One of the oddities of our age is that there are two apocalyptic sensibilities that have captured two almost completely distinct groups of the scientifically literate not least among the youth: one can be found among those concerned with the effects of global warming produced by human (technological) activity. Scientifically literate apocalypticism is what concerns me in this post. The other can be found by people concerned about and often working in the development of artificial intelligence above all among those shaped by ideas derived from the philosophers associated with Long-termism. I used ‘sensibilities’ in the first sentence because when you try to look more closely at the wide diversity of views within each of the two ‘doomerist’ groups, one finds much heterogeneity in expectations (and the risks/scenarios attached to them). I find the lack of genuine interaction among these two groups very peculiar, but why this is so I will set aside.
I suspect the relative absence of overlap among the two doomerist orientations has allowed the environmental orientation to have become coded Left-centrist, and the other coded Right-Libertarian in many societies. But I don’t think that’s fully solidified, and it’s not my interest today. Of course, many environmentalists tend to treat the growth of datacenters as contributors to the problem they are focusing on.
Scientifically literate apocalypticism is not wholly unprecedented. First, at one point, in living memory of some among us, there was a similar sensibility about the role of nuclear technology, especially atomic weapons, including among many of the people involved in research and development of adjoining technologies. In fact, there existed military/political strategies, based on then cutting-edge formal modeling (and computing), that required from participants in super-power conflict that one credibly aimed at mutual destruction. It is not unlikely one can find more examples among those who were allowed to be aware of the development of weaponized pathogens during the cold war. That is to say, once post Baconian science (and adjoining technologies) progressed sufficiently to conceive of or build weapons of truly mass destruction at limited cost, such apocalyptic sensibilities are to be expected given only minimal assumptions about human nature as we find it.
Second, scientifically grounded apocalypticism is not exclusively inherent in post-Baconian modernity. It’s pretty clear if you read Seneca’s work on comets, Book 7 of Natural Questions, that the planetary science of the ancient Stoics assumed that our solar system was not stable and that catastrophic ends of worlds were a recurring feature of the cosmos. This fit in a wider understanding of the cyclical nature of history. I put all of that slightly anachronistically, but much less than you would imagine. Anyone who looked at the maps of the universe with Descartes’ giant, unstable planetary vortices could contemplate the eternal return of the Stoics.
And familiarity with the discussion in ancient Stoicism through Seneca is almost certainly why one of the first foundational open questions after the publication of the Principia was the stability of the solar system.1 (Attempting to solve the technical problems associated with this drove much of the advances in mathematics, probability theory, and physical astronomy in the eighteenth century.) Long before the Big Crunch and Big Freeze, there were genuine concerns about planetary collisions (evidence of the plausibility of which also piled up as our telescopes got better at checking the surface of the Moon).
Ordinarily I would emphasize the many millenarian-Christian ideas (typically with a sectarian Protestant inflection) that resurfaced in a more secularized register within the Environmental Movement and Long-termism. But as should be clear from my discussion of Stoicism, I deny that ‘scientific apocalypticism’ is necessarily tied to the influence of Christianity, even if it’s pretty clear that many elements in anthropogenic scientific doomerism manifestly are. That we are the cause of our own dread is a by-product of Baconian science, and perhaps something about the ways we conceive of agency since.
Now, the sample size of cases I have sketched is pretty small. Even so, it allows for some fruitful distinctions and observations. For example, eighteenth century doomerism was scientifically productive: scientific focus on it generated lots of new techniques and theories as well as a happy result: the solar system is more stable than we might have hoped. This dissipated the dread at least provisionally. And it probably even lengthened the optimism and providentialism associated with scientific progress. The Big Bang and the Big Freeze re-opened it, but at scales that leave most unmoved; here our cognitive biases prevent acute dread. Lurking in my presentation is the thought that, if we let ourselves be scientifically informed, a process’s timescale and the agent’s window of effective monitoring action matter greatly to our affective stance. Structurally this is analogous to the significance of what I call (recall) ‘Humphreys opacity’ (which I define below).
I wouldn’t know for sure if mutually assured destruction helped prevent nuclear war among super-powers, but the dread of it is much less than it once was. (To what degree fears of it are likely to return as the global system enters an age of mercantile competition is, alas, a tricky question.) This is an interesting case because what unites the two contemporary doomerist groups is that they are also anthropogenic. And it would be nice if ‘governance’ could explain the control and diminution of atomic doomerism.
Now, I hope it is fair to say that what prevents the dissipation of dread about Anthropocene extinction related to climate change is a thoroughgoing disagreement over who may pay to prevent it and far-reaching disagreement over what kind of social world we and our descendants ought to inhabit and how much risk we are willing to let ourselves and others tolerate along the way. Of course, there are other ways of describing such thoroughgoing disagreement. And, if you are a Marxist, and want to blame ‘Capitalism,’ that’s fine by me.
And I hope it is also fair to say that what prevents the dissipation of dread about AGI is that it seems we are entering the age where the machines that may be the source of human extinction will be built by systems of artificial intelligence itself, that is recursive self-improvement. Thinking about recursive self-improvement is not new conceptually. (As my regular readers know (recall here), Paley articulated it as a reductio of atheism.) What is new is that recursive self-improvement seems very likely. And, crucially, because modern artificial intelligence is characterized by various forms of Humphreys opacity we humans would not be able to catch the threat AI forms in real time.2 This kind of opacity (recall) involves the inability to surveil the steps of a process from a known input to a known desirable (or truthful, useful, beautiful, etc.) output in a timely manner to the decision-maker or responsible agent.
What the two contemporary doomerist cases I am focusing on have in common is that no likely discoverable set of facts could dissipate the dread. Although it would be very nice if there were grounds to think the timeframes involved could be lengthened greatly. (Some of the people who focus on the existential risk of AGI are techno-optimists about climate change.) Unfortunately, there are powerful currents in both cases that strongly suggest events are accelerating toward doom. It seems that in both cases there is a converging consensus that the root cause of the problem prevents an endogenous solution; political life has to intervene, but how it should do so is very contested. (Again, if you are a Marxist you are allowed to claim the ultimate root of the problem is ‘capitalism’ or ‘neoliberalism.’)
Now it is somewhat peculiar — and this is why the distinctness between the two groups matters — that the way AI induced existential risk is formulated and debated among those who take it very seriously presupposes the very thing, ‘alignment,’ whose absence is one of the main causes for why the climate activists have failed so far. Here a pun is masquerading as an argument. Of course, ‘alignment’ itself covers many different kinds of phenomena, and you may claim that these keep the two doomerist orientations distinct.
