I came across an academic article yesterday that was published in March. Entitled 'The AI Layoff Trap', it was authored by Brett Hemenway Falk and Gerry Tsoukalas, who are both based at well-respected US business schools. I should note that this, in part, motivated a post already made this morning on the cost of AI investment. The two posts are linked as a result.
This is the abstract from the paper, written in typical academic style:
If AI displaces human workers faster than the economy can reabsorb them, it risks eroding the very consumer demand firms depend on. We show that knowing this is not enough for firms to stop it. In a competitive task-based model, demand externalities trap rational firms in an automation arms race, displacing workers well beyond what is collectively optimal. The resulting loss harms both workers and firm owners. More competition and “better” AI amplify the excess; wage adjustments and free entry cannot eliminate it. Neither can capital income taxes, worker equity participation, universal basic income, upskilling, or Coasian bargaining. Only a Pigouvian automation tax can. The results suggest that policy should address not only the aftermath of AI labor displacement but also the competitive incentives that drive it.
So what does that mean? The argument goes something like this, and I have ignored all the maths that underpins and focus on the claims made instead.
First, the claim is that AI-driven automation creates what might be called a demand externality. In other words, as firms replace workers with machines, they reduce wage income, but because wages are the primary driver of consumption, this weakens aggregate demand in the economy as a result. Critically, though, no individual firm bears the full cost of that lost demand: instead, it is dispersed across the entire economy. Each firm, therefore, ignores at least a part of the damage it creates, blaming others for it instead.
Second, this then leads to a predictable outcome. Firms think that they are acting rationally in their own interests, but collectively they generate what the authors describe as a "Nash equilibrium of over-automation" (which I admit is a new term to me). In effect, they behave as if they are trapped in a classic Prisoner's Dilemma model. Each of them has an incentive to automate to protect their profit margins and market share, even though all would be better off if automation were slower. The result is excessive worker displacement, reduced total profits, and lower returns for both labour and capital.
Third, competition actually makes this problem worse, not better. The more fragmented the market, the less any one firm internalises the demand loss it creates. As a result, competitive pressure accelerates automation beyond the socially optimal level. In this case, market forces amplify inefficiency rather than correct it.
Fourth, the paper tests a range of commonly proposed responses and finds most of them inadequate. The paper found that policies such as universal basic income, upskilling, or additional capital taxation may cushion the effects of automation, but they do not address the underlying externality it creates. They might boost demand or redistribute income, but they do not change the incentive to automate excessively. Even profit-sharing and worker participation schemes cannot fully resolve the issue, as the externality is systemic and diffuse.
Fifth, the only policy that directly targets the problem is a Pigouvian tax on automation (click the link for a new glossary entry explaining what a Pigouvian tax is). The authors claim that taxing each act of labour displacement (or redundancy) at its social cost would align private incentives with collective welfare. In principle, this could reduce automation to its socially optimal level. The tax revenue could then be recycled to support incomes, retraining, or public services, reinforcing demand.
Sixth, the paper then introduces an important qualification. If displaced workers are rapidly reabsorbed into higher-paid roles, which the author's model captures with a parameter that determines how income is recycled into demand, the externality can diminish or even reverse. In that case, automation may be welfare-enhancing. However, the model's evidence suggests that this condition is demanding and unlikely to hold in most real-world labour markets, especially in the short to medium term.
The overall conclusion is clear. Left to its own devices, an AI-driven economy is likely to over-automate, destroying demand and reducing well-being. Market mechanisms will not, in most circumstances, self-correct this outcome. In fact, they intensify it.
If that diagnosis is right, the message is clear and is that policy on AI has to move beyond attempts to cushion the effects of automation and must instead confront its causes.
The implication is straightforward, even if politically challenging. If we want an economy in which technological progress improves well-being rather than undermines it, then we cannot leave the pace and direction of automation solely to private decision-making.
Link this issue to that on climate change, and it is apparent that the current AI euphoria is widely misplaced; governments are wrong to think that AI is a solution to any of their problems, and the "tech bros" driving AI automation are not the source of our salvation but are much more likely to be the source of our destruction.
The question then is, who will say that?
Thanks for reading this post.
You can share this post on social media of your choice by clicking these icons:
There are links to this blog's glossary in the above post that explain technical terms used in it. Follow them for more explanations.
You can subscribe to this blog's daily email here.
And if you would like to support this blog you can, here:
Buy me a coffee!
Might it be that an unstated, but operating, basic “economic” plan is to make as many people unemployed as possible in a World which is increasingly uninhabitable?
”Sometimes I wonder whether the world is being run by smart people who are having us on or by imbeciles who really mean it?” (From Laurence J. Peter).com
Thank you for sharing this – I’m surrounded at work by people unquestioningly in love with AI.
Overall, AI is gaining ground because it addresses a fundamental worry about the cost of caring – there is not enough money apparently to go around to care and do it properly (except the huge sums being invested in the supply side of AI you also mention this morning). But on the demand side, ‘lack of money’ the need to make savings etc., rules and drives the AI forward.
Bluntly, it is this obsession with the ‘lack of money’ – something that is totally artificial and ignores the facts of money creation that needs to be addressed and when it is, maybe the right questions about AI will be asked.
Nothing interfaces better with humans than other human beings (intra-species). We should view AI as a different species in terms of an interface, but the problem is we are taught to see it as an equivalent solution not (and I mean at best) a complementary one. We rely too heavily on the infallibility of those who programme it as well.
