There’s a truism that technology is good - even if it creates winners and losers, it improves the world. Toby Ord argues that the conclusions about the benefits of technology is sensitive to the end of humanity - but this jumps over the transitions by starting from the assumption^[1] that “long-term progress in science, technology, and values have tended to make people’s lives longer, freer, and more prosperous.” That is, looking back historically, the net impact misses the immense immediate harms of large scale technological changes that can last for generations.

As I’ll explain, the largest technological revolutions in human history are arguably the agricultural revolution and the industrial revolution. In both cases, the vast majority of those immediately affected were harmed, not helped. Of course, the longer term impact was positive; those benefits are not in question^[2] - not that those alive during the transition should have cared.

The two obvious examples

The invention of agriculture led to increased food availability and around ten thousand years of greatly worsened health and lifespans^[3]. The wealthiest and most powerful people benefited immensely from the population explosion, and from the wars that larger populations enabled and required; the population suffered from both malnutrition, and that same increase in the scale of violence^[4].

The invention of industry was more beneficial to the consumer - but not to those directly involved. In 1840, over a third of the British population worked in a factory. This was bad, in part directly due to factory worker deaths, but also due to pollution and disease. Mortality shot up over the middle of the 1800s - the famed “urban penalty”, especially among children, albeit partially offset by reduced deaths because of sanitation later in the century^[5].

More Data?

These two examples are potentially misleading, because they are the motivating rationale for the view, rather than evidence being examined for that. So we should look more widely. Of course, this leads to having written the bottom line before looking for evidence, but I’ll look externally for the arguments, and provide the context so others can account for the filtering. So I asked ChatGPT; “What would you list as the five largest technological transformations that humanity had ever undergone?” The list includes the obvious two above^[6], and a follow-up^[7] laid out the case for the negative impacts:

1. Agriculture / the Neolithic transition - Sedentism, surplus, population growth, hierarchy, states, disease burdens, property, warfare at scale. Near-term farmers often got worse diets, more infectious disease, more labor, lower stature, worse dental health, and more hierarchy.
2. Writing and external symbolic storage - Administration, law, history, mathematics, scripture, bureaucracy, long-distance coordination. Early writing mostly helped palaces, temples, tax systems, accounting, property claims, labor control, and bureaucracy before it helped ordinary people read novels or do science. So the near-term “users” benefited, but many affected subjects may have faced more legible extraction and administration. Evidence from early Mesopotamia links writing with larger government buildings and multi-level bureaucracies.
3. Metallurgy, especially iron - Tools, weapons, plows, empires, deforestation, intensified agriculture, military expansion. Bronze matters too, but iron’s scale and availability make it more transformative. Better tools helped agriculture and craft production, but weapons, fortifications, conquest, inequality, and elite control plausibly dominated early experience for many. The case is less clean because metal tools also had immediate productive benefits, but the war-and-hierarchy channel is very real.
4. The fossil-fuel industrial revolution - Steam, coal, factories, railroads, mechanized production, urbanization, modern economic growth, climate change. Early industrialization brought urban crowding, dangerous factory work, child labor, pollution, disease, and artisan displacement. The “standard of living” debate is contested, but mortality and health in rapidly growing industrial towns are a serious mark against the near-term story.
5. Electricity + computation + telecommunications - I’d bundle these reluctantly as the “information-electrical stack”: telegraph, telephone, radio, electric grids, computers, internet, AI. This led to surveillance, labor displacement, attention capture, military command/control, financial acceleration, and dependence on fragile networks^[8].

This is far from a comprehensive survey, though I think it shows that the net impact of technological changes is at least often and plausibly negative over the short term, and this changes to positive later on. But there is a glaring exception, and it bears consideration.

Some Technologies Are Good, Actually

There have been a couple of revolutionary changes in medicine and public health over the past couple centuries. The vaccine revolution, the advent of modern sanitation, and infection control each include a strong case that they were immediately beneficial, and stayed that way indefinitely^[9]. Refrigeration, washing machines, and bicycles^[10] are arguably more examples in this class. So some technologies really are just positive - but we need to ask which ones.

I think there’s a simple explanation; directly good things are good, but many other classes of transformative change end up disruptive in ways that hurt before they can help^[11]. Technologies that have first order impacts on coordination and production, or that empower groups in other ways, tend to differentially benefit the powerful in ways that are harmful to others, either directly or indirectly^[12].