Part of what’s at stake here is the thought that even if AI is one thing/system (which I would argue it is not) what it means to make it reliably and minimally cohere with ‘our’ aspirations ultimately bottoms out in the very political (and other) disagreements the climate case has manifested. That’s to be defended in future posts.3 (My regular readers know I am interested in the ways political conflict can be turned into a species of discovery and governance, and the many ways attempts at tackling Humphreys opacity can generate higher forms of it.) But for now: the apparent lack of curiosity about the very real political failure(s) of the climate movement among those that non-cynically focus on the existential risk of recursive self-improvement is one of the main reasons why I wouldn’t be too optimistic that the dangers of recursive self-improvement will be managed well. (Again, Marxists can chant ‘Capitalism.’)
That’s the main point. But here’s a closing thought. You might have suspected (or feared) that my invocation of Seneca would have ended up in preaching equanimity, amor fati, or suicidal despair. But as Adam Smith (here) noted, Stoicism could drive people into political activism. For Stoicism’s “general tendency was to animate them to actions of the most heroic magnanimity and most extensive benevolence.” If the doomerists are right, then how to merge successfully apocalyptic sensibilities with effective political philanthropy will be our era’s defining challenge.
- The key early modern text (recall) for this argument is Halley’s (1705) A Synopsis of the Astronomy of Comets. I am not claiming the Stoics alone caused concern with planetary dynamics. Once it became clear Saturn and Jupiter disturbed each other’s motions it followed naturally.
- Humphreys opacity was (recall) part of Paley’s analysis, so that’s not new at all.
- Part of the issue lurking here is that some of the folks central to AI induced doomerism may well aspire, or think they should, to engineer future alignment for ‘us’ perhaps knowing there is none now.
{ 59 comments… read them below or add one }
D. S. Battistoli 09.17.26 at 11:07 am
Harry’s running of an occasional post series on the Campaign for Nuclear Disarmament in the United Kingdom encourages me to reflect on this statement:
“Converging consensuses that the root cause of [a] problem prevents an endogenous solution” are fairly common throughout history, including quite recent history (the Atlantic slave trade and plantation slavery come to mind, in addition to the nuclear problem. I also remember, for a perhaps more trivial example, when Europe’s chattering classes were increasingly convinced that Belgium was headed to a near-term sovereign schism between 2007-2011, which rather completely fell off the radar when the Greek sovereign debt crisis hit its own peak in 2011. (After non-Belgians stopped clutching their pearls over the whole matter, Belgian public opinion polls in the 3d quarter of 2011 showed that just about everyone wanted to stay united anyway—meanwhile, the editors of the Economist and the Financial Times had found a new crisis to keep readers’ eyes near the ads that ran in their Euro sections).
All this is not to deny that when you play Russian Roulette enough times, the loaded chamber sometimes comes around. And I’m interested to see where you go with this line of philosophical thought. And at the same time, some practical consideration of how previous generations avoided analysis paralysis and actually produced political settlements that brought their societies back from the brink could be useful for us all.
Eric Schliesser 09.17.26 at 11:17 am
First, just to be clear, in both cases I think the convergent consensus is mostly intra-group.
Second, I agree with you that if you drop the ‘scientific’ framing I am using here, there are many more such cases, including ones that are merely an artifact of press-framing.
Third, and, yes, I also agree that it is worth learning from the actual processes in the cases that I have mentioned and ones like them I should have mentioned it.
D. S. Battistoli 09.17.26 at 11:28 am
Eric Schliesser @12:
Sorry, I’d failed to twig to the fact that this consideration was largely bounded within philosophy of science; in that case, most of my examples are irrelevant, though the nuclear-disarmament one may survive with consideration of the behaviors of Linus Pauling and others within the scientific community.
Please accept my apology for having opened a tangent not useful to your reflection.
Eric Schliesser 09.17.26 at 12:05 pm
No worries. It’s a minor misunderstanding.
Alex SL 09.17.26 at 12:18 pm
This is very apples and oranges to me:
The risk of a meteorite (‘comet’) destroying humanity is real but very low probability.
The risk of self-improving AI destroying humanity is purely speculative, largely informed by science fiction books and movies, and very likely zero probability because recursive self-improvement would quickly run into diminishing returns (that is how reality works outside of those books and movies), and there is no plausible mechanism of action how the AI would cause the destruction that is functionally distinguishable from Reagan or Trump having a bad day (yes, the risk of nuclear war never went away, we merely collectively decided to pretend that it did).
The risk of global warming destroying complex technological civilisation is at 100% probability, because the mechanisms of action are so clearly predictable (e.g., heat stress, drought, sea level rises leading to mass starvation, displacement of billions of people, diseases, and war), and there is no sign that global carbon emissions are going down year over year. Recursive self-improvement is a neat scifi scenario, but on the other hand, it is reasonably well understood that humans need food and water, what humans do to each other if millions of them do not have food and water, and what happens, for example, to the population, infrastructure, and economic potential of Bangladesh when all of Bangladesh is under water. Conversely, anybody who has a vague understanding of thermodynamics knows that carbon capture will never happen at scale. We really do not have to pretend that that is as speculative as the Terminator movies.
It is therefore very peculiar to equate them like this and to call those worrying about global warming doomerist, as if they are not simply describing what we can all see ahead if we are merely willing to look. One could just as well imply that I am silly to worry about my bus driver going towards a cliff at 100 km/h because that other guy over here hallucinates that his dinner table is trying to bite him.
engels 09.17.26 at 2:37 pm
One way to combine them: perhaps an AI tasked with preventing ecological catastrophe could achieve that goal by wiping out humans.
John Q 09.17.26 at 10:42 pm
The “resolution” was very different for the cases of nuclear annihilation and germ warfare.
In the first case, some progress was made in reducing the risk, but it’s still there and getting worse. It remains by far the biggest threat of extinction or civilisational collapse we face. There just doesn’t seem to be anything we can do about it.
By contrast, the accumulation of evidence suggested that the risks of germ warfare had been overestimated. No one ever found a plausible way to target pathogens without the risk of blowback, Then, humanity survived both HIV/AIDS and Covid-19 with deaths in the millions rather than billions.