Looks an interesting article. Presumably the argument could apply to any type of technical progress which increases productivity of labour and enables firms to produce more with less labour?<p>
There have been periods of mass unemployment in the past – some driven by demand-suppressing government policy – and then mitigated ex-post by the New Deal and Keynes inspired public spending .<p>
Presumably this ‘Pigouvian tax’ could have been imposed on all the oil companies to mitigate the disastrous externalities they have imposed on us and the planet.?
I think the traditional term for this is a “tragedy of the commons”
Except that the real tragedy of the commons was that they were appropriated by capitalists. Up to that point, they had been managed pretty well by the people who lived off them. In some parts of the world, they still are.
See Elinor Ostrom: https://en.wikipedia.org/wiki/Elinor_Ostrom
Agreed
I thought of including her in the economists list
This, of course, assumes that “AI” can do the things claimed. The hundreds of billions invested in AI (by which I mean large language models or LLMs) suggests that the “Tech Bros” think this is the case. I beg to differ. I don’t think LLMs can do anything like as much as is claimed. They are, essentially, word completion/prediction on steroids.
LLMs are, undoubtedly useful. I have found them very helpful for a number of tasks including programming. But I doubt they will engender the sea change envisaged. I think they are more likely to have similar effect to word processors, spreadsheets, grammar checkers etc. Helpful, yes. Displacing some labour, yes. Paradigm changing, not so much.
Previously we have had such “innovations” as the Metaverse (remember that), and quantum computing, even the dot com boom. There is something in all of these – but not nearly as much as proponents claim.
Dear Tim,
I think LLM-s are just the entry level AI what we can understand somewhat. The issue is not that they will talk us out from the workforce. I think the issue is the pattern recognition which could them competitive in a lot of other job areas than costumer service.
If I think about, we humans do the same, just recognising patterns most of the time and acting accordingly. The only difference we can understand things, the current AI not.
Language might have been the easiest to crack for a computer because of the available data found on the internet.
Tim/Limerin – you have touched upon something that I have noticed about even my work related AI solutions: even those seem to be based on extractive, marketing, surveillance type algorithms. Sharepoint (Sh**point as I call it) will always tell you what everyone else is looking at as you wade through it to find what you want. And it too seems to list things that you want in some sort of ‘market order’, so you have to process what you want out of a choice – too much choice. Sharepoint is also monopolistic in nature – you have to use it regularly – dare I say ‘exclusively’ to get the most of it – in other words it is built on repetition by domination and knowledge of the user – working through documents is one thing; working through emails and pdfs with it is another story – it is very slow at processing these and is no good for example as an audit file – it would take hours to get through it.
I think a really useful AI would be one not based on surveillance and marketing – that would need to be a totally fresh start in my view.
Amazing how much twaddle can be written by so called ‘professors and consultants.
The question is will corporation and institutions do away with people if ai and automations make it cheaper and super productive for them? The answer to that is YES, to hell with the consequences.
The powers that be are indifferent to others and this Ties up with the articles earlier in the week…
Dismantle social security before the Ai/automation employment massacre and it’s tough luck when it happens as ‘we can’t afford to help people’
Civil disorder? Not with digital ID and no jury trails. That won’t be able to build pensions quick enough. They won’t be jobs in prisons either as it will all be robots!
Once enough people begin to realise that software has made them surplus to requirement, i suspect there will be a rebirth of the luddites…
* would like to add fro my previous post that I am not being dismissive of the work that these gentlemen have done BUT a lot of prosaic rumination is being done here when, like climate change, we are lost in discussion that delays action.
Action that needs to be taken, urgently.
It’s hard to imagine a practical way to implement a Pigouvian tax based on how many people a company has not employed. Good luck in calculating that! Maybe a metric based on wage-bill (excluding the top tier?) vs turnover could work?
I don’t think it is possible. This was an academic exercise, literally.
What happened to the sharing buttons please?
I don’t know.
I still have them.
This might be one of those “have you rebooted lately?” moments
This article is about Chat GPT, so perhaps is not that closely related to what your post is about. But I do think it shows just how little the tech bros understand what they are creating in the mad rush to get ahead of competitors.
https://openai.com/index/where-the-goblins-came-from/
[…] This felt particularly relevant in the light of the posts I made yesterday on AI and its cost, and the implications it might have for employment. […]
Fascinating and great synopsis. Somewhat worrying given the total lack of preparation policy makers and legislators have for the incoming storm. As a software engineer and startup co-founder with a team using advanced AI every day (more specifically generative transformer models / LLMs) I can tell you that the capabilities of these models are already well beyond those of even the most talented junior engineers. Comparing this to previous technology advancements maybe is missing a crucial difference. The people at Google, Anthropic and Open AI who are writing these tools, are themselves using the same tools for improving the capabilities of the tools, creating a compounding effect and exponential growth in capabilities. I’m not sure if previous technological advancements had this same dynamic – or at least not as quickly reinforced so I’d be cautious to draw too many parallels to how economies absorbed this kind of change before.
You explain precisely why these models can blow up
[…] AI automation will not be our salvation […]
[…] to do that, but might, in fact, be wholly destructive, both of our climate and of our economy, by destroying employment within it, alongside the whole structure of our society as a […]