abundant, machines might be used to allow the humans more leisure. But leisure without basic needs being met is not a good thing, so that would only really help if those who owned the machines paid the others not to work. For all their hints at Universal Basic Income, we can be pretty sure that the Broligarchy are not serious about being taxed more to pay people not to work.

Does AI improve productivity?

So the fundamental question of value we need to ask is whether AI in the workplace uses less energy in the same time or less time for the same energy to complete a task as a human would.

In what follows, I will focus on the use of genAI in general office work, where it is being pushed as a time saver. For example, a recent trial in the UK Civil Service concluded it could save an average of 26 minutes a day. Is that actually an improvement? After all, we are often told that AI can save teachers’ and doctors’ time, but so would having more teachers and doctors, thereby requiring each to teach/treat fewer pupils/patients. If there really is a shortage of labour in those professions (rather than a shortage of people willing to work in those atrocious conditions which would be resolved by improving the conditions by employing more people to do the same overall work), then it might be the case that energy is more abundant than labour.

But that doesn't seem to be the case for the sorts of office job where AI is being touted as an efficiency. So let's do some sums. Please note: there is no sense in which these numbers are accurate or comprehensive, the idea is to get a very rough estimate, continually making assumptions in favour of the machine, and see where we come out.

How much energy does an office worker use to write an email?

First assumption: 2000 kcal a day. How does that translate into labour time? Well, second assumption: if someone works 8 hours a day, the other 16 hours are necessary for that work to be possible. This assumption is massively in favour of the machine, of course, but that is what we are trying to do. But it does have the advantage of taking into account - when we think about cost - that workers are paid to live not just the price of their food.

With these two assumptions, we get an energy use per minute of an office worker of (2000/8/60=) 4.16 kcal. (Since the idea is not to replace the worker but 'save time' for them, I am not including the costs of the office building, heating, IT etc. That would have to be factored in with a net reduction of workforce.)

Third assumption: it takes an office worker 10 minutes to write a 100 word email. Again, this is not intended to be precise but a rough ballpark, erring on the side of too long to again favour the machine. So that is ~41 kcal, or ~47 wh.

How much energy does it take genAI to write an email?

Now, how do we compare this to genAI? We need a base figure like our 2000 kcal a day for a person. For that I will take the claim that ChatGPT uses 140wh to write a 100 word email. This is probably not accurate, but it is also likely not totally inaccurate. So it will be my fourth assumption. Note that this doesn't include the energy used to train the model, so it is again weighted in favour of the machine.

In other words, the human is around 3x more energy efficient.

But what about time?

genAI is much quicker than a human in writing that email, probably taking considerably less than a minute. Now if this 10x time-saving was achieved with the same energy use, that would be a clear increase in productivity. But it is not.

Let's make a fifth assumption in favour of the genAI: it saves 10 minutes of human time by writing that email. But it uses almost 30 minutes of human energy. Is that a productivity gain?

The genAI tool does the task so quickly because it breaks it down into 1000s of sub-tasks which can be run at once. Computationally, this is called 'Parallel Distributed Processing', and it is the key property of a Neural Net, as we call these AIs now. What makes them so clever, of course, is that they break these tasks down in ways we do not understand through a process we call 'Deep Learning'. The details don't matter for present purposes, but it allows us a better way to understand how to evaluate the speed of the AI compared to humans completing the same task.

The question we don't know the answer to, and can't know, is: If we could break down the task of writing a 100 word email into 30 sub-tasks which could be performed in parallel by 30 humans, how long would it take? While we can't know the answer to this because we cannot break down the task into 30 sub-tasks which could be performed in parallel by humans, we do know (well, it is one of our assumptions) that a single human could do it in 10 minutes. So it is not unreasonable to guess that - in this impossible scenario - it would take 30 humans just 20 seconds.

I have made several assumptions, each of which could be questioned, but along the way I have tried to favour the machine in how I made those assumptions. From those assumptions we can conclude that there is one clear sense in which humans are faster at writing short emails than genAI. We think the opposite because we compare the effort of a single human with the effort of a single AI, rather than comparing the effort in each case, which is hidden from our view in the case of the AI. When we measure effort by energy consumption and take into account the possibility of parallel distributed (human) processing, we get a very different perspective.

How to decide whether to implement genAI

If the argument above is even vaguely close to the right conclusion, you should use genAI in office work if and only if labour is scarce and energy abundant. How does this play out for different values of 'you'?

For an individual business or corporation, 'scarce' and 'abundant' here mean 'expensive' and 'cheap'. Now most such entities regard the cost of office workers as (very) expensive, especially compared to the energy costs of genAI, which they don't pay directly and which are often subsidised (partly because the genAI providers are selling below cost to gain market share, partly because those providers often get hidden subsidies from government to build data centres which are the main energy consumption points, and partly because in most countries, the energy infrastructure is subsidised by government).

However, for nation states, the calculation is very different. Most have an ample labour supply for the sort of office work we are talking about using genAI for. And most of the countries where this is likely to happen have issues of adequacy of their energy supply. The UK meets both those conditions: office labour (for writing emails!) is abundant² and energy is scarce. From this macro perspective, where 'you' is a government, it seems that the right thing to do is change the cost structure for businesses: make it cheaper to employ office workers and make the energy used by genAI more expensive. That will improve national productivity, by getting the work done by the most energy efficient machine for the job - a human.³

Of course, the current UK government has done exactly the opposite, by increasing the cost of employing people, subsidising energy, and reducing the commercial barriers to building data centres.