The Artificial Elephant in the Room

The start of the next large scale transformation of society has begun^[13]. If AI is a normal technology, it’s normal in the sense of agriculture, industrialisation, or public health^[14]. The question is what this looks like, and I think the above gives us a few places to think more carefully about the impacts.

First, how strong is this base rate? We should obviously be skeptical of expert predictions intuition, given their track record. Traditional judgemental forecasting best practices starts with the base rate, and adjusts on that basis - but this suffers from a reference class problem; should our base rate be for technological revolutions, which the above assumes, or should we be asking about the emergence of a new smarter species, akin to evolutionary transitions^[15]? Even if it’s the correct base rate, we have such a small n, with so many differences, that our views about what is changing could easily overwhelm the base rate.

Second, if AI follows the trend, we would ask how long the negative period lasts. Agriculture’s temporary misery was 10,000 years, the Industrial revolution’s negative phase was closer to 100, and it seems plausible AI’s harmful phase will be continuing the trend on a logistic scale, so that we might be in for an annus horribilis to end all others^[16]. Of course, the depth of the trough is incredibly significant! We could see mass disruption, unemployment, riots, national and international collapses due to a complete end of trust in objective truth not literally perceived. If we pull through, this might be followed by a benefit that is almost unimaginable - but if the transition, however brief, includes mass-scale biological weapons, nuclear war, or other global catastrophes, the median human would (and will, and loudly does) oppose any such transition.

Lastly, we can look at current impacts, rather than base rates. I mentioned above that the impact of some of these technologies started out positive before the massive disruptions, and I think we can see the clouds on the horizon. So far, AI has done a tremendous amount for us, from eliminating the need to deal with parking attendants or toll booths via license plate recognition, to giving us AI boyfriends. It’s speeding up research, it’s finding vulnerabilities, and it’s accelerating software companies with little effect on overall employment, at least yet.

This has been touted as evidence, but it seems like very weak evidence, especially given any structural model of AI impacts which is more complicated than picking specific final outcomes over time and projecting them linearly. And I could argue here that this doesn’t solve for equilibrium^[17]. But the other reason to make such arguments is the previous examples; the immediate impacts of agriculture were more food. The immediate impacts of industry were jobs and mass manufactured goods. It wasn’t until the second order effects kicked in that we had problems.

Conclusions and ways I might be wrong

I’ve tried to lay out the reason that technological optimism on the basis of history is legitimate over longer time frames, but insufficient or even counter-evidence in the shorter term. I think the case is clear, and I think it takes a certain myopia to confidently believe that even in good worlds where things turn out wonderfully, we’ll think that there wasn’t a period of time where AI was obviously net negative.

How confident am I in this prediction? Not very strongly, but I’d put 2:1 odds that, conditional on transhumanists agreeing that we got at least a moderately good future with ASI in 2040, that there is widespread agreement that there was some period of time where it was clearly net negative.

There are a few ways this could fail. Obviously, my statement was a conditional, and we could fail to reach the conditional, we might stop before we build ASI as the world realizes how dangerous it is and how little we know about how to keep it safe^[18]. If the conditional is fulfilled, I might still be wrong; I could also be too pessimistic about the timeline, and the negative period could last weeks instead of years, and we could move past the time of perils quickly enough to get lucky in not having an bioengineered plagues or nuclear wars during that time. Alternatively, we could imagine that ASI materializes so quickly that it could even outpace the disruption. For example, an AI-foom occurs, and we get insanely lucky in getting aligned ASI. In such a case, perhaps it solves unemployment via UBI before eliminating jobs, and that same week it replaces the consequent lack of meaning with something humans appreciate instead, so we move directly into a Banksian post-scarcity future.

If all of these ways I could be wrong sound like science fiction, I apologize, but that’s what I expect them to be. In reality, I think that those of us alive today don’t want this technological revolution to occur.

1. ^{^}

   It also assumes total utilitarianism and no time preference, and those of us living through such a transition who are less than fully impartial (like myself,) or those who argue for something less than full total utilitarianism, a group that now includes Will MacAskill, have reason to object.

2. ^{^}

   These transitions were both physical, and as Morris suggests, moral. But those followed, or were instrumental in leading to, the benefits, they didn’t cause them.

3. ^{^}

   The tail end of the agricultural revolution was the iron age, which was arguably its own revolution - see below.

4. ^{^}

   It’s possible that the net result was a reduction in overall violence, which Pinker argues - but that’s based on ethnographic studies of hunter-gatherers today, where the resource contention is already fierce - exactly the thing that agriculture made happen. See WrongBot’s 2010 post.

