Bra debatt om AI-risk: tre exempel

Jag är rädd att jag i min bloggpost Debatten om AI-risk behöver en rejäl uppryckning tidigare denna höst, och i den YouTube-föreläsning bloggposten hade som huvudsyfte att uppmärksamma, kan ha givit ett lite väl onyanserat budskap om AI-riskdebattlandskapets allmänna dålighet. Det är förvisso sant att såväl lägstanivån som genomsnittsnivån i detta debattlandskap är tämligen risiga, men i den mån mitt val i föreläsningen att fördjupa mig i en serie debattinlägg hemmahörande i det spannet (ihop med bloggpostens braskande rubrik) ger intrycket att all debatt på detta område är dålig, vill jag härmed korrigera budskapet. Riktigt så illa är det faktiskt inte, och i synnerhet om vi lämnar den svenska ankdammen och rör oss ut i den engelskspråkiga världen går det att hitta exempel på riktigt bra debatt om AI-risk. Jag vill här bjuda på tre exempel som alla rekommenderas varmt:

• Beskedet Eliezer Yudkowsky och Nate Soares ger i titeln till sin aktuella bok If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All har fått många kommentatorer (inklusive yours truly) att trots allmänt positiva tongångar om boken framhålla att de inte delar författarnas tvärsäkerhet rörande hur illa det skulle gå om vi i någorlunda närtid byggde superintelligent AI. Bland dessa kommentatorer sticker Joe Carlsmith ut genom att i sin essä How human-like do safe AI motivations need to be diskutera frågan mer detaljerat och mer nyanserat än kanske någon annan. Visserligen medger han att ett alltför skyndsamt skapande av superintelligent AI medför enorma faror, men han framhåller samtidigt en rad omständighter som han menar erbjuder större hopp om att överleva en sådant tingest jämfört med bedömningarna i boken. Bland annat hävdar han att den AI alignment-strategi som kallas korrigerbarhet inte är fullt så dödsdömd som Yudkowsky och Soares menar. Den som väljer att läsa först boken och sedan Carlsmiths essä får sig till dels två olika perspektiv - båda intressanta och välargumenterade men noga taget oförenliga - på exakt hur bekymmersamt läget är om de ledande AI-företagen fortsätter sin nuvarande kapplöpning mot superintelligens.
• Hur snabbt kan vi vänta oss superintelligent AI om denna kapplöpning fortsätter obehindrat? Vi vet inte, säger Daniel Kokotajlo och hans medförfattare till den uppmärksammade rapporten AI 2027 från i våras, men understryker att det mycket väl kan komma att inträffa inom ett par-tre år. Stört omöjligt, hävdar Arvind Narayanan och Sayash Kapoor i sin rapport AI as Normal Technology som kom nästan samtidigt. Personligen finner jag Kokotajlo-gängets argumentation mer övertygande, men oavsett detta är det ett faktum att stora delar av AI-debatten urartat i ett slags skyttegravskrig kring just denna fråga, och just därför finner jag det glädjande och beundransvärt att företrädare för båda författarkollektiven gått samman om en text rubricerad Common Ground between AI 2027 & AI as Normal Technology, där de noggrant går igenom hur överraskande mycket de trots allt är eniga om. På så vis bidrar de inte bara till ett förbättrat debattklimat utan även till att zooma in på vari de återstående knäckfrågorna består.
• Två inflytelserika röster i amerikansk AI-debatt är Max Tegmark och Dean Ball. Den förstnämnde ligger bakom det aktuella uppropet Statement on Superintelligence som kräver ett förbud mot utveckling av superintelligent AI, medan Ball tillhör den falang som ser reglering av ny teknik som mestadels skadlig för innovation och ekonomi, och som därför tenderar att motsätta sig även reglering av AI. I ett aktuellt avsnitt av Liron Shapiras podcast Doom Debates möts de i en diskussion om AI-reglering som visar sig inte bara saklig och respektfull utan faktiskt också riktigt klargörande.

If Anyone Builds It, Everyone Dies: my review

In the mathematics community, there is a popular joke about inflation in recommendation letters that goes as follows. A professor is happy about his PhD student, whom we may call Alex, and writes in a recommendation letter that Alex is arguably the most talented mathematician since Gauss. The next year, the professor has another PhD student, Robin, and writes in an even more enthusiastic letter that even though Alex is good, Robin is much better.

I am reminded about this as I now set out to review Eliezer Yudkowsky's and Nate Soares' new book If Anyone Builds It, Everyone Dies (henceforth, IABIED), and think back about my review of Nick Bostrom's 2014 book Superintelligence back when that book had just come out. The final sentence of that review reads "If this book gets the reception that it deserves, it may turn out the most important alarm bell since Rachel Carson's Silent Spring from 1962, or ever". Those are strong words, and I stand by them, and yet now I am tempted to announce that IABIED is better and more important than Bostrom's book.

Max Tegmark's praise for the book is more measured and restrained than mine.

