Preventing an AI-related catastrophe

Why do we think that reducing risks from AI is one of the most pressing issues of our time? In short, our reasons are:

1. Even before getting into the actual arguments, we can see some cause for concern — as many AI experts think there’s a small but non-negligible chance that AI will lead to outcomes as bad as human extinction.
2. We’re making advances in AI extremely quickly — which suggests that AI systems could have a significant influence on society, soon.
3. There are strong arguments that “power-seeking” AI could pose an existential threat to humanity⁷ — which we’ll go through below.
4. Even if we find a way to avoid power-seeking, there are still other risks.
5. We think we can tackle these risks.
6. This work is neglected.

We’re going to cover each of these in turn, then consider some of the best counterarguments, explain concrete things you can do to help, and finally outline some of the best resources for learning more about this area.

If you’d like, you can watch our 10-minute video summarising the case for AI risk before reading further:

1. Many AI experts think there’s a non-negligible chance AI will lead to outcomes as bad as extinction

In May 2023, hundreds of prominent AI scientists — and other notable figures — signed a statement saying that mitigating the risk of extinction from AI should be a global priority.

So it’s pretty clear that at least some experts are concerned.

But how concerned are they? And is this just a fringe view?

We looked at four surveys of AI researchers who published at NeurIPS and ICML (two of the most prestigious machine learning conferences) from 2016, 2019, 2022 and 2023.⁸

It’s important to note that there could be considerable selection bias on surveys like this. For example, you might think researchers who go to the top AI conferences are more likely to be optimistic about AI, because they have been selected to think that AI research is doing good. Alternatively, you might think that researchers who are already concerned about AI are more likely to respond to a survey asking about these concerns.⁹

All that said, here’s what we found:

In all four surveys, the median researcher thought that the chances that AI would be “extremely good” was reasonably high: 20% in the 2016 survey, 20% in 2019, 10% in 2022, and 10% in 2023.¹⁰

Indeed, AI systems are already having substantial positive effects — for example, in medical care or academic research.

But in all four surveys, the median researcher also estimated small — and certainly not negligible — chances that AI would be “extremely bad (e.g. human extinction)”: a 5% chance of extremely bad outcomes in the 2016 survey, 2% in 2019, 5% in 2022 and 5% in 2023. ¹¹

In the 2022 survey, participants were specifically asked about the chances of existential catastrophe caused by future AI advances — and again, over half of researchers thought the chances of an existential catastrophe was greater than 5%.¹²

So experts disagree on the degree to which AI poses an existential risk — a kind of threat we’ve argued deserves serious moral weight.

This fits with our understanding of the state of the research field. Three of the leading companiess developing AI — DeepMind, Anthropic and OpenAI — also have teams dedicated to figuring out how to solve technical safety issues that we believe could, for reasons we discuss at length below, lead to an existential threat to humanity.¹³

There are also several academic research groups (including at MIT, Cambridge, Carnegie Mellon University, and UC Berkeley) focusing on these same technical AI safety problems.¹⁴

It’s hard to know exactly what to take from all this, but we’re confident that it’s not a fringe position in the field to think that there is a material risk of outcomes as bad as an existential catastrophe. Some experts in the field maintain, though, that the risks are overblown.

Still, why do we side with those who are more concerned? In short, it’s because there are arguments we’ve found persuasive that AI could pose such an existential threat — arguments we will go through step by step below.

It’s important to recognise that the fact that many experts recognise there’s a problem doesn’t mean that everything’s OK because the experts have got it covered. Overall, we think this problem remains highly neglected (more on this below), especially as billions of dollars a year are spent to make AI more advanced.⁴

2. We’re making advances in AI extremely quickly

Before we try to figure out what the future of AI might look like, it’s helpful to take a look at what AI can already do.

Modern AI techniques involve machine learning (ML): models that improve automatically through data input. The most common form of this technique used today is known as deep learning.

Probably the most well-known ML-based product is ChatGPT. OpenAI’s commercialisation system — where you can pay for a much more powerful version of the product — led to revenue of over $2 billion by the end of 2023, making OpenAI one of the fastest growing startups ever.

If you’ve used ChatGPT, you may have been a bit underwhelmed. After all — while it’s great at some tasks, like coding and data analysis — it makes lots of mistakes. (Though note that the paid version tends to perform better than the free version.)

But we shouldn’t expect the frontier of AI to remain at the level of ChatGPT. There has been huge progress in what can be achieved with ML in only the last few years. Here are a few examples (from less recent to more recent):

• AlphaStar, which can beat top professional players at StarCraft II (January 2019)
• MuZero, a single system that learned to win games of chess, shogi, and Go — without ever being told the rules (November 2019)
• GPT-f, which can solve some Maths Olympiad problems (September 2020)
• AlphaFold 2, a huge step forward in solving the long-perplexing protein-folding problem (July 2021)
• Gato, a single ML model capable of doing a huge number of different things (including playing Atari, captioning images, chatting, and stacking blocks with a real robot arm), deciding what it should output based on the context (May 2022)
• Midjourney V6 (December 2023), Stable Diffusion XL (July 2023), DALL-E 3 (August 2023) and Imagen 2 (December 2023), all of which are capable of generating high-quality images from written descriptions
• Sora (February 2024), a model from OpenAI that can create realistic video from text prompts
  *And large language models, such as GPT-4, Claude, and Gemini — which we’ve become so familiar with through chatbots — continue to surpass benchmarks on maths, code, general knowledge, and reasoning ability.¹⁶

If you’re anything like us, you found the complexity and breadth of the tasks these systems can carry out surprising.

And if the technology keeps advancing at this pace, it seems clear there will be major effects on society. At the very least, automating tasks makes carrying out those tasks cheaper. As a result, we may see rapid increases in economic growth (perhaps even to the level we saw during the Industrial Revolution).

If we’re able to partially or fully automate scientific advancement we may see more transformative changes to society and technology.¹⁷

That could be just the beginning. We may be able to get computers to eventually automate anything humans can do. This seems like it has to be possible — at least in principle. This is because it seems that, with enough power and complexity, a computer should be able to simulate the human brain. This would itself be a way of automating anything humans can do (if not the most efficient method of doing so).

And as we’ll see in the next section, there are some indications that extensive automation may well be possible through scaling up existing techniques.

