In 2015, I wrote an essay entitled, “Will
robots replace human labour and reduce real wage levels?” That essay still
seems to attract quite a lot of attention even though some of the thoughts in it
are now out of date. I did not anticipate the impact of Large Language Models (LLMs).
One of the things I like about my previous essay is the use of
down-to-earth examples to illustrate processes. So, I will try to continue in
that vein.
Horses, tractors and robots
My previous essay included some paragraphs, quoted
from Nick Rowe, illustrating how the replacement of human labour by
non-human labour is an ongoing process rather than a new phenomenon:
“Horses were once like robots. Horses could do a lot of the
same work that humans could do. Humans and horses can pull things, if you feed
them. But then mechanical horses, called tractors, were invented, that could
pull heavier things with cheaper food. Tractors pushed horses' wages below
subsistence, so the horse population declined.
The robot horse displaced horses, just as horses displaced
humans from all the jobs where humans pulled things. But humans, unlike horses,
can do lots of other jobs beside pulling things. Humans are very versatile.
Horses can't really do anything except pull things. So humans switched to doing
other jobs, while horses couldn't. And the marginal product of labour, and
hence wages in those other jobs, increased. Horses and tractors were
complementary factors to human labour in those other jobs.
But that won't happen if robots are invented that really are
just like humans, and can do all the jobs that humans can do. Robots that are
just like humans would be just like slaves, rather than like tractors and
horses.”
My comment in 2015:
What we are seeing now is robots that are displacing humans
from a range of activities and freeing them to do things that robots can’t do -
just as horses did. There are adjustment problems for people in the affected
industries, but the impact on average real wages is likely to be positive. Over
time, superior robots are likely to be invented that will replace the initial
series of robots, just as tractors displaced horses. If robots can eventually
reproduce like crazy, their capacity to live off “the smell of an oily rag”
might mean that wages in many industries in which humans are currently employed
will be driven below human subsistence levels.
However, it seems unlikely that robots will ever be viewed
by humans as close substitutes for human labour with respect to all attributes
relevant to all economic activities. My guess is that many humans will show a
strong preference for some goods with a high human labour input e.g. home
produced food, restaurant meals and beverages that are served by humans, live
music by local musicians, handicrafts and works of art produced by humans, and
some manufactured goods that individual humans have designed specifically for
themselves or friends and relatives.
My bottom line is that over the next few decades the impact
of robots in replacing human labour is likely to be a relatively small part of
the total impact of technological change on the quality of life. Rather than
worrying about robots replacing human labour perhaps we should be more
concerned that the rate of technological progress may be slowing down.
Why do my thoughts of 2015 need revisiting?
In 2015 I was thinking mainly about robots as physical
substitutes for human labour, and I posed the choice as one between two models:
robots as something close to slave labour (a pure increase in labour supply,
competing directly with humans) or robots as something more like horses and
tractors (displacing humans from some tasks while freeing them, and raising
their marginal productivity, in others). I concluded that robots were unlikely
to be close substitutes for humans across all the attributes relevant to
production, and that the wage effects of automation would therefore be a
comparatively minor part of the overall story over the following few decades.
Large language models (LLMs) were not on my radar in 2015.
They change the shape of the question, because they are far closer to
general-purpose cognitive substitutes than any robot arm or self-driving
vehicle. An LLM does not just replace one narrow task; it can draft, summarise,
code, translate, analyse and advise across an enormous range of activities that
previously required a trained human. In that sense, the "close
substitute" horn of my original dilemma is more relevant to cognitive and
creative work than I gave it credit for. I now think the economic impact of
LLMs will be massive - considerably larger than the impact of robots in
physical production that I was focused on a decade ago.
State plainly, LLMs are closer to general-purpose cognitive
substitutes than anything I had in view a decade ago. The AI revolution is very
much like an influx of low-cost labour that will replace human labour.
However, it is premature to jump to the conclusion that most
of the human population is about to become unemployed and will need to find some
source of income other than wages and salaries if they are to survive. In what
follows I discuss a range of factors that impinge on the speed and unevenness
of dispersion of the new technology before turning my attention to the profound
issues associated with checking and validation of AI output.