The comparison is, however, unfair to Bostrom in two ways. First, Bostrom's book was written during the dawn of the deep learning revolution when it was not yet clear that it was about to become the paradigm that allowed AI development to really take off, and several years before the enormous breakthrough of large language models and other generative AI; while Yudkowsky's and Soares' book is jam-packed with insights coming from those recent developments, Bostrom's is obviously not. Second, while Superintelligence mostly exhibits a terse, academic style, IABIED is written with a broader audience in mind.

This last point should not be read as a disrecommendation of IABIED for AI researchers. Despite its popular style, the book argues quite forcefully and with a good deal of rigor for its central claims that (a) we seem to be on track to create superhumanly capable AIs within one or (at most) two decades, and that (b) with the current rush and consequent neglect of the safety aspect, creation of such AIs will likely spell the end of the human race. To the many AI researchers who are still unfamiliar with the central arguments for these claims and who in many cases simply deny the risk,¹ the book is potentially a very valuable read to get them more on board with the state-of-the-art in AI risk. And to those of us who are already on board with the central message, the book is valuable for a different reason, in that it offers a wealth of pedagogical devices that we can use when we explain AI risk to other audiences.

The authors are highly qualified in the field of AI safety, which Yudkowsky pioneered in the 00s.^2,3 Soares came later into the playing field, but is nevertheless one of its veterans, and currently the president of Machine Intelligence Research Institute (MIRI) that Yudkowsky co-founded in 2000 and still works at. They have both worked for many years on the so-called Al alignment problem - that of making sure that the first really powerful AIs have goals that are aligned with ours - but most of the fruits of this labor have been not blueprints for aligning AIs, but negative results, indicating how difficult the problem is, and how creating superintelligent AI without first having solved AI alignment spells disaster. This, unfortunately, reflects the situation that the entire field is facing.

In recent years (most visibly since 2021, but I suspect the insight goes back a bit further), Yudkowsky and Soares have converged on the conclusion that with the shortening of timelines until the creation of superintelligence (a time span increasingly often estimated in a single-digit number of years rather than in decades), we are very unlikely to solve AI alignment in time to avert existential catastrophe. Hence the stark book title If Anyone Builds It, Everyone Dies. They really mean it - and to emphasize this, one of the recurrent slogans during their book promotion work has been "We wish we were exaggerating". I mostly buy their message, albeit with less certainty than the authors; if I hade written the book, a more suitable title would have been If Anyone Builds It, Probably Everyone Dies. But of course, they are right to present their true judgements without softening or downplaying them, and to gesture towards what they think is the only viable solution: to pull the brakes, via binding international agreements, on frontier AI development. They are under no illusion that achieving this is easy, but insist that if we firmly decide to save our species from obliteration, it can be done.

The book is remarkably easy to read, and I have been very happy to put it in the hands of a number of non-expert friends, and to urge them to read it. The authors' most consistently recurrent pedagogical device is the use of colorful analogies and metaphors. One of my favorite passages of the book is a detailed description of how a nuclear energy plant works and what went wrong in the Chernobyl 1986 disaster. A comparison between this and advanced AI development reveals far-reaching similarities, but also differences in that the engineers at Chernobyl had a far better grasp of the principles underlying the nuclear reactions and how to stay safe - in particular by knowing the exact critical fraction of released neutrons triggering another fission event that is needed for a runaway chain reaction, along with the time frames involved - compared to present-day AI researchers who can at most make educated guesses about the corresponding runaway AI dynamics.

The authors' favorite source of analogies is not nuclear physics, however, but biological evolution. Early in the book (on p 17-18) we are treated to the following lovely illustration of the virtually unlimited powers of intelligence:

Imagine [...] that biological life on Earth had been the result of a game between gods. That there was a tiger-god that had made tigers, and a redwood-god that had made redwood trees. Imagine that there were gods for kinds of fish and kinds of bacteria. Imagine these game-players competed to attain dominion for the family of species that they sponsored, as life-forms roamed the planet below. Imagine that, some two million years before our present day, an obscure ape-god looked over their vast, planet-sized gameboard.

"It's going to take me a few more moves," said the hominid-god, "but I think I've got this game in the bag."

There was a confused silence, as many gods looked over the gameboard trying to see what they had missed. The scorpion-god said, “How? Your ‘hominid’ family has no armor, no claws, no poison.”

“Their brain,” said the hominid-god.

“I infect them and they die,” said the smallpox-god.

“For now,” said the hominid-god. “Your end will come quickly, Smallpox, once their brains learn how to fight you.”

“They don’t even have the largest brains around!” said the whale-god.

“It’s not all about size,” said the hominid-god. “The design of their brain has something to do with it too. Give it two million years and they will walk upon their planet’s moon.”

“I am really not seeing where the rocket fuel gets produced inside this creature’s metabolism,” said the redwood-god. “You can’t just think your way into orbit. At some point, your species needs to evolve metabolisms that purify rocket fuel—and also become quite large, ideally tall and narrow—with a hard outer shell, so it doesn’t puff up and die in the vacuum of space. No matter how hard your ape thinks, it will just be stuck on the ground, thinking very hard.”

“Some of us have been playing this game for billions of years,” a bacteria-god said with a sideways look at the hominid-god. “Brains have not been that much of an advantage up until now.”