Current trends show rapid progress in the capabilities of ML systems

There are three things that are crucial to building AI through machine learning:

1. Good algorithms (e.g. more efficient algorithms are better)
2. Data to train an algorithm
3. Enough computational power (known as compute) to do this training

Epoch is a team of scientists investigating trends in the development of advanced AI — in particular, how these three inputs are changing over time.

They found that the amount of compute used for training the largest AI models has been rising exponentially — doubling on average every six months since 2010.

That means the amount of computational power used to train our largest machine learning models has grown by over one billion times.

Epoch also looked at how much compute has been needed to train a neural network to have the same performance on ImageNet (a well-known test data set for computer vision).

They found that the amount of compute required for the same performance has been falling exponentially — halving every 10 months.

So since 2012, the amount of compute required for the same level of performance has fallen by over 10,000 times. Combined with the increased compute used for training, that’s a lot of growth.

Finally, they found that the size of the data sets used to train the largest language models has been doubling roughly once a year since 2010.

It’s hard to say whether these trends will continue, but they speak to incredible gains over the past decade in what it’s possible to do with machine learning.

Indeed, it looks like increasing the size of models (and the amount of compute used to train them) introduces ever more sophisticated behaviour. This is how things like GPT-4 are able to perform tasks they weren’t specifically trained for.

These observations have led to the scaling hypothesis: that we can simply build bigger and bigger neural networks, and as a result we will end up with more and more powerful artificial intelligence, and that this trend of increasing capabilities may increase to human-level AI and beyond.

If this is true, we can attempt to predict how the capabilities of AI technology will increase over time simply by looking at how quickly we are increasing the amount of compute available to train models.

But as we’ll see, it’s not just the scaling hypothesis that suggests we could end up with extremely powerful AI relatively soon — other methods of predicting AI progress come to similar conclusions.

When can we expect transformative AI?

It’s difficult to predict exactly when we will develop AI that we expect to be hugely transformative for society (for better or for worse) — for example, by automating all human work or drastically changing the structure of society.¹⁸ But here we’ll go through a few approaches.

One option is to survey experts. Data from the 2023 survey of 3000 AI experts implies there is 33% probability of human-level machine intelligence (which would plausibly be transformative in this sense) by 2036, 50% probability by 2047, and 80% by 2100.¹⁹ There are a lot of reasons to be suspicious of these estimates,⁹ but we take it as one data point.

Ajeya Cotra (a researcher at Open Philanthropy) attempted to forecast transformative AI by comparing modern deep learning to the human brain. Deep learning involves using a huge amount of compute to train a model, before that model is able to perform some task. There’s also a relationship between the amount of compute used to train a model and the amount used by the model when it’s run. And — if the scaling hypothesis is true — we should expect the performance of a model to predictably improve as the computational power used increases. So Cotra used a variety of approaches (including, for example, estimating how much compute the human brain uses on a variety of tasks) to estimate how much compute might be needed to train a model that, when run, could carry out the hardest tasks humans can do. She then estimated when using that much compute would be affordable.

Cotra’s 2022 update on her report’s conclusions estimates that there is a 35% probability of transformative AI by 2036, 50% by 2040, and 60% by 2050 — noting that these guesses are not stable.²⁰

Tom Davidson (also a researcher at Open Philanthropy) wrote a report to complement Cotra’s work. He attempted to figure out when we might expect to see transformative AI based only on looking at various types of research that transformative AI might be like (e.g. developing technology that’s the ultimate goal of a STEM field, or proving difficult mathematical conjectures), and how long it’s taken for each of these kinds of research to be completed in the past, given some quantity of research funding and effort.

Davidson’s report estimates that, solely on this information, you’d think that there was an 8% chance of transformative AI by 2036, 13% by 2060, and 20% by 2100. However, Davidson doesn’t consider the actual ways in which AI has progressed since research started in the 1950s, and notes that it seems likely that the amount of effort we put into AI research will increase as AI becomes increasingly relevant to our economy. As a result, Davidson expects these numbers to be underestimates.

Holden Karnofsky, co-CEO of Open Philanthropy, attempted to sum up the findings of others’ forecasts. He guessed in 2021 there was more than a 10% chance we’d see transformative AI by 2036, 50% by 2060, and 66% by 2100. And these guesses might be conservative, since they didn’t incorporate what we see as faster-than-expected progress since the earlier estimates were made.

All in all, AI seems to be advancing rapidly. More money and talent is going into the field every year, and models are getting bigger and more efficient.

Even if AI were advancing more slowly, we’d be concerned about it — most of the arguments about the risks from AI (that we’ll get to below) do not depend on this rapid progress.

However, the speed of these recent advances increases the urgency of the issue.

(It’s totally possible that these estimates are wrong – below, we discuss how the possibility that we might have a lot of time to work on this problem is one of the best arguments against this problem being pressing).

3. Power-seeking AI could pose an existential threat to humanity

We’ve argued so far that we expect AI to be an important — and potentially transformative — new technology.

We’ve also seen reason to think that such transformative AI systems could be built this century.

Now we’ll turn to the core question: why do we think this matters so much?

There could be a lot of reasons. If advanced AI is as transformative as it seems like it’ll be, there will be many important consequences. But here we are going to explain the issue that seems most concerning to us: AI systems could pose risks by seeking and gaining power.

We’ll argue that:

1. It’s likely that we’ll build AI systems that can make and execute plans to achieve goals
2. Advanced planning systems could easily be ‘misaligned’ — in a way that could lead them to make plans that involve disempowering humanity
3. Disempowerment by AI systems would be an existential catastrophe
4. People might deploy AI systems that are misaligned, despite this risk

Thinking through each step, I think there’s something like a 1% chance of an existential catastrophe resulting from power-seeking AI systems this century. This is my all things considered guess at the risk incorporating considerations of the argument in favour of the risk (which is itself probabilistic), as well as reasons why this argument might be wrong (some of which I discuss below). This puts me on the less worried end of 80,000 Hours staff, whose views on our last staff survey ranged from 1–55%, with a median of 15%.

It’s likely we’ll build advanced planning systems

We’re going to argue that future systems with the following three properties might pose a particularly important threat to humanity:²¹

1. They have goals and are good at making plans.

   Not all AI systems have goals or make plans to achieve those goals. But some systems (like some chess-playing AI systems) can be thought of in this way. When discussing power-seeking AI, we’re considering planning systems that are relatively advanced, with plans that are in pursuit of some goal(s), and that are capable of carrying out those plans.