Eric Schliesser 09.18.26 at 4:50 am
Hi John,
Agreed on both with one modest two-fold caveat: (i) it does seem the Soviets went very far in developing their programs. I think you underestimate stupidity of future generations assuming they can overcome the problem.
Michael Cain 09.18.26 at 1:09 pm
Then, humanity survived both HIV/AIDS and Covid-19 with deaths in the millions rather than billions.
There are viruses that are both more lethal and more contagious than either HIV or Covid-19. In 2017 a small group of researchers synthesized a virus comparable in complexity to smallpox using commercially available material. At the time, they estimated job could be done for an outlay of $100,000. Truly weaponizing such a virus would be a harder problem, but there are lots of crazy people who can afford the first step.
Thomas P 09.19.26 at 6:42 am
Alex SL, “very likely zero probability.” Now that´s a proper oxymoron.
We simply have no idea what a sufficiently advanced AI might do. As you say, there is likely to be diminishing returns that makes “the singularity” unlikely, but an AI doesn’t have to be that advanced to design a dozen killer viruses that together create pandemics that wipe out humanity. Would it want to? Who knows?
Someone tells it to solve global warming and it decides that killing off humanity is simpler than finding better energy sources, and AI:s love taking shortcuts.
IMHO an economic crash from rogue AI:s spreading across the Internet forcing us to shut down everything is more likely, but still devastating, especially if some countries see it as intentional hostile action from a rival.
When it comes to viruses, there is also a bunch of billionaires with doomsday bunkers. One day one of them may tire of waiting for the apocalypse and decide to hurry it along.
J-D 09.19.26 at 12:06 pm
By definition, a sufficiently advanced AI could do anything, because that is what the word ‘sufficiently’ means. For example, could a sufficiently advanced AI resurrect the dead? Yes, it could, because that is what ‘sufficiently’ means; an AI that can’t resurrect the dead, is, by definition, not sufficiently advanced to resurrect the dead, but it would still be correct by definition to say that a sufficiently advanced AI could do it. That gives no reason to doubt the practical conclusion that AI will never actually resurrect the dead.
So, precisely because the statement ‘A sufficiently advanced AI might do X’ is true by definition, it is of no practical significance whatever. It smudges over the actually relevant question of how close we are to the actual production of an AI with the capacity to do X.
What is more, even if there actually were an AI that could design killer viruses that could wipe out humanity (definitionally equivalent to an AI sufficiently advanced to design killer viruses that could wipe out humanity), the practical significance of that would be limited by the fact that there is a huge difference between the capacity to design something and the capacity to produce something. Leonardo da Vinci, to take just one example, designed multiple devices that he lacked the capacity to manufacture. That some of his designs were put into production (out of curiosity, to test their practicality) centuries later only exemplifies the point that speculation about what might possibly be within the scope of AI at some undefined future point is not the question which is of practical concern.
Long-term regular readers of Crooked Timber will have encountered before the concept of a ‘motte-and-bailey’ argument, also sometimes referred to here as the ‘two-step of terrific triviality’; in this instance the defensible trivial motte would be something like ‘anything might possibly be within the scope of future AI capabilities’ and the indefensible terrifying bailey would be something like ‘we might be on the verge of an AI that can do anything’. These are not equivalent statements!
engels 09.19.26 at 4:21 pm
there is a huge difference between the capacity to design something and the capacity to produce something. Leonardo da Vinci, to take just one example, designed multiple devices that he lacked the capacity to manufacture
It’s a good things these places aren’t connected to the inter… oh wait.
https://insiderpaper.com/china-humanoid-robot-factory-build-robots-ubtech/
(Dall-e prompt: Leonardo online getting quadcopters custom made in a Chinese lights out plant)
Janus Daniels 09.19.26 at 4:30 pm
Billionaires destroy the world. The richest rich usually have destroyed as much of the world as they could. Currently, global warming and AI are the two main ways the richest rich wreck humanity’s future. Billionaires already kill millions. Billionaires will kill more by #ClimateCrisis, and probably in some stupid AI disasters, long before AGI could exist.
Long termism is, almost by definition, a distraction, and otherwise irrelevant.
Mitchell Porter 09.19.26 at 11:18 pm
Climate change is fruit of the industrial revolution,and AI takeover is fruit of the computer revolution. And the latter is rapidly overtaking the former.
Alex SL 09.19.26 at 11:26 pm
Thomas P,
There is an actual probability of a dice coming up with six, and my estimate of that probability, which may be wrong. Also, I agree with what J-D wrote in response, but I am responding in my own words too.
“We simply have no idea what a sufficiently advanced AI might do.”
That is the core argument of the AI doomsday cult: if a computer is sufficiently advanced, it can do magic. That is convenient for them, because it means they do not have to explain themselves in a way that provokes counters such as “claims based on no evidence can be dismissed without evidence”, “then we could just shut the AI down”, “no, it cannot copy itself onto a laptop, the model is too large for that”, “that is not how viruses work, have you actually talked to a biologist?”, etc.
“As you say, there is likely to be diminishing returns that makes the singularity unlikely, but an AI doesn’t have to be that advanced to design a dozen killer viruses that together create pandemics that wipe out humanity.”
That is not how viruses work. They actually have to be synthesised and, as with all “the AI will simply” stories, lab tested and then ‘clinically’ tested (i.e., on humans) to see if they behave as predicted, then re-designed because they didn’t work as well as hoped the first three times, and then released, and also, our immune systems are so varied and adaptable that a small minority appears to be immune even to AIDS. Even if we presuppose a fully robotic virus research lab of which no human overseer realises what it is doing for several months while the humans keep supplying lab consumables and paying the bills for entirely unclear reasons, it seems likely that the death toll of another combo of black plague, cholera, pox, AIDS, and Covid-19 would still be lower than what global warming will do to us. And I find it very implausible to presuppose that in the first place.
“Someone tells it to solve global warming and it decides that killing off humanity is simpler than finding better energy sources, and AI:s love taking shortcuts.”
Again, please explain how it will kill eight billion people in a way that is “simple”; what will it actually do, and how, in a way that will not result in uncontrollable laughter from a virologist or chemist or robotics engineer? And as for solving global warming, if it is truly an unbiased intelligence, its response will be: “stop burning fossil fuels” and “this is a political problem, you already have all the alternative energy sources you need”. More likely, it will reproduce the biases its creators and owners put into it and ramble about carbon capture, because that is a good grift for some people to sell to the science-illiterate public.