Human interaction
As argued in my original essay, some activities that depend
on human interaction may remain largely immune. People want to be served,
taught, cared for, negotiated with and entertained by other people.
Yet, this point should not be overstated. A lot of
"human interaction" jobs are information-transfer jobs with a human
face on them - for example, customer service, a lot of teaching, first-line
medical triage, even therapy. People are apparently willing to accept an AI
substitute when it's cheaper, faster, or less judgmental. AI companions,
tutoring bots, and mental-health chatbots are already being adopted.
Human advantages
Trades that depend on physical presence and situational
judgement - builders, plumbers, electricians, nurses - are unlikely to face
serious robotic competition over the next decade or two, whatever progress is
made in software.
Diffusion lags
Use of AI in physical production processes is usually a
costly and slow process. Diffusion into actual production processes can take
many years. Restructuring enterprises – investing capital, redesigning
workflows, retraining staff - is likely to take considerably longer than development
of the knowhow and systems design which makes automation feasible.
By contrast, however, adoption of AI can occur rapidly in
many service sector industries. Individual employees began using AI on their
laptops last year, with no capital expenditure and no sign-off from anyone.
That is arguably why uptake has already been faster than any prior
general-purpose technology. In service sector industries, the labour-market
effect may arrive informally and quietly, well ahead of any restructuring.
Costly inputs
The relevance of the "smell of an oily rag" idea
in my original piece - the thought that machines might reproduce and operate so
cheaply that they would drive human wages toward subsistence – depends on the
time period we are considering. During the next few years, electric power and compute
will be costly; the marginal cost of running systems at scale is real and substantial.
In the longer term, however, the situation will look different.
Computational and financial expense of generating outputs from trained
models is falling rapidly. Within a few decades, what looks now like an
unusually capable but costly input is likely to look more like a flood of
costless labour.
Potential for entrepreneurship
We should take account of the potential for enterprising
individuals, rather than large incumbent firms, to exploit these tools quickly
- using AI to launch new products and services. That will create new employment
opportunities, but not necessarily for the same people who lost the old
ones, at anything like the same wage, on anything like the same timescale.
The more important point is that the new technology may have
potential to help large numbers of people to find hidden entrepreneurial
capabilities. As I noted in a
recent essay, entrepreneurs may market new products to satisfy needs that
consumers previously didn’t know they had. For the most part, those
innovations are likely to be relatively minor and only of local or regional
significance, but they could add substantially to the incomes of the new entrepreneurs
and, in aggregate, make a substantial contribution to national income.
Complementarity
The "horses and tractors" story is still relevant
because humans are using AI as a tool to raise their personal productivity. Adding
an LLM app to a personal computer is like adding a hydraulic arm to a tractor –
the hydraulic arm enables the driver to become more productive because he can
use the tractor to lift things rather than just pull things.
Nevertheless, general-purpose AI has potential to displace
humans in a wide range of routine cognitive tasks over the next decade or so.
And, over the longer term, the market adjustment mechanism can be expected to
have progressively less scope to switch humans to other jobs.
Arguably, like most of the other points discussed above, the
complementarity point seems more relevant to considering the impact of AI over
the next few years than to considering what is likely to happen in the decades
ahead.
It is now time to introduce a more profound consideration –
verification.
Who will be responsible for checking what AI does?
Imagine that vast numbers of aliens suddenly arrive from
outer space and are willing and able to do nearly all the cognitive and
physical work that humans do at much lower cost than humans. However, even
though these alien agents are friendly to humans, they cannot always be trusted
to produce outcomes that meet appropriate quality standards and respect individual
rights (of humans). Adverse outcomes will sometimes occur because the agents are
fallible or because they may be willing to “cut corners” in pursuit of goals. (Those
are traits they share with humans.) At other times, adverse outcomes will arise
because some humans will instruct the agents to cut corners or infringe the
rights of others. The use of agents does not remove the need for principals to check
and validate outputs.