5. ^{^}

   Credit John Snow’s discovery of the source for Cholera in 1854, followed by the London sewers being built in the 1870s and the public health act in 1875.

6. ^{^}

   ChatGPT hedges, of course; “The big ambiguity is whether to count language and fire. If yes, they probably dominate the list… but for an essay about ‘technological advances’ in historical societies, I’d probably use the first list and discuss fire/language as preconditions rather than transformations within civilization.” I exclude them, since we don’t have data, and they arguably created modern humans, rather than being invented by them. Perhaps this is cheating, given the likely positive impacts, at least long term - but that’s partly the problem, as we have no idea what the transitions looked like!

7. ^{^}

   “Don't count language and fire. Using the list that you gave, is there a plausible case that each of them was net negative for most of the people who were immediately affected in the near term by the new technologies?”
   This is an admittedly leading question, revised from the first attempt: “Don't count language and fire. Using the list that you gave, what was the impact in the near term of the new technologies?” which ignored the negatives: “Agriculture - More people per land area. Writing - Memory outside the human brain. Iron metallurgy - Cheap strong tools and weapons. Fossil fuels - Energy beyond muscle, wind, water, and wood. Electricity/computation/telecom - Instant coordination and automated information processing.”

8. ^{^}

   ChatGPT is very aware of my previous work, which biases things further.

9. ^{^}

   Even the tiny negatives from each only emerged much later, after the primary positive impacts. Too-successful hygiene, infection control, and sterilization increased allergy/autoimmune risks in the past decade or two, and a generation of peanut allergies were one visible (but relatively minor) result. There are even a couple of small scale examples for vaccines; formalin-inactivated RSV vaccine didn’t work and enhanced RSV instead, and something similar happens with dengue vaccines. And, of course, there’s incorrectly inactivated polio vaccines and the 1977 flu pandemic, but for vaccines, those were all failures of the technology, rather than net negative impacts of the technology working - though this makes a critical footnote for considering potential AI catastrophes.

10. ^{^}

    ChatGPT suggested that the printing press belonged in this list. That seems wrong, given the near term religious fragmentation, war, propaganda, and destabilization more broadly that preceded the fuller realization of the immense benefits of public literacy.

11. ^{^}

    It’s not obvious why bicycles wouldn’t be disruptive, until you realize they weren’t invented until after trains, so they weren’t driving much large scale change, they were primarily personal transit.

12. ^{^}

    The internet seems like a good edge case, where it was initially empowering to individuals, albeit mostly a selection of high-status and/or well educated ones, and only later mostly empowered the global elite and large corporations. This may be typical; I imagine similar dynamics for agriculture lasting years or decades before millenia of net-negative impacts. Similarly, small factories, initial use of industrial equipment, and small-scale industrialization have a better claim to having immediate positive impacts for labor saving than the negatives that followed in later years and lasted decades. (The fact that AI is currently largely empowering fits into this pattern as well, but I’ll get to this later.)

13. ^{^}

    If you disagree, and want to regurgitate irrelevant claims that were made before ChatGPT was released, find somewhere else to parrot your views. I’m not interested.

14. ^{^}

    To be clear, this doesn’t disagree with Narayanan and Kapoor’s description of AI as a normal technology; “AI is like previous technological revolutions in human affairs.” They even use the Industrial revolution as an explicit comparator. It disagrees with their prediction, but it does so only for reasons that were laid out above, around how technological revolutions have actually changed the world. If AI is not a normal technology, at least in some ways, as Den Ball argued, the implications are different.

15. ^{^}

    This argues for a big transition, but perhaps not for a bad ending?

16. ^{^}

    And yes, if Eliezer Yudkowsky is right (which seems entirely plausible) this would quite literally be the end.

17. ^{^}

    That is, if productivity skyrockets then the downstream impacts will include employment changes, if research gets far cheaper to produce, the down impact of that research won’t increase linearly, etc. This point should be disqualifying of anyone who argues that the current impact on jobs is sufficient to conclude AI’s net impact on jobs will be negligible - you need a more sophisticated argument, with many more assumptions, to make that claim.

18. ^{^}

    Or, for the sake of completeness in listing unlikely but not impossible edge cases, we might have already hit the increasingly implausible but widely asserted cliff of AI capabilities a bit below human level for general intelligence, and AI might only be as transformative as, say, renewable energy or satellites.