2. They have excellent strategic awareness.

   A particularly good planning system would have a good enough understanding of the world to notice obstacles and opportunities that may help or hinder its plans, and respond to these accordingly. Following Carlsmith, we’ll call this strategic awareness, since it allows systems to strategise in a more sophisticated way.

3. They have highly advanced capabilities relative to today’s systems.

   For these systems to actually affect the world, we need them to not just make plans, but also be good at all the specific tasks required to execute those plans.

   Since we’re worried about systems attempting to take power from humanity, we are particularly concerned about AI systems that might be better than humans on one or more tasks that grant people significant power when carried out well in today’s world.

   For example, people who are very good at persuasion and/or manipulation are often able to gain power — so an AI being good at these things might also be able to gain power. Other examples might include hacking into other systems, tasks within scientific and engineering research, as well as business, military, or political strategy.

These systems seem technically possible and we’ll have strong incentives to build them

As we saw above, we’ve already produced systems that are very good at carrying out specific tasks.

We’ve also already produced rudimentary planning systems, like AlphaStar, which skilfully plays the strategy game Starcraft, and MuZero, which plays chess, shogi, and Go.²²

We’re not sure whether these systems are producing plans in pursuit of goals per se, because we’re not sure exactly what it means to “have goals.” However, since they consistently plan in ways that achieve goals, it seems like they have goals in some sense.

Moreover, some existing systems seem to actually represent goals as part of their neural networks.²³

That said, planning in the real world (instead of games) is much more complex, and to date we’re not aware of any unambiguous examples of goal-directed planning systems, or systems that exhibit high degrees of strategic awareness.

But as we’ve discussed, we expect to see further advances within this century. And we think these advances are likely to produce systems with all three of the above properties.

That’s because we think that there are particularly strong incentives (like profit) to develop these kinds of systems. In short: because being able to plan to achieve a goal, and execute that plan, seems like a particularly powerful and general way of affecting the world.

Getting things done — whether that’s a company selling products, a person buying a house, or a government developing policy — almost always seems to require these skills. One example would be assigning a powerful system a goal and expecting the system to achieve it — rather than having to guide it every step of the way. So planning systems seem likely to be (economically and politically) extremely useful.²⁴

And if systems are extremely useful, there are likely to be big incentives to build them. For example, an AI that could plan the actions of a company by being given the goal to increase its profits (that is, an AI CEO) would likely provide significant wealth for the people involved — a direct incentive to produce such an AI.

As a result, if we can build systems with these properties (and from what we know, it seems like we will be able to), it seems like we are likely to do so.²⁵

Advanced planning systems could easily be dangerously ‘misaligned’

There are reasons to think that these kinds of advanced planning AI systems will be misaligned. That is, they will aim to do things that we don’t want them to do.²⁶

There are many reasons why systems might not be aiming to do exactly what we want them to do. For one thing, we don’t know how, using modern ML techniques, to give systems the precise goals we want (more here).²⁷

We’re going to focus specifically on some reasons why systems might by default be misaligned in such a way that they develop plans that pose risks to humanity’s ability to influence the world — even when we don’t want that influence to be lost.²⁸

What do we mean by “by default”? Essentially, unless we actively find solutions to some (potentially quite difficult) problems, then it seems like we’ll create dangerously misaligned AI. (There are reasons this might be wrong — which we discuss later.)

Three examples of “misalignment” in a variety of systems

It’s worth noting that misalignment isn’t a purely theoretical possibility (or specific to AI) — we see misaligned goals in humans and institutions all the time, and have also seen examples of misalignment in AI systems.²⁹

Example 1: Winning elections

The democratic political framework is intended to ensure that politicians make decisions that benefit society. But what political systems actually reward is winning elections, so that’s what many politicians end up aiming for.

This is a decent proxy goal — if you have a plan to improve people’s lives, they’re probably more likely to vote for you — but it isn’t perfect. As a result, politicians do things that aren’t clearly the best way of running a country, like raising taxes at the start of their term and cutting them right before elections.

That is to say, the things the system does are at least a little different from what we would, in a perfect world, want it to do: the system is misaligned.

Example 2: The profit incentive

Companies have profit-making incentives. By producing more, and therefore helping people obtain goods and services at cheaper prices, companies make more money.

This is sometimes a decent proxy for making the world better, but profit isn’t actually the same as the good of all of humanity (bold claim, we know). As a result, there are negative externalities: for example, companies will pollute to make money despite this being worse for society overall.

Again, we have a misaligned system, where the things the system does are at least a little different from what we would want it to do.

Example 3: Specification gaming in existing AI systems

DeepMind has documented examples of specification gaming: an AI doing well according to its specified reward function (which encodes our intentions for the system), but not doing what researchers intended.

In one example, a robot arm was asked to grasp a ball. But the reward was specified in terms of whether humans thought the robot had been successful. As a result, the arm learned to hover between the ball and the camera, fooling the humans into thinking that it had grasped the ball.³⁰

So we know it’s possible to create a misaligned AI system.

Why these systems could (by default) be dangerously misaligned

Here’s the core argument of this article. We’ll use all three properties from earlier: planning ability, strategic awareness, and advanced capabilities.

To start, we should realise that a planning system that has a goal will also develop ‘instrumental goals’: things that, if they occur, will make it easier to achieve an overall goal.

We use instrumental goals in plans all the time. For example, a high schooler planning their career might think that getting into university will be helpful for their future job prospects. In this case, “getting into university” would be an instrumental goal.

A sufficiently advanced AI planning system would also include instrumental goals in its overall plans.

If a planning AI system also has enough strategic awareness, it will be able to identify facts about the real world (including potential things that would be obstacles to any plans), and plan in light of them. Crucially, these facts would include that access to resources (e.g. money, compute, influence) and greater capabilities — that is, forms of power — open up new, more effective ways of achieving goals.

This means that, by default, advanced planning AI systems would have some worrying instrumental goals:

• Self-preservation — because a system is more likely to achieve its goals if it is still around to pursue them (in Stuart Russell’s memorable phrase, “You can’t fetch the coffee if you’re dead”).
• Preventing any changes to the AI system’s goals — since changing its goals would lead to outcomes that are different from those it would achieve with its current goals.
• Gaining power — for example, by getting more resources and greater capabilities.

Crucially, one clear way in which the AI can ensure that it will continue to exist (and not be turned off), and that its objectives will never be changed, would be to gain power over the humans who might affect it (we talk here about how AI systems might actually be able to do that).