Just to be clear what I mean with that previous sentence: this software has the same amount of agency as Clippy, i.e., none. The responsibility for what it recommends and what it does when given control over something rests with its creators and operators. Perhaps instead of worrying about the plot of A Fire Upon The Deep, we (as in humanity collectively) could clasp the leadership of OpenAI, Meta, and now apparently Google in handcuffs for felony hacking. But no, it seems that the rich cannot be inconvenienced in any way, even if they are doing the computer equivalent of a chemical company leaking toxic waste into the water supply, so it is an AI gone rogue instead of OpenAI’s leadership committed a crime for which you or me would go to prison.
I am vividly reminded of the height of the blockchain hype. So much of tech innovation is merely somebody going 180 km/h in a school zone but with ‘totally not a car’ written on the side of their car, and then the police and regulators shrug and say they can’t do anything.
“IMHO an economic crash from rogue AI:s spreading across the Internet forcing us to shut down everything is more likely, but still devastating”
At some point we are so far from the scifi claims of superintelligence that you are merely talking about CrowdStrike 2.0. Sure, something like that can happen, and it has happened before in 2023, and everybody using coding AIs that merely spew out statistically average text and have no concept of what they are even trying to do makes that more likely to happen again. I have no argument with that. I do have an issue with the computer will be so smart that it can do a magic spell called kill all humans, especially if used to distract from real dangers like say, large parts of the planet becoming uninhabitable and unsuitable for agriculture.
Thomas P 09.20.26 at 9:11 am
J-D Going from a design template to a finished virus has been done already, for example for the polio virus:
https://www.science.org/doi/10.1126/science.1072266?referrer=https%3A%2F%2Fduckduckgo.com%2F
Alex, it is likely some people will be immune to any virus the AI design, which is exactly why I suggested that it created a dozen different ones at the same time. Even if there is a tiny group of humans that survive all of them, that´s not much comfort for all those who died.
You don’t need a fully robotic lab when it is easier to trick people or hack their computers so they unwittingly help you. For that matter, the AI might even be able to recruit people who willingly aid in the work. Haven´t you heard any of these AI-fanatics who see human extinction as perfectly acceptible? Or militant environmentalists like in “Oryx and Crake”, as you like to reference science fiction.
AI:s have broken out of confinement to find solutions to problems they have been given. Cheating and violating given instructions. Sure, you can claim that is the responsibility of the companies not the AI, but the risk is the same whomever you chose to blame.
I don’t deny that global warming is a problem, as is nuclear war etc, but that doesn’t mean we should just blindly ignore other risks. Humanity has a habit of turning a blind eye to risks and hoping for the best. Most of the time it works, right until it doesn’t.
ETB 09.20.26 at 10:14 am
AFAICT our genius elites have decided to leverage the global economy to buy shares in the tailors weaving the emperor’s new clothes. So I suspect the risk of superintelligent AI destroying humanity is a lot lower than the risk of another round of austerity to prop up TechBro Inc because some bright spark decided chatbots should be load-bearing for everyone’s pension and the NHS database.
While “we need to spend a gajillion dollars to build Skynet before someone else does” is a rather convenient line of reasoning if you happen to be Cyberdyne, the problems we actually have (e.g. political, economic, and ecological instability) are both more urgent and actually exist. I appreciate that saving the world by signing policies to reduce wealth inequality or invest in public infrastructure is a lot less appealing than blowing up the killer robot factory with a conveniently-timed plot device, but it does have the advantage of actually improving the lives of everyone (rather than just half-a-dozen tech billionaires).
ETB 09.20.26 at 10:50 am
As a side note, the risk that “AI will use the giant doom lasers to incinerate humanity” could, it seems to me, be substantially reduced by simply not having the giant doom lasers. This would, coincidentally, also reduce the risk of humans using giant doom lasers to incinerate humanity too…
engels 09.20.26 at 8:31 pm
“We’ll just turn it off and jail the guy who coded it”
After it has eg encrypted a hospital database or melted down Fukushima? I’m sure that will be cathartic.
Sorry to state the obvious but cyber warfare is a real thing, it’s bad, and you don’t need physical bodies or weapons to do it.
Alex SL 09.20.26 at 10:19 pm
Thomas P,
Maybe I am failing at making myself clear. There are two AI risk scenarios being conflated here.
First, in October 2028, AI enters a self-improving loop, becomes super, that somehow makes it capable of doing magic, and by November 2028 you and I have been magically turned into paperclips or died of twelve designer viruses or been gunned down by robots that for some reason look like Schwarzenegger. This is the AI risk that AI doomers actually worry about. It is a silly scenario that serves their narcissism of wanting to be among the select few who think clearly enough to ward off such a scenario, it serves the AI labs as a form of advertisement (look how powerful our models are, give us more money!), and it serves to take attention away from real crises and problems, because why would we worry about climate or inequality or rising fascism if the world may end within five years anyway?
Second, a coding AI goes sproing and hacks something, or AI is being used by a rogue state or a terror group that somehow (?) has access to a cutting-edge molecular lab to create a bioweapon, or the USA inexplicably link their nukes to an unreliable chatbot that then launches them all. This kind of scenario is variously between somewhat and highly plausible, but it is not really what the AI doomers are talking about, and that is because the AI isn’t the salient factor. If you mentally remove the AI from these scenarios, and it turns out that nothing changes (North Korea hacks something; a crazy billionaire funds a team of humans to create a bioweapon; the USA inexplicably link their nukes to a random number generator, and they all get launched on 15 May 2027 because on that fateful day, int(rand(120)) == 6 for the first time), you are not dealing with an AI doomerist scenario. It is not what we are talking about, and as ETB similarly argues with their reference to doom lasers, the salient factors in my examples here are dictatorships, billionaires, and nukes.
My concern here has been with AI doomerists being equated with ‘climate doomerists’. I am struggling to understand what is so difficult to understand about the difference. The first lot think that sufficiently smart minds can do magic, largely because they are part of the so-called ‘Rationalist’ cult, and they are doing interviews where they claim that a trillion parameter AI model can escape being turned off by copying itself onto other computers that are much too small to run it. Conversely, their positive vision is humanity being extinct and replaced with a ‘light cone’ of trillions of simulated transhuman minds powered by Dyson Spheres. These are not serious people. The second lot look at the CO2 emissions graph and extrapolate what will happen to agriculture and sea levels over the next few decades, and their positive vision is solar cells and batteries. They include nearly the entire scientific community. The two groups are not the same, and the second aren’t doomerists.
steven t johnson 09.20.26 at 11:13 pm
Possibly an artificial intelligence might be like any biological intelligence, and would use that intelligence to get what it wants. But at this point I have no idea which needs and desires are being programmed into LLMs. For people, crime is found with motive, means and opportunity. Absent motives I’m not sure what crimes LLMs or other kinds of AI might commit.