That scenario is similar to the one explored in a paper
entitled, “Some Simple Economics of AGI” by Christian Catalini,
Xiang Hui and Jane Wu, published by
SSRN, February 24, 2026. The authors’ core thesis is that artificial
general intelligence (AGI) makes output cheap and abundant; human verification
becomes the scarce factor; and the economy’s bottleneck shifts from doing
to checking.
Catalini et al describe a dynamic process leading to adverse
outcomes because the availability of humans with necessary oversight skills doesn’t
expand sufficiently as the cost to automate collapses:
“The current "human-in-the-loop" equilibrium is
unstable. It is eroded from below as apprenticeship pathways collapse (the
Missing Junior Loop), shrinking human expertise precisely when oversight
becomes most valuable. It is eroded from within as experts codify their own
obsolescence (the Codifier's Curse), converting experience into training data.
As capabilities outpace oversight, deploying unverified systems becomes
privately rational -introducing a "Trojan Horse" externality of
misaligned output. Using AI to verify AI only manufactures false confidence as
correlated blind spots propagate. Left unmanaged, these forces pull toward a
Hollow Economy of explosive nominal output but decaying human agency.”
The authors also present a more optimistic scenario:
“Yet this outcome is not inevitable. The answer is not a
retreat into obsolescence, but a radical elevation of human purpose. By scaling
verification infrastructure alongside agentic capabilities, the forces that
threaten collapse become the catalyst for unbounded discovery, experimentation,
and execution - powering an Augmented Economy.”
Nevertheless, after I had finished reading the article, I
was left with the impression that that authors think we are heading rapidly towards
a Hollow Economy. In their Conclusion they write:
“Autonomy is fundamentally outpacing oversight. Open-source
agent swarms are already operating autonomously at scale—integrated with
payments, infrastructure provisioning, and live user data—producing the first
visible instances of unverified agentic output in the wild. The same
infrastructure that enables accidental harm presents the ultimate attack
surface for deliberate abuse, granting malicious actors the leverage to scale
catastrophic harm at the marginal cost of compute. Inside major software firms,
AI produces a material share of new code, a share executives openly project
rising toward a majority in the near future.” (The authors cite references that
have been omitted from this quote.)
Catalini et al have provided an important warning about what
might happen, but it seems to me that a more nuanced view might be
appropriate, taking account of the potential for a diversity of responses by
individual firms and the incentives of the legal framework.
Expect diverse responses
I don’t think Catalini et al have suggested explicitly that intense
competition will cause firms to engage in “a race to the bottom”. However, that
idea surfaces frequently in public discussion of competition and may underlie
fears about the emergence of a Hollow Economy. A problem with race to the
bottom theory is that it assumes that all firms will make the same assessments
about the strategies most likely to enhance their profitability. The theory
ignores the possibility that while some firms will see advantages in reducing
standards, others will see reputational advantages in maintaining high
standards.
It seems likely that while some firms will seek to profit by
cutting corners on validation of output, others will focus on enhancing longer
term profitability by seeking to build inhouse expertise in checking and validation
of output. Interestingly, by drawing attention to the potential for a widespread
validation problem to emerge, the paper by Catalini et al may encourage more firms
to pursue an inhouse training strategy.
It should also be noted that even though AI is likely to evolve
rapidly to become more efficient at verification, the existence of a
technically perfect verifier does not solve the accountability problem. As one paper explains, high-stakes
domains require a human who understands the content and accepts responsibility
for its use, even if the AI's check is flawless.
Having introduced the concept of accountability, we now need
to consider how the choices which firms make are likely to be influenced by the
incentives provided by the legal and political environment in which they operate.
Legal liability
There is a good reason why I have chosen the Roman legal maxim,
“He who acts through another does the act himself” as the epigraph for this
essay. That maxim has modern legal application in laws which establish that
principals are legally responsible for the authorized acts and contracts
performed by their agents.
Existing tort and regulatory frameworks already hold humans
responsible when AI-assisted outcomes go wrong. This complicates claims that unverified
deployment of AI is likely to be highly profitable. In any context where firms
currently adopt rigorous checking practices to reduce potential legal liability
for damage to others, they will continue to have an incentive to maintain
rigorous checking practices. Insurance companies offering cover for legal
liability may refuse cover to firms which cut corners.