What’s more, the AI systems we’re considering have advanced capabilities — meaning they can do one or more tasks that grant people significant power when carried out well in today’s world. With such advanced capabilities, these instrumental goals will not be out of reach, and as a result, it seems like the AI system would use its advanced capabilities to get power as part of the plan’s execution. If we don’t want the AI systems we create to take power away from us this would be a particularly dangerous form of misalignment.

In the most extreme scenarios, a planning AI system with sufficiently advanced capabilities could successfully disempower us completely.

As a (very non-rigorous) intuitive check on this argument, let’s try to apply it to humans.

Humans have a variety of goals. For many of these goals, some form of power-seeking is advantageous: though not everyone seeks power, many people do (in the form of wealth or social or political status), because it’s useful for getting what they want. This is not catastrophic (usually!) because, as human beings:

• We generally feel bound by human norms and morality (even people who really want wealth usually aren’t willing to kill to get it).
• We aren’t that much more capable or intelligent than one another. So even in cases where people aren’t held back by morality, they’re not able to take over the world.

(We discuss whether humans are truly power-seeking later.)

A sufficiently advanced AI wouldn’t have those limitations.

It might be hard to find ways to prevent this sort of misalignment

The point of all this isn’t to say that any advanced planning AI system will necessarily attempt to seek power. Instead, it’s to point out that, unless we find a way to design systems that don’t have this flaw, we’ll face significant risk.

It seems more than plausible that we could create an AI system that isn’t misaligned in this way, and thereby prevent any disempowerment. Here are some strategies we might take (plus, unfortunately, some reasons why they might be difficult in practice):³¹

• Control the objectives of the AI system. We may be able to design systems that simply don’t have objectives to which the above argument applies — and thus don’t incentivise power-seeking behaviour. For example, we could find ways to explicitly instruct AI systems not to harm humans, or find ways to reward AI systems (in training environments) for not engaging in specific kinds of power-seeking behaviour (and also find ways to ensure that this behaviour continues outside the training environment).

  Carlsmith gives two reasons why doing this seems particularly hard.

  First, for modern ML systems, we don’t get to explicitly state a system’s objectives — instead we reward (or punish) a system in a training environment so that it learns on its own. This raises a number of difficulties, one of which is goal misgeneralisation. Researchers have uncovered real examples of systems that appear to have learned to pursue a goal in the training environment, but then fail to generalise that goal when they operate in a new environment. This raises the possibility that we could think we’ve successfully trained an AI system not to seek power — but that the system would seek power anyway when deployed in the real world.³²

  Second, when we specify a goal to an AI system (or, when we can’t explicitly do that, when we find ways to reward or punish a system during training), we usually do this by giving the system a proxy by which outcomes can be measured (e.g. positive human feedback on a system’s achievement). But often those proxies don’t quite work.³³ In general, we might expect that even if a proxy appears to correlate well with successful outcomes, it might not do so when that proxy is optimised for. (The examples above of politicians, companies, and the robot arm failing to grasp a ball are illustrations of this.) We’ll look at a more specific example of how problems with proxies could lead to an existential catastrophe here.

  For more on the specific difficulty of controlling the objectives given to deep neural networks trained using self-supervised learning and reinforcement learning, we recommend OpenAI governance researcher Richard Ngo’s discussion of how realistic training processes lead to the development of misaligned goals.

• Control the inputs into the AI system. AI systems will only develop plans to seek power if they have enough information about the world to realise that seeking power is indeed a way to achieve its goals.

• Control the capabilities of the AI system. AI systems will likely only be able to carry out plans to seek power if they have sufficiently advanced capabilities in skills that grant people significant power in today’s world.

But to make any strategy work, it will need to both:

• Retain the usefulness of the AI systems — and so remain economically competitive with less safe systems. Controlling the inputs and capabilities of AI systems will clearly have costs, so it seems hard to ensure that these controls, even if they’re developed, are actually used. But this is also a problem for controlling a system’s objectives. For example, we may be able to prevent power-seeking behaviour by ensuring that AI systems stop to check in with humans about any decisions they make. But these systems might be significantly slower and less immediately useful to people than systems that don’t stop to carry out these checks. As a result, there might still be incentives to use a faster, more initially effective misaligned system (we’ll look at incentives more in the next section).

• Continue to work as the planning ability and strategic awareness of systems improve over time. Some seemingly simple solutions (for example, trying to give a system a long list of things it isn’t allowed to do, like stealing money or physically harming humans) break down as the planning abilities of the systems increase. This is because, the more capable a system is at developing plans, the more likely it is to identify loopholes or failures in the safety strategy — and as a result, the more likely the system is to develop a plan that involves power-seeking.

Ultimately, by looking at the state of the research on this topic, and speaking to experts in the field, we think that there are currently no known ways of building aligned AI systems that seem likely to fulfil both these criteria.

So: that’s the core argument. There are many variants of this argument. Some have argued that AI systems might gradually shape our future via subtler forms of influence that nonetheless could amount to an existential catastrophe; others argue that the most likely form of disempowerment is in fact just killing everyone. We’re not sure how a catastrophe would be most likely to play out, but have tried to articulate the heart of the argument, as we see it: that AI presents an existential risk.

There are definitely reasons this argument might not be right! We go through some of the reasons that seem strongest to us below. But overall it seems possible that, for at least some kinds of advanced planning AI systems, it will be harder to build systems that don’t seek power in this dangerous way than to build systems that do.

Disempowerment by AI systems would be an existential catastrophe

When we say we’re concerned about existential catastrophes, we’re not just concerned about risks of extinction. This is because the source of our concern is rooted in longtermism: the idea that the lives of all future generations matter, and so it’s extremely important to protect their interests.

This means that any event that could prevent all future generations from living lives full of whatever you think makes life valuable (whether that’s happiness, justice, beauty, or general flourishing) counts as an existential catastrophe.

It seems extremely unlikely that we’d be able to regain power over a system that successfully disempowers humanity. And as a result, the entirety of the future — everything that happens for Earth-originating life, for the rest of time — would be determined by the goals of systems that, although built by us, are not aligned with us. Perhaps those goals will create a long and flourishing future, but we see little reason for confidence.³⁴

This isn’t to say that we don’t think AI also poses a risk of human extinction. Indeed, we think making humans extinct is one highly plausible way in which an AI system could completely and permanently ensure that we are never able to regain power.

People might deploy misaligned AI systems despite the risk

Surely no one would actually build or use a misaligned AI if they knew it could have such terrible consequences, right?