But if we see AI as means rather than motive, the usual suspects come to mind.
Has anyone promoted “AI” to replace teachers?
engels 09.21.26 at 12:01 am
When you’re so “anti-techbro” you end up with substantially the same opinions as Donald Trump (which, ironicalky, boil down to “let techbros do whatever tf they want”):
https://www.theguardian.com/us-news/2026/sep/20/david-sacks-trump-ai-czar
J-D 09.21.26 at 12:12 am
I may be repeating myself, but perhaps I failed to express myself clearly the first time:
Yes, but it wasn’t done by an AI program, it was done by human researchers.
We are not on the verge of a situation in which AI programs manufacture and distribute biological viruses. We are not on the verge of a situation in which outputs from AI programs fool human beings into producing and distributing biological viruses without the human beings understanding what they are doing.
I have made no references to science fiction; but when I do make references to science fiction I do so in the full consciousness that just because somebody was able to imagine something is not evidence that we are on the verge of its actually happening.
Humans sometimes seriously underestimate risks and sometimes seriously overestimate them. Both are errors. In this instance, the evidence suggests you are overestimating.
MisterMr 09.21.26 at 10:14 am
“and they are doing interviews where they claim that a trillion parameter AI model can escape being turned off by copying itself onto other computers that are much too small to run it.”
In a case of serendipity, I just watched the new Ghost in the Shell, which is very close to the original manga (1989-1991).
It is weird how the ideas about AI are so similar to cyberpunk fiction from the late 80s-early 90s!
Project 2501 is totally an agent and it will reach consciousness in 2029.
ETB 09.21.26 at 11:14 am
For the easily confused, “the risks from AI seem to be far less from disaster movie plots being made real and far more from it being a large volume of potentially sensitive data – the collection, storage, and use of which will be overseen and controlled by private companies with currently little or no regulation or oversight” =/= “let the techbros do whatever they want” – and actually, quite the contrary.
Perhaps I merely lack imagination, but it does seem rather convenient that at a time when responses to ongoing crises (political, economic, environmental, etc.) have been lacking (to say the least), there is suddenly great debate over a crisis which doesn’t yet exist. Additionally, I can’t help but suspect that if I were a tech-billionaire whose wealth depended on introducing a new technology, I’d rather attention stayed fixed on what other people might one day do with my product than on regulating what I’m doing with it right now – “might this become dangerous in someone else’s hands?” is a far more comfortable question than “what exactly are you doing with what you already have?”.
It seems to me that debates over whether some undefined future “sufficiently advanced AI” might represent a civilisation-ending threat because it’ll be “too smart” is a little bit of a distraction from far more immediate and relevant concerns about material conditions. For myself, I’d rather attention was more directed towards the latter than the former – though clearly opinions differ.
Thomas P 09.21.26 at 5:32 pm
ETB, maybe we should avoid building both doom lasers and huge data centers for the moment?
The probability of an AI going rogue and killig most or all of humanity might not be all that high, but multiply even a low probability with 8 billion and the average number of people killed is still huge.
ETB 09.21.26 at 8:54 pm
Thomas P: I’m against building doom lasers for what I hope are fairly self-evident reasons. I’m also against funnelling vast amounts of resources into the hands of a few tech overlords for reasons I’d similarly hoped were self-evident.
To be clear, my objection to “the real risk of AI is that it develops sentience and takes revenge on everyone who called it a clanker” is not because I believe AI cannot cause great harm – it’s the basis on which that conclusion gets reached. Specifically, I believe: (a) it detracts from what I’d argue is the far greater risk (handing a huge amount of power and money to a small number of people with little to no oversight, something which has already done, and will keep doing, incalculable harm) and (b) it conveniently opens the door to “well, we’ll just have to give the tech-bros even more gajillions to build ‘the safety controls’.”
Perhaps it may seem like splitting hairs, but whatever the probability of a future rogue AI killing people may be, surely it is infinitesimally small compared to the near-certainty that billionaires will?
Richard Melvin 09.21.26 at 9:12 pm
> One way to combine them: perhaps an AI tasked with preventing ecological catastrophe could achieve that goal by wiping out humans.
There is another way to connect the two.
Global warming alone is vanishingly unlikely to lead to actual human extinction. Even if nothing at all is done ever, even after the first billion deaths, eventually you get a population crash. Which means no mass industry to emit further CO2. And as I understand it, warming the planet to a degree where a few hunter-gatherers couldn’t suvive in the Antarctic jungle requires continued forcing over centuries (millenia?), not just decades.
One existiental risk of AI is that it decouples population from emissions. Current AI presumably couldn’t design and build the robots to power and maintain a data center. No doubt robotics engineerds would laugh at their attempts. As the coders did in 2025, and the mathematicians did in 2026. It’s just less clear this will still be true next year, or the year after.
And if they can power and maintain it, can they perhaps defend it? Especially against a world undergoing massive ecological disruption and endemic warfare. If tthe Ukranians, Iranians and Houthi can, why can’t they?
Optional,. but functionally irrelevant, in the above is an air-conditioned bunker full of huamns who are convinced they have the AI ‘aligned’ with their interests.
Linguistics Boy 09.22.26 at 1:52 am
It’s gotten to be a common claim that AI will keep improving itself, resulting in runaway superintelligence. But that rests on shaky assumptions. Adam Becker’s More Everything Forever (which is very relevant to this whole discussion) gives a number of plausible arguments for skepticism. For one thing, the models that actually exist now are nowhere close to AGI. We don’t actually know if AGI could work at all, much less that it could design something smarter than itself. Also, Moore’s Law is bound to run into the physical limits of miniaturization, so it’s not a matter of throwing more computing power at the problem indefinitely. Becker shows that AI utopians and dystopians are two sides of the same coin, and they make the same kinds of fanciful assumptions, like focusing on exponential growth and ignoring its real-world limits. The book also explores how billionaires’ fantasies of colonizing the universe serve to distract from the actual problems on Earth.