Regulation
Government regulation impinges on a wide range of activities
in modern economies. Firms using AI will still incur penalties if they fail to
meet regulatory requirements. For good or ill, that places an increased burden on
firms to have checking mechanisms in place to ensure that compliance occurs.
In respect of some industries, e.g. medicine, law and
education, government regulation typically determines who may supply services. For
example, if an individual uses AI to diagnose the nature of an illness they
suffer from, they are likely to be required to consult a human doctor to check
that diagnosis before they can obtain pharmaceuticals to treat the illness.
In recent years, the concept of "meaningful human
control" has entered legal-political discussion in response to worries
that autonomous systems create a "responsibility gap" where lethal or
harmful outcomes occur but no human can properly be held morally responsible
for them. In a
paper published in 2018, Filippo Santoni de Sio and Jeroen van den Hoven presented
a philosophical account of conditions that need to apply for humans to remain ultimately
in control of autonomous weapons and other autonomous systems. The authors
suggest systems should be able to respond to a “tracking” condition (relating
to relevant moral reasons of the humans designing and deploying the system and
relevant facts in the environment in which the system operates) and a “tracing”
condition (to enable the outcome of its operations to be traced back to at
least one human along the chain of design and operation). The authors also
suggest that meaningful human control requires training systems to improve
users’ understandings of the risks and responsibilities associated with
operating them.
For present purposes, the important point is that meaningful
human control requires substantial human inputs.
It is unfortunately also necessary to note that advances in
AI have added to the capability of predatory individuals and organizations (including
authoritarian governments) to purposefully cause financial and physical harm. In
that context, humans may need to focus more attention on protecting themselves
from predation.
So, will AI replace humans?
After all that I have written above, perhaps the most
appropriate conclusion would be that I am not a prophet, so I don’t claim to know
whether AI will replace humans. That is certainly the case if we are talking about
what might happen in the 22nd century or beyond.
However, if we are talking about the next few decades, it is
possible to make some relevant points:
·
The economic impact of LLMs is likely to be
massive. This is a major step in the ongoing process of replacement of human
labour by non-human labour.
·
The impact of LLMs will be uneven. In the
services sector, there is likely to be substantial displacement of human labour
engaged in routine cognitive activities; trades that depend on the physical
presence of humans will be less affected, and some activities involving human
interaction will be largely immune. Automation is likely to continue to proceed
at a moderate pace in manufacturing and agriculture.
·
A major human input will still be required to verify
AI output. Firms will need verification to protect their reputations.
·
Existing legal and regulatory frameworks hold
humans responsible when AI assisted outcomes go wrong – providing additional incentives for human
verification.
Even after taking account of the human labour required to
verify output, we are obviously still talking about substantial displacement of
human labour. However, that doesn’t necessarily translate into a situation
where there will be growing numbers of intelligent humans either enjoying a
life of leisure (or sitting in their homes twiddling their thumbs).
This brings me back to the bottom line of my 2015 essay - that
the wage impact of automation would be a relatively small part of the total
impact of technological change on the quality of life. For the next decade or
two, I think that claim still holds, for the reasons set out above: uneven
diffusion, the persistence of human-interaction and physical-presence work, and
the ongoing need for human verification. Over a longer horizon, as compute
costs continue to fall and verification itself becomes easier to scale, I am
considerably less confident the claim will hold - which only means my 2015 self
was answering a shorter-run question than he realised.
A massive growth of productivity offers potential for a
massive expansion in human leisure, but that outcome isn’t inevitable. It is
worth remembering in that amid the Great Depression of the 1930s John Maynard
Keynes predicted (in an essay entitled “Economic
Possibilities for Our Grandchildren”) that by now we would be able to
satisfy our needs without having to work more than three hours a day. Average
incomes rose more than Keynes predicted and working hours fell much more
gradually. Keynes had no way to predict the emergence of new products to
satisfy needs that consumers didn’t know they had.
It remains an open question of whether humans will use the growth
of productivity resulting from AI for increased leisure, or whether they will continue
to work long hours to earn the income required to satisfy needs that they don’t
yet know they have.
No comments:
Post a Comment