Unfortunately, there are at least two reasons people might create and then deploy misaligned AI — which we’ll go through one at a time:³⁵

1. People might think it’s aligned when it’s not

Imagine there’s a group of researchers trying to tell, in a test environment, whether a system they’ve built is aligned. We’ve argued that an intelligent planning AI will want to improve its abilities to effect changes in pursuit of its objective, and it’s almost always easier to do that if it’s deployed in the real world, where a much wider range of actions are available. As a result, any misaligned AI that’s sophisticated enough will try to understand what the researchers want it to do and at least pretend to be doing that, deceiving the researchers into thinking it’s aligned. (For example, a reinforcement learning system might be rewarded for certain apparent behaviour during training, regardless of what it’s actually doing.)

Hopefully, we’ll be aware of this sort of behaviour and be able to detect it. But catching a sufficiently advanced AI in deception seems potentially harder than catching a human in a lie, which isn’t always easy. For example, a sufficiently intelligent deceptive AI system may be able to deceive us into thinking we’ve solved the problem of AI deception, even if we haven’t.

If AI systems are good at deception, and have sufficiently advanced capabilities, a reasonable strategy for such a system could be to deceive humans completely until the system has a way to guarantee it can overcome any resistance to its goals.

2. There are incentives to deploy systems sooner rather than later

We might also expect some people with the ability to deploy a misaligned AI to charge ahead despite any warning signs of misalignment that do come up, because of race dynamics — where people developing AI want to do so before anyone else.

For example, if you’re developing an AI to improve military or political strategy, it’s much more useful if none of your rivals have a similarly powerful AI.

These incentives apply even to people attempting to build an AI in the hopes of using it to make the world a better place.

For example, say you’ve spent years and years researching and developing a powerful AI system, and all you want is to use it to make the world a better place. Simplifying things a lot, say there are two possibilities:

1. This powerful AI will be aligned with your beneficent aims, and you’ll transform society in a potentially radically positive way.
2. The AI will be sufficiently misaligned that it’ll take power and permanently end humanity’s control over the future.

Let’s say you think there’s a 90% chance that you’ve succeeded in building an aligned AI. But technology often develops at similar speeds across society, so there’s a good chance that someone else will soon also develop a powerful AI. And you think they’re less cautious, or less altruistic, so you think their AI will only have an 80% chance of being aligned with good goals, and pose a 20% chance of existential catastrophe. And only if you get there first can your more beneficial AI be dominant. As a result, you might decide to go ahead with deploying your AI, accepting the 10% risk.

4. Even if we find a way to avoid power-seeking, there are still risks

So far we’ve described what a large proportion of researchers in the field⁷ think is the major existential risk from potential advances in AI, which depends crucially on an AI seeking power to achieve its goals.

If we can prevent power-seeking behaviour, we will have reduced existential risk substantially.

But even if we succeed, there are still existential risks that AI could pose.

There are at least two ways these risks could arise:

• We expect that AI systems will help increase the rate of scientific progress.³⁶ While there would be clear benefits to this automation — the rapid development of new medicine, for example — some forms of technological development can pose threats, including existential threats, to humanity. This technological advancement might increase our available destructive power or make dangerous technologies cheaper or more widely accessible.
• We might start to see AI automate many – or possibly even all – economically important tasks. It’s hard to predict exactly what the effects of this would be on society. But it seems plausible that this could increase existential risks. For example, if AI systems are highly transformative, then their use (or potential use) could possibly create insurmountable power imbalances. Even the threat of this might be enough. For example, a military might feel pushed to create transformative automated weapons because it knows or believes its enemies are doing so, even if this dynamic benefits no one.

We know of several specific areas in which advanced AI may increase existential risks, though are likely others we haven’t thought of.

Bioweapons

In 2022, Collaborations Pharmaceuticals — a small research corporation in North Carolina — were building an AI model to help determine the structure of new drugs. As part of this process, they trained the model to penalise drugs that it predicted were toxic. This had just one problem: you could run the toxicity prediction in reverse to invent new toxic drugs.

Some of the deadliest events in human history have been pandemics. Pathogens’ ability to infect, replicate, kill, and spread — often undetected — make them exceptionally dangerous.

Even without AI, advancing biotechnology poses extreme risks. It potentially provides opportunities for state actors or terrorists to create mass-casualty events.

Advances in AI have the potential to make biotechnology more dangerous.

For example:

1. Dual-use tools, like the automation of laboratory processes, could lower the barriers for rogue actors trying to manufacture a dangerous pandemic virus.³⁷ The Collaborations Pharmaceuticals model is an example of a dual-use tool (although it’s not particularly dangerous).

2. AI-based biological design tools could enable sophisticated actors to reprogram the genomes of dangerous pathogens to specifically enhance their lethality, transmissibility, and immune evasion.³⁸

If AI is able to advance the rate of scientific and technological progress, these risks may be amplified and accelerated — making dangerous technology more widely available or increasing its possible destructive power.³⁹

In the 2023 survey of AI experts, 73% of respondents said they had either “extreme” or “substantial” concern that in the future Al will let “dangerous groups make powerful tools (e.g. engineered viruses).”⁴⁰

Intentionally dangerous AI agents

Most of this article discusses the risk of power-seeking AI systems that arise unintentionally due to misalignment.

But we can’t rule out the possibility that some people might intentionally create rogue AI agents that seek to disempower humanity. It might seem hard to imagine, but extremist ideologies of many forms have inspired humans to carry out radically violent and even self-destructive plans.⁴¹

Cyberweapons

AI can already be used in cyberattacks, such as phishing, and more powerful AI may cause greater information security challenges (though it could also be useful in cyberdefense).

On its own, AI-enabled cyberwarfare is unlikely to pose an existential threat to humanity. Even the most damaging and costly societal-scale cyberattacks wouldn’t approach an extinction-level event.

But AI-enabled cyberattacks could provide access to other dangerous technology, such as bioweapons, nuclear arsenals, or autonomous weapons. So there may be genuine existential risks posed by AI-related cyberweapons, but they will most likely run through another existential risk.

The cyber capabilities of AI systems are also relevant to how a power-seeking AI could actually take power.

Other dangerous tech

If AI systems generally accelerate the rate of scientific and technological progress, we think it’s reasonably likely that we’ll invent new dangerous technologies.

For example, atomically precise manufacturing, sometimes called nanotechnology, has been hypothesised as an existential threat — and it’s a scientifically plausible technology that AI could help us invent far sooner than we would otherwise.