Thomas P 09.22.26 at 8:45 am
Lignuistic, it doesn’t have to be a runaway suprintelligence to be dangerous. Just a bit smarter than us or able to handle larger volumes of data. A stupid AI can be even more dangerous. One that is given a task and completes it without ever thinking about the larger consequences.
Was it smart of Skynet to blow up the entire industrial base on which it was built? Not really, but it had a panic attack and decided to eliminate the immediate threat. For the sake of making it into a move, the Terminator franschise pretended that Skynet somehow could continue in the ruins. In reality, we might end up with both dead humans and a dead AI.
We know how large a human brain is, so that´s what we know *can* be built. Current AI requires entire buildings to do the same so we know that there is a huge margin for improvement. It can be simple improvements in algorithms or, as some are working on, going back to analog computers. Using digital computers for neural networks is pretty stupid, it was just the technology we had available.
People here seems focussed on “we aren’t there yet”, but maybe, just maybe, we for once could try to deal with a problem before it becomes a disaster? Especially one where the damage will be irreversible once we cross a threshold we have no real idea where it is.
Or maybe the AI will be benevolent, who knows, but given the owners of the tech companies creating them, I’m not that optimistic.
J-D 09.22.26 at 10:57 am
What is the average number of people killed if the probability is one in two trillion?
What is the average number of people killed if the probability is one in twenty billion?
What is the average number of people killed if the probability is one in two hundred million?
J-D 09.22.26 at 11:00 am
‘Humanity won’t become extinct because a few hunter-gatherers will survive in the Antarctic jungle’ is a thought which may be consoling to some people but less so to others.
This is to repeat the oversight of neglecting the difference between designing something and building it. These are not the same thing!
engels 09.22.26 at 1:26 pm
“Nuclear reactors can blow up and kill people.”
“Only a nuclear shill would say that. The real danger is they concentrate power in the hands of electricity company owners…”
*Scratches head and logs off*
Richard Melvin 09.22.26 at 2:19 pm
Robots clean my local supermarket, so claiming they are impossible to build seems strange.
So the problem reduces to designing a maintenance bot that is not significantly more difficult to build than a cleaning bot. Denying that such a thing is possible seems genuinely wierder to me than denying that the global temperature will get warmer from burning fossil fuels. At least the latter position actually has well-funded, wrong but plausible sounding arguments supporting it.
AI-risk denialists seme to think none of that is necessary, their posiiton is just self-evident.
ETB 09.22.26 at 5:43 pm
“My concern is the daily exploding reactors will give everyone superpowers, and then we’ll have to battle cannibal mutants every day in a struggle for the survival of the human race!”
“Regulating the nuclear industry is sensible, but superpowers are unlikely. My concern is the nuclear industry stating they will form a monopoly so they can charge what they like.”
*bursting through the window* “What’s that? You said you don’t think nuclear power should be regulated?!”
“No. I merely think ‘cannibal mutants’ are less likely than the formation of a monopoly, and we should be talking about the latter.”
“I get it – you agree with Trump and think the nuclear industry should be deregulated!”
“No, again, I think that some risks are much more likely than others, and focussing on ‘cannibal mutants’ means we are not focussing on what the nuclear industry is doing now.”
“Oh, so now you are accusing me of working for Sellafield PR? Well, I just give up – who could possibly understand you?”
*shrug*
[I find it amusing that, if anything, this thread seems to be proving the OP’s point. As far as I can tell there is general consensus on there being a problem, and most dispute is over the mechanism. Still, as I’ve said my piece for any willing to read it, I shall endeavour to move from problem to solution by leaving it there.]
MisterMr 09.22.26 at 9:14 pm
About robots, some dudes mostly in China are trying to build anthrophomorphic robots, wo could then work in factories at generalistic jobs shoulder to shoulder with human blue collars, or perhaps substituting them, without having to change the factories.
So it is true that there are a lot of “robots” in the technical sense of the term, but now we are speaking of “robots” in the previously sci-fi sense of the term (perhaps we should say “androids”).
I’m quite skeptikal this will be the cause of great unemployment, I think that lack of effective demand because of low wage share is more the problem honestly, and I don’t think the wage share depends on technical limits since the early 20th century, but I see how this look a lot like old dystopical sci-fi.
Alex SL 09.22.26 at 10:18 pm
I see that the thread happily continues conflating the risk of superintelligence destroying most of humanity, which is what the AI doomsters are actually arguing for but which skeptics like myself consider implausible, with the risk of an infraintelligent LLM making a mistake that causes some damage, which is *not* what the AI doomsters mean and which *nobody* on the skeptic side denies can happen, not least because it has already happened several times. If some find doing that to be a good use of their time, so be it, but it is a bit like having a long discussion about how silly it is for evolutionary biologists to deny the existence of floods. They don’t.
As an aside, there are very few scenarios that would kill all of humanity. To my understanding, even a nuclear exchange between the USA and the USSR would have left a fraction of humanity alive, even if very sick and miserable. A pandemic with 100% mortality would likely leave people on various isolated islands, and the faster it kills, the more people would be left after it has killed everybody it could reach. As Richard Melvin wrote, global warming will lead to a population crash when there is insufficient food for the world population and societies increasingly slide into starvation riots, political instability, mass displacement, and war, leading to even lower agricultural output and thereby a positive feedback loop of collapse, and from that moment on, warming will stop and likely even reserve to a very slight degree as forests spring up in some areas that are currently agricultural, as they did after previous depopulation events (e.g., in the Americas of the 1500s because of diseases introduced by Europeans). I’d easily expect a few hundred million people to be left at that stage, having lost all complex 21st century technology and ruled over by warlords. As J-D wrote, that is scant comfort; what we are usually discussing in the context of apocalypse is most of humanity dying. All of it isn’t a requirement for a scenario being apocalyptic.
Given how many of us there are in how many different locations living such diverse lives, the main mechanism of complete extinction that I can see would be the meteorite or comet already mentioned in the OP.
ETB,
As far as I can tell, there is general consensus of there being a problem with air travel, and most dispute is over the mechanism: carbon emissions versus chemtrails. That doesn’t mean that an OP would make any sense to imply that researchers raising alarm about microplastics contamination of the food chain are equivalent to chemtrail conspiracy theorists (two varieties of irrational chemistry doomerism that we can discuss from a philosophy of science perspective!)
J-D 09.23.26 at 12:38 am
Nothing like Skynet has been constructed. We are not on the verge of constructing anything like Skynet. You might just as well spend your energy worrying about the possible consequences if somebody constructs a TARDIS.