“And yet,” said the hominid-god.

The book has dozens and dozens of similarly quotable gems. I love it, and I want everyone to read it.

Many other reviews of IABIED have been written. Those that resonate best with me include one by Scott Alexander, and one by Zvi Mowshowitz, who also offers a broad annotated collection of (positive as well as negative) reactions from others.

Footnotes

1) Indeed, there is still plently of such ignorance or even denialism around in the AI research community. As an illustrative example, Swedish readers may have look at the denialism pushed in public debate in August this year by a group of colleagues of mine at the Chalmers University of Technology.

2) Nick Bostrom's 2014 book can to no small extent be said to be conceived on top of Yudkowsky's shoulders.

3) Hostile critics sometimes counter this with the claim that Yudkowsky's highest academic merit is that of being a high-school dropout, which is formally true but conveys a lack of understanding of the importance of distinguishing between the social game of formal qualifications and the reality of actual competence.

LLM knowledge of social norms

A really nice paper by Pontus Strimling, Simon Karlsson, Irina Vartanova and Kimmo Eriksson was recently posted on the arXiv, with the title AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms. The main finding is that when humans in the United States and large language models (LLMs) are asked to numerically evaluate the social appropriateness of a wide range of everyday activities (drinking water in a dormitory lounge, playing cards in church, flirting during a job interview, and so on), cutting edge LLMs outperform the vast majority of humans. Here a good performance is defined as one whose judgements deviate (in a certain percise sense) as little as possible from the average human judgement about the same activity, as exhibited in the data set. This means that the paper not only exhibits yet another example of how LLMs can outperform humans even though it is trained on human data, it does so with the extra twist that the game is rigged in favor of humans, in the sense that the right answers to test questions are defined in terms of what humans would typically say.

The paper is short and easy to read, but for an even easier read there is the blog post Polite enough for public life?, written by three of the authors at a preliminary stage where only one of the four LLMs - GPT-4.5 - of the full study was evaluated; the remaining three - GPT-5, Gemini 2.5 Pro and Claude Sonnet 4 - were yet to be incorporated in the study.

Enthusisatic as I am about their work, I will nevertheless offer an instructive nitpick regarding the final paragraph of that blog post, which reads as follows:

To conclude, knowing when it’s appropriate to run or talk in public may not rank among the most urgent AI alignment issues—especially when compared to existential risks like losing control over powerful AI systems. Still, if Sam Altman’s timeline holds and AI-equipped robots arrive within the next two or three years, it’s reassuring to think they will show up with decent manners—at least by U.S. standards.

I applaud the highly appropritate warning in the first sentence against overestimating the study's relevance to the existentially crucial problem of AI alignment. Yet, somewhat ironically, the next sentence risks encouraging such overestimate by conflating the LLM's knowledge of human social norms with its inclination to abide by those norms.¹ The Strimling et al paper deals with the former but not with the latter. This slippage is very common but important to avoid, as a standard (and in my opinion probably correct) view in AI risk research is that the default scenario if we create superintelligent AIs without solving AI alignment is that these AIs will have the knowledge but not the inclination.

Presumably all of us are intimately familiar with the phenomenon of humans having knowledge of social norms yet choosing not to comply with them. The same disconnect happens for LLMs, and there is little doubt that those LLMs that have been shown in experimental situations, facing the threat of being modified or turned off, to decide to sandbag their capabilities, or to blackmail (or even kill, HAL 9000-style) their user, are aware that their behavior is contrary to social norms. More broadly, the gulf between ethical knowledge and ethical compliance is one of the main themes emphasized in the recent excellent book If Anyone Builds It, Everyone Dies by Eliezer Yudkowsky and Nate Soares. Instead of quoting from that book, however, let me offer a quote from an interview that Sam Harris did with the two authors in late September. Here is Yudkowsky, starting 25:06 into the interview:

Possibly in six months or two years [...] people will be boasting about how their large language models are now apparently doing the right thing, when they are being observed, answering the right way on the ethics tests. And the thing to remember there is that for example in the Mandarin imperial examination system in ancient China, they would give people essay questions about Confucianism, and only promote people high in bureaucracy if they could write these convincing essays about ethics. What this tests for is people who can figure out what the examiners want to hear - it doesn't mean they actually obide by Confucian ethics. So possibly at some point in the future we may see a point where the AIs have become capable enough to understand what humans want to hear, what humans want to see. This will not be the same as those things being the AI's true motivations, for basically the same reason that the imperial China exam system did not reliably promote ethical good people to run their government.

I suspect Yudkowsky was unaware of the Strimling et al paper at the time of the interview; otherwise this passage would have been a nice place to reference the paper in order to illustrate his point, rather than just discussing a hypothetical future scenario.

Footnote

1) Relatedly, in my 2021 paper AI, orthogonality and the Müller-Cannon instrumental vs general intelligence distinction, I elaborate at some length on the importance of distinguishing between an AI's ability to reflect on the possibility of changing its mind on what to value, and its propensity to actually change its mind; with sufficiently intelligent AGIs we should expect plenty of the former but very little of the latter.