In The Precipice, Toby Ord estimated the chances of an existential catastrophe by 2120 from “unforeseen anthropogenic risks” at 1 in 30. This estimate suggests there could be other discoveries, perhaps involving yet to be understood physics, that could enable the creation of technologies with catastrophic consequences.⁴²

AI could empower totalitarian governments

An AI-enabled authoritarian government could completely automate the monitoring and repression of its citizens, as well as significantly influence the information people see, perhaps making it impossible to coordinate action against such a regime.

AI is already facilitating the ability of governments to monitor their own citizens.

The NSA is using AI to help filter the huge amounts of data they collect, significantly speeding up their ability to identify and predict the actions of people they are monitoring. In China, AI is increasingly being used for facial recognition and predictive policing, including automated racial profiling and automatic alarms when people classified as potential threats enter certain public places.

These sorts of surveillance technologies seem likely to significantly improve — thereby increasing governments’ abilities to control their populations.

At some point, authoritarian governments could extensively use AI-related technology to:

• Monitor and track dissidents
• Preemptively suppress opposition to the ruling party
• Control the military and dominate external actors
• Manipulate information flows and carefully shape public opinion

Again, in the 2023 survey of AI experts, 73% of respondents expressed “extreme” or “substantial” concern that in the future authoritarian rulers could “use Al to control their population.”⁴⁰

If a regime achieved a form of truly stable totalitarianism, it could make people’s lives much worse for a long time into the future, making it a particularly scary possible scenario resulting from AI. (Read more in our article on risks of stable totalitarianism.)

AI could worsen war

We’re concerned that great power conflict could also pose a substantial threat to our world, and advances in AI seem likely to change the nature of war — through lethal autonomous weapons⁴³ or through automated decision making.⁴⁴

In some cases, great power war could pose an existential threat — for example, if the conflict is nuclear. Some argue that lethal autonomous weapons, if sufficiently powerful and mass-produced, could themselves constitute a new form of weapon of mass destruction.

And if a single actor produces particularly powerful AI systems, this could be seen as giving them a decisive strategic advantage. Such an outcome, or even the expectation of such an outcome, could be highly destabilising.

Imagine that the US was working to produce a planning AI that’s intelligent enough to ensure that Russia or China could never successfully launch another nuclear weapon. This could incentivise a first strike from the actor’s rivals before these AI-developed plans can ever be put into action.

This is because nuclear deterrence can benefit from symmetry between the abilities of nuclear powers, in that the threat of a nuclear response to a first strike is believable and therefore a deterrent to a first strike. Advances in AI, which could be directly applied to nuclear forces, could create asymmetries in the capabilities of nuclear-armed nations. This could include improving early warning systems, air defence systems, and cyberattacks that disable weapons.

For example, many countries use submarine-launched ballistic missiles as part of their nuclear deterrence systems — the idea is that if nuclear weapons can be hidden under the ocean, they will never be destroyed in the first strike. This means that they can always be used for a counterattack, and therefore act as an effective deterrent against first strikes. But AI could make it far easier to detect submarines underwater, enabling their destruction in a first strike — removing this deterrent.

Many other destabilising scenarios are likely possible.

A report from the Stockholm International Peace Research Institute found that, while AI could potentially also have stabilising effects (for example by making everyone feel more vulnerable, decreasing the chances of escalation), destabilising effects could arise even before advances in AI are actually deployed. This is because one state’s belief that their opponents have new nuclear capabilities can be enough to disrupt the delicate balance of deterrence.

Luckily, there are also plausible ways in which AI could help prevent the use of nuclear weapons — for example, by improving the ability of states to detect nuclear launches, which would reduce the chances of false alarms like those that nearly caused nuclear war in 1983.

Overall, we’re uncertain about whether AI will substantially increase the risk of nuclear or conventional conflict in the short term — it could even end up decreasing the risk. But we think it’s important to pay attention to possible catastrophic outcomes and take reasonable steps to reduce their likelihood.

Other risks from AI

We’re also concerned about the following issues:

• Existential threats that result not from the power-seeking behaviour of AI systems, but from the interaction between AI systems. (In order to pose a risk, these systems would still need to be, to some extent, misaligned.)
• Other ways we haven’t thought of that AI systems could be misused — especially ones that might significantly affect future generations.
• Other moral mistakes made in the design and use of AI systems, particularly if future AI systems are themselves deserving of moral consideration. For example, we might (inadvertently) create sentient AI systems, which could then suffer in huge numbers. We think this could be extremely important, so we’ve written about it in a separate problem profile.

So, how likely is an AI-related catastrophe?

This is a really difficult question to answer.

There are no past examples we can use to determine the frequency of AI-related catastrophes.

All we have to go off are arguments (like the ones we’ve given above), and less relevant data like the history of technological advances. And we’re definitely not certain that the arguments we’ve presented are completely correct.

Consider the argument we gave earlier about the dangers of power-seeking AI in particular, based off Carlsmith’s report. At the end of his report, Carlsmith gives some rough guesses of the chances that each stage of his argument is correct (conditional on the previous stage being correct):

1. By 2070 it will be possible and financially feasible to build strategically aware systems that can outperform humans on many power-granting tasks, and that can successfully make and carry out plans: Carlsmith guesses there’s a 65% chance of this being true.
2. Given this feasibility, there will be strong incentives to build such systems: 80%.
3. Given both the feasibility and incentives to build such systems, it will be much harder to develop aligned systems that don’t seek power than to develop misaligned systems that do, but which are at least superficially attractive to deploy: 40%.
4. Given all of this, some deployed systems will seek power in a misaligned way that causes over $1 trillion (in 2021 dollars) of damage: 65%.
5. Given all the previous premises, misaligned power-seeking AI systems will end up disempowering basically all of humanity: 40%.
6. Given all the previous premises, this disempowerment will constitute an existential catastrophe: 95%.

Multiplying these numbers together, Carlsmith estimated that there’s a 5% chance that his argument is right and there will be an existential catastrophe from misaligned power-seeking AI by 2070. When we spoke to Carlsmith, he noted that in the year between the writing of his report and the publication of this article, his overall guess at the chance of an existential catastrophe from power-seeking AI by 2070 had increased to >10%.⁴⁵

The overall probability of existential catastrophe from AI would, in Carlsmith’s view, be higher than this, because there are other routes to possible catastrophe — like those discussed in the previous section — although our guess is that these other routes are probably a lot less likely to lead to existential catastrophe.