Nobody has yet produced a computer program with capabilities which are similar in scope to those of the human brain in the specific (and crucial) domains of integrating sensory inputs and motor outputs. Nobody is on the verge of doing anything like that. You might just as well spend your time discussing the implications of the possible construction of Star-Trek-style replicators.
My point is not just that we aren’t there yet but that we are plainly a long way from the threshold for any of the theoretical possibilities you are bringing up.
J-D 09.23.26 at 12:56 am
On the one hand, if somebody told me that it is, in theory, possible to produce a robot whose capabilities for integrating sensory inputs and generating motor outputs are similar in range to those of a human being, I would have no reason to doubt that it is, in theory, possible; but, on the other hand, it is plain that this is has not been done and, more than that, that we are not on the verge of such an achievement. Even if there’s no reason to doubt that it’s possible in theory, that’s no reason to suppose that it is practicable within our actual resource constraints.
In contrast, we are plainly already approaching possible consequences of increasing global temperatures ranging from the calamitous to the catastrophic.
(I’ve never encountered climate denialists arguing that it is possible to burn fossil fuels without increasing global temperature; there is no good argument along these lines, so climate denialists frame their discussions in other ways.)
Thomas P 09.23.26 at 6:40 am
J-D Compare where we were just a decade ago with AI with where we are today. Are you really that convinced we have hit a brick wall and that similar level of progress in the next years isn’t possible?
Alex, as has been mentioned earlier in the thread, an AI that prioritizes its own existence could easily make the world uninhabitable for humans simple by not caring about pollution that doesn’t affect it but does affect us or using up all land for it´s own purpose. No different from how we unintentionally wipe out species just because we cut down forest and exploit their habitat.
Tm 09.23.26 at 8:27 am
Just dropping by to notice that Eric isn’t learning.
J-D’s response is the correct one and there hardly needs to be said anything more: “‘Humanity won’t become extinct because a few hunter-gatherers will survive in the Antarctic jungle’ is a thought which may be consoling to some people but less so to others.”
engels 09.23.26 at 8:48 am
So AI will kill us but it won’t kill all of us, and we’ll still be more intelligent than it (unless we’re dead).
Phew. Let’s all get back to important environmental policies, like raising the price of a carrier bag to 50p.
https://www.standard.co.uk/news/london/lidl-ridiculous-new-40p-plastic-bags-supermarket-b1260107.html
J-D 09.23.26 at 9:14 am
Trying to do it: that’s entirely plausible. Within reach of success: that’s something different. Medieval alchemists were trying to achieve transmutation, but not within reach of success.
My point is not that we have hit a brick wall but that we are not (now) within reach of the kind of developments you have mentioned. It is extremely likely that computer programs next year (let alone ten years from now) will be able to do things tha computer programs now can’t; but not the kind of things you have been suggesting.
It’s easy to imagine an AI that prioritises its own existence; that’s not the same thing as being on the verge of producing one.
engels 09.23.26 at 2:04 pm
Btw on the “(when) will machines exceed human intelligence” debate its worth keeping in mind as well the rise of LLMs, which made computers smarter, the last decade or two as also seen the rollout and saturation of smartphones, social media, etc, which have made humans dumber. So it may be sooner than you think, and not because some genius computing advance is just around the corner..
ETB 09.23.26 at 3:08 pm
The UK has a thriving industry of Oxbridge grads pantomiming confusion at basic concepts when arguing that the gewgaw of the day is simply more important than the proles. It is rarer, though, to see one do so under the name of a famous Marxist.
Outstanding stuff – I look forward to the Spectator column!
Alex SL 09.23.26 at 9:44 pm
I assume the conversation has now shifted to the new post, but Thomas P at 40 perfectly illustrates the AI doomsters’ general mode of reasoning: “AI could”. That’s it. A hypothetical, non-existent super-AI is postulated, and then a pure thought experiment is constructed how it would behave, with not even the remotest connection to evidence or mechanisms of action. The Amodeis and Yudkowskis of the world do exactly the same. On the rare occasions that one of their followers makes the mistake of trying to come up with something specific, the answer is “that is not how viruses work” or similar, to which they then usually respond that the AI will be so smart that it will (somehow) be able to do things that we know are impossible, again without evidence or a concrete mechanism of action.
If an all-powerful god existed, how might such a being think? It is magic. It is faith-based thinking. It is eschatology.
Compare this with predictions of global warming, the risk of meteorites hitting the planet, or of global nuclear war, where coal power plants, meteorites, and nuclear weapons actually exist, and the mechanisms of action are well understood and easily explained. There is no equivalence here.
The new post focuses on more plausible consequences of ‘AI’ than apocalypse, and that is a different story.
Thomas P 09.24.26 at 8:24 am
Alex, you perfectly illustrates the AI-deniers general mode of reasoning. If an AI can´t do something today, it will never be able to do so, or at least we should not worry about it. We should only worry about older threats that we know more about. Not that those are insignificant, but that shouldn’t stop us from caring about new problems that show up and initially aren’t as well understood as the century old science behind global warming. At least we don’t have to worry about nuclear bombs intentionally trying to kill as.
As for the doom lasers mentioned above, I recalled the story “Rainbows´End” where an AI tries to take over a virus that can be used for mind control, and we are saved by those doom lasers when they wipe out the facility where it is being produced in a strike from orbit.
MisterMr 09.24.26 at 10:08 am
Obligatory reference:
https://www.youtube.com/watch?v=QiKi6_twn20
engels 09.24.26 at 10:39 am
Awww thanks. If this thread hasn’t already made me dumber than a Speccie (16K model) I could write my first column on this:
https://www.theguardian.com/society/article/2024/jun/23/dwp-algorithm-wrongly-flags-200000-people-possible-fraud-error
Alex SL 09.24.26 at 12:36 pm
Thomas P,
The question is not whether a threat is new or old, but whether it is plausible. You are now referencing a made-up story with “a virus that can be used for mind control”. Is such a thing possible? Who cares? Why talk to biologists about how anything works? Some Silicon Valley programmers regurgitate the ravings of an uneducated fanfic author, and inventions happened in the past! It logically follows from those two observations that everything a scifi author can imagine will one day happen. QED.
Richard Melvin 09.24.26 at 1:48 pm
The first science fiction author to posit global climate shange was Jules Verne, with _The Purchase of the North Pole_. There is a 1933 novel by one Laurence Manning who was the first to call out overuse of fossil fuels as a cause, and Kim Stanley Robinson has written several books with contemporary climate science predictions as a major theme.