For another estimate, in The Precipice, philosopher and advisor to 80,000 Hours Toby Ord estimated a 1-in-6 risk of existential catastrophe by 2120 (from any cause), and that 60% of this risk comes from misaligned AI — giving a total of a 10% risk of existential catastrophe from misaligned AI by 2120.

A 2021 survey of 44 researchers working on reducing existential risks from AI found the median risk estimate was 32.5% — the highest answer given was 98%, and the lowest was 2%.⁴⁶ There’s obviously a lot of selection bias here: people choose to work on reducing risks from AI because they think this is unusually important, so we should expect estimates from this survey to be substantially higher than estimates from other sources. But there’s clearly significant uncertainty about how big this risk is, and huge variation in answers.

All these numbers are shockingly, disturbingly high. We’re far from certain that all the arguments are correct. But these are generally the highest guesses for the level of existential risk of any of the issues we’ve examined (like engineered pandemics, great power conflict, climate change, or nuclear war).

That said, I think there are reasons why it’s harder to make guesses about the risks from AI than other risks – and possibly reasons to think that the estimates we’ve quoted above are systematically too high.

If I was forced to put a number on it, I’d say something like 1%. This number includes considerations both in favour and against the argument. I’m less worried than other 80,000 Hours staff — our position as an organisation is that the risk is between 3% and 50%.

All this said, the arguments for such high estimates of the existential risk posed by AI are persuasive — making risks from AI a top contender for the most pressing problem facing humanity.

5. We can tackle these risks

We think one of the most important things you can do would be to help reduce the gravest risks that AI poses.

This isn’t just because we think these risks are high — it’s also because we think there are real things we can do to reduce these risks.

We know of two main ways people work to reduce these risks:

1. Technical AI safety research
2. AI governance and policy work

There are lots of ways to contribute to this work. In this section, we discuss many broad approaches within both categories to illustrate the point that there are things we can do to address these risks. Below, we discuss the kinds of careers you can pursue to work on these kinds of approaches.

Technical AI safety research

The benefits of transformative AI could be huge, and there are many different actors involved (operating in different countries), which means it will likely be really hard to prevent its development altogether.

(It’s also possible that it wouldn’t even be a good idea if we could — after all, that would mean forgoing the benefits as well as preventing the risks.)

As a result, we think it makes more sense to focus on making sure that this development is safe — meaning that it has a high probability of avoiding all the catastrophic failures listed above.

One way to do this is to try to develop technical solutions to prevent the kind of power-seeking behaviour we discussed earlier — this is generally known as working on technical AI safety, sometimes called just “AI safety” for short.

We discuss this path in more detail here:

Career review of technical AI safety research

Approaches

There are lots of approaches to technical AI safety, including:

• Scalably learning from human feedback. Examples include iterated amplification, AI safety via debate, building AI assistants that are uncertain about our goals and learn them by interacting with us, and other ways to get AI systems to report truthfully what they know.
• Threat modelling. An example of this work would be demonstrating the possibility of (allowing us to study) dangerous capabilities, like deceptive or manipulative AI systems. This approach splits into work that evaluates whether a model has dangerous capabilities (like the work of METR in evaluating GPT-4), and work that evaluates whether a model would cause harm in practice (like Anthropic’s research into the behaviour of large language models and this paper on goal misgeneralisation). It can also include work to find ‘model organisms of misalignment’, in the hope of better understanding their dangers.
• Work to figure out how to control powerful AI systems – preventing them from causing harm even if they are unsafe. Read more in this blogpost from the team at Redwood Research working on control.
• Interpretability research. This work involves studying why AI systems do what they do and trying to put it into human-understandable terms. For example, this paper examined how AlphaZero learns chess, and this paper looked into finding latent knowledge in language models without supervision. This category also includes mechanistic interpretability — for example, Zoom In: An Introduction to Circuits by Olah et al.. For more, see this survey paper, as well as Hubinger’s a transparency and interpretability tech tree, and Nanda’s A Longlist of Theories of Impact for Interpretability for overviews of of how interpretability research could reduce existential risk from AI.
• Other anti-misuse research to reduce the risks of catastrophe caused by misuse of systems. For example, this work includes training AIs so they’re hard to use for dangerous purposes. (Note there’s lots of overlap with the other work on this list).
• Research to increase the robustness of neural networks. This work involves ensuring that the sorts of behaviour neural networks display when exposed to one set of inputs continues when exposed to inputs they haven’t previously been exposed to, in order to prevent AI systems changing to unsafe behaviour. See section 2 of Unsolved Problems in AI safety for more.
• Work to build cooperative AI. Find ways to ensure that even if individual AI systems seem safe, they don’t produce bad outcomes through interacting with other sociotechnical systems. For more, see Open Problems in Cooperative AI by Dafoe et al. or the Cooperative AI Foundation. This seems particularly relevant for the reduction of ‘s-risks.’
• More generally, there are some unified safety plans. For more, see Hubinger’s 11 possible proposals for building safe advanced AI, or Karnofsky’s How might we align transformative AI if it’s developed very soon.⁴⁷

See Neel Nanda’s overview of the AI alignment landscape for more details.

Read more about technical AI safety research below.

AI governance and policy

Reducing the gravest risks from AI will require sound high-level decision making and policy, both at AI companies themselves and in governments.

As AI has advanced and drawn increasing interest from customers and investors, governments have shown an interest in regulating the technology. Some have already taken significant steps to play a role in managing the development of AI, including:

• The US and the UK have each established their own national AI Safety Institutes.
• The European Union has passed the EU AI Act, which contains specific provisions for governing general-purpose AI models that pose systemic risks.
• The UK and then South Korea have hosted the first two AI Safety Summits (in 2023 and 2024), a series of high-profile summits aiming to coordinate between different countries, academics, researchers, and civil society leaders.
• China has implemented regulations targeting recommendation algorithms, synthetic AI content, generative AI models, and facial recognition technology.
• The US instituted export controls to reduce China’s access to the most cutting-edge chips used in AI development.

Much more will need to be done to reduce the biggest risks — including continuous evaluation of the AI governance landscape to assess overall progress.

We discuss this career path in more detail here:

Career review of AI strategy and policy careers

Approaches

People working in AI policy have proposed a range of approaches to reducing risk as AI systems get more powerful.