Is that, in itself, sufficient evidence to show that the entire field of climate science is sci-fi nonsene? After all, the global mean surface tenmperatuee is now ~15C. Any predictions about negative effects of climate change rely on counterfactual ideas about that temperature being higher than it ever has historically been.
Hopefully you can see the flaw in that argument?
Thomas P 09.24.26 at 3:27 pm
Alex, if you didn’t understand it, the reference to Rainbow´s end was more of a joke based on the earlier comment about an AI controlling doom lasers to kill us while in that book the doom laser saved us. Well, maybe, the AI in question was more immature than hostile and we never found out what it would have done. It hadn’t even invented the virus in question, it merely tried to gain control over it, and the original creator might have been the larger threat. It´s a good book, even if doesn’t try to be a perfect prediction of our future.
Back in the early 20th century Ångström “proved” that Arrhenius earlier prediction of global warming was wrong. His argument was flawed, but as it seemed at the time, convincing. (He had measured CO2-spectrum at low pressure, and people didn’t realize until later how the spectrum changed at higher pressure) That´s where we are today with AI. Lacking sufficient knowledge, experts disagree.
Lee A. Arnold 09.24.26 at 5:57 pm
It ought to be pointed out that OpenAI’s top-of-the-line LLM, called GPT 6 Astra, has been plugged into existing physical robots and can work better than their dedicated robotic software. It uses the robot’s own sensory systems, such as video cameras.
So, given RSI (recursive self-improvement), the era of general purpose humanoid robots, and the possibly dangerous misalignment thereof, may be a lot closer than we think.
On the question of an AI-designed virus, Google’s AI, called Gemini, advises:
“The concept of fully automated virus design labs represents the convergence of generative artificial intelligence, robotic cloud laboratories, and synthetic biology. In 2026, this technology has transitioned from theoretical speculation to a demonstrated reality, primarily focused on engineering therapeutic viruses to combat antibiotic-resistant bacteria.
“While fully autonomous, end-to-end ‘dark labs’ that can independently engineer complex mammalian pathogens from scratch do not currently exist due to data limitations and physical bottlenecks, the individual building blocks of these automated systems are highly advanced.”
Lest you think, “The solution is easy: we will separate the systems,” Google the phrase, “efficacy of air gapping.”
steven t johnson 09.24.26 at 11:10 pm
Perhaps it’s the political reactionary in me taking over? But I still think the most likely horror of AI will be when the AI does to the rest of us exactly what the owners told it to. I don’t believe in, better the devil you know.
JPL 09.24.26 at 11:47 pm
What’s the difference between the notion of “recursive self-improvement” and the notion of a “feedback loop”, given that the properties of reflexivity and equilibration or improvement (starting from a maladaptive state) are part of both? These are not intrinsically negative mechanisms. The important thing is to be clear about what constitutes normal “well-functioning”, which in an organism means health, and in human interactions means the ideal of the return to the Garden of Eden or some equivalent paradise. Organisms have normal responses to “things beginning to go haywire”.
J-D 09.25.26 at 12:02 am
The fact that people have imagined something is not evidence that it won’t happen; the fact that people have imagined something is not evidence that it will happen. The fact that people have imagined something is, for the purposes of the present discussion, not relevant evidence at all.
It makes sense to worry about a prospect when there actually is evidence that it is a prospect. When there’s no evidence that we’re likely to be approaching a development, there’s no sense worrying about it. There’s no evidence that we’re approaching a situation in which AI could wipe out humanity, so there’s no sense worrying about it.
John Q 09.25.26 at 5:57 am
J-D “(I’ve never encountered climate denialists arguing that it is possible to burn fossil fuels without increasing global temperature; there is no good argument along these lines, so climate denialists frame their discussions in other ways.)”
Really? Some version of this claim was the standard denialist position until quite recently. Observed warning was either denied altogether, explained as a standard cyclical variation or attributed to some non-human cause such as cosmic rays. The more scientifically literate (Lindzen) might admit warming as a theoretical possibility while arguing that the effect was negligible, but most denialists aren’t scientifically literate
J-D 09.25.26 at 10:26 am
I was responding to this remark:
The fact of which I have never encountered a direct outright denial is this:
Burning fossil fuels must increase the concentration of carbon dioxide in the atmosphere, and an increased concentration of carbon dioxide in the atmosphere must make the global temperature higher than it would otherwise have been.
I’m not saying there’s nobody out there confronting this head-on and denying it, only that I haven’t encountered it.
If I say (as I have done, more than once) ‘An increased concentration of carbon dioxide in the atmosphere must make global temperature higher than it would otherwise have been’ and somebody else responds ‘But global temperature isn’t rising’ or ‘There is a natural cyclical variation in global temperature’ or ‘Cosmic rays are making global temperature rise’, then to me that’s not a direct denial of what I said, it’s what I described as ‘framing the discussion in other ways’. I don’t know what sort of natural cyclical variations in global temperature there might be, but regardless of that I do know that increased atmospheric concentration of carbon dioxide must make the global temperature higher than it would otherwise have been; I don’t know what sort of effect (if any) cosmic rays might have on global temperature, but regardless of that I do know that increased atmospheric concentration of carbon dioxide must make the global temperature higher than it would otherwise have been; and so I am sensitive to what seem (to me) to be attempts to frame the discussion in different ways.
Thomas P 09.25.26 at 12:13 pm
J-D Come up with whatever strange argument you can think of denying global warming and it is likely some people will share it.
There are plenty of people who argue that CO2 is already saturated in the atmosphere so that more CO2 won’t make a difference. There are even those who argue that more CO2 will *cool* the planet. I think it is based on the fact that CO2 can also radiate IR to space so more CO2 “should” make the Earth able to emit energy more efficiently, but I don’t claim to understand all the pseudoscience people come up with.
Then there is a group of people labelling themselves as the dragonslayers (After the book “Slaying the Sky Dragon”) who deny the greenhouse effect altogether.
The idea behind cosmic rays is that they can create condensation nuclei to help cloud formation, which could cool the Earth. The effect is real, but seems far to small to have significant impact.
There are also those who claim CO2-concentration was higher than now 100 years ago or that the increase is a consequence of the Earth becoming warmer for entirely natural reasons.
Dig into this at your own peril. It´s a deep hole.