We don’t necessarily endorse all the ideas below, but what follows is a list of some prominent policy approaches that could be aimed at reducing the largest dangers from AI:⁴⁸

• Responsible scaling policies: some major AI companies have already begun developing internal frameworks for assessing safety as they scale up the size and capabilities of their systems. These frameworks introduce safeguards that are intended to become increasingly stringent as AI systems become more potentially dangerous, and ensure that AI systems’ capabilities don’t outpace companies’ abilities to keep systems safe. Many argue that these internal policies are not sufficient for safety, but they may represent a promising step for reducing risk. You can see versions of such policies from Anthropic, Google DeepMind, and OpenAI.
• Standards and evaluation: governments may also develop industry-wide benchmarks and testing protocols to assess whether AI systems pose major risks. The non-profit METR and the UK AI Safety Institute are among the organisations currently developing these evaluations to test AI models before and after they are released. This can include creating standardised metrics for an AI systems’s capabilities and potential to cause harm, as well as propensity for power-seeking or misalignment.
• Safety cases: this practice involves requiring AI developers to provide comprehensive documentation demonstrating the safety and reliability of their systems before deployment. This approach is similar to safety cases used in other high-risk industries like aviation or nuclear power.⁴⁹ You can see discussion of this idea in a paper from Clymer et al and in a post from Geoffrey Irving at the UK AI Safety Institute.
• Information security standards: we can establish robust rules for protecting AI-related data, algorithms, and infrastructure from unauthorised access or manipulation — particularly the AI model weights. Rand released a detailed report analysing the security risks to major AI companies, particularly from state actors.
• Liability law: existing law already imposes some liability on companies that create dangerous products or cause significant harm to the public, but its application to AI models and risk in particular is unclear. Clarifying how liability applies to companies that create dangerous AI models could incentivise them to take additional steps to reduce risk. Law professor Gabriel Weil has written about this idea.
• Compute governance: governments may regulate access to and use of high-performance computing resources necessary for training large AI models. The US restrictions on exporting state-of-the-art chips to China is one example of such a policy, and others are possible. Companies could also be required to install hardware-level safety features directly into AI chips or processors. These could be used to track chips and verify they’re not in the possession of anyone who shouldn’t have them, or for other purposes. You can learn more about this topic in our interview with Lennart Heim and in this report from the Center for a New American Security.
• International coordination: Fostering global cooperation on AI governance to ensure consistent standards. This could involve treaties, international organisations, or multilateral agreements on AI development and deployment. We discuss some related considerations in our article on China-related AI safety and governance paths.
• Societal adaptation: it may be critically important to prepare society for the widespread integration of AI and the potential risks it poses. For example, we might need to develop new information security measures to protect crucial data in a world with AI-enabled hacking. Or we may want to implement strong controls to prevent handing over key societal decisions to AI systems.⁵⁰
• Pausing scaling if appropriate: some argue that we should currently pause all scaling of larger AI models because of the dangers the technology poses. We have featured some discussion of this idea on our podcast, and it seems hard to know when or if this would be a good idea. If carried out, it could involve industry-wide agreements or regulatory mandates to pause scaling efforts when necessary.

The details, benefits, and downsides of many of these ideas have yet to be fully worked out, so it’s crucial that we do more research. And this list isn’t comprehensive — there are likely other important policy interventions and governance strategies worth pursuing.

We also need more forecasting research into what we should expect to happen with AI, such as the work done at Epoch AI.

6. This work is neglected

In 2022, we estimated there were around 400 people around the world working directly on reducing the chances of an AI-related existential catastrophe (with a 90% confidence interval ranging between 200 and 1,000). Of these, about three quarters worked on technical AI safety research, with the rest split between strategy (and other governance) research and advocacy.⁵ We also estimated that there were around 800 people working in complementary roles, but we’re highly uncertain about this figure.³

In The Precipice, Ord estimated that there was between $10 million and $50 million spent on reducing AI risk in 2020.

That might sound like a lot of money, but we’re spending something like 1,000 times that amount⁴ on speeding up the development of transformative AI via commercial capabilities research and engineering at large AI companies.

To compare the $50 million spent on AI safety in 2020 to other well-known risks, we’re currently spending several hundreds of billions per year on tackling climate change.

Because this field is so neglected and has such high stakes, we think your impact working on risks from AI could be much higher than working on many other areas — which is why our top two recommended career paths for making a big positive difference in the world are technical AI safety and AI policy research and implementation.

What do we think are the best arguments against this problem being pressing?

As we said above, we’re not totally sure the arguments we’ve presented for AI representing an existential threat are right. Though we do still think that the chance of catastrophe from AI is high enough to warrant many more people pursuing careers to try to prevent such an outcome, we also want to be honest about the arguments against doing so, so you can more easily make your own call on the question.

Here we’ll cover the strongest reasons (in our opinion) to think this problem isn’t particularly pressing. In the next section we’ll cover some common objections that (in our opinion) hold up less well, and explain why.

Arguments against working on AI risk to which we think there are strong responses

We’ve just discussed the major objections to working on AI risk that we think are most persuasive. In this section, we’ll look at objections that we think are less persuasive, and give some reasons why.

What you can do concretely to help

As we mentioned above, we know of two main ways to help reduce existential risks from AI:

1. Technical AI safety research
2. AI governance and policy work

The biggest way you could help would be to pursue a career in either one of these areas, or in a supporting area.

The first step is learning a lot more about the technologies, problems, and possible solutions. We’ve collated some lists of our favourite resources here, and our top recommendation is to take a look at the technical alignment curriculum from AGI Safety Fundamentals.

Technical AI safety

If you’re interested in a career in technical AI safety, the best place to start is our career review of being an AI safety researcher.

If you want to learn more about technical AI safety as a field of research — e.g. the different techniques, schools of thought, and threat models — our top recommendation is to take a look at the technical alignment curriculum from AGI Safety Fundamentals.

It’s important to note that you don’t have to be an academic or an expert in AI or AI safety to contribute to AI safety research. For example, software engineers are needed at many places conducting technical safety research, and we also highlight more roles below.

You can see a list of key organisations where you might do this kind of work in the full career review.

AI governance and policy work

If you’re interested in a career in AI governance and policy, the best place to start is our AI governance and policy career review.

You don’t need to be a bureaucrat in a grey suit to have a career in AI governance and policy — there are roles suitable for a wide range of skill sets. In particular, people with technical skills in machine learning and related fields are needed for governance work (although those skills are certainly not necessary).

We split this career path into six different kinds of roles:

1. Government roles
2. Research
3. Industry work
4. Advocacy and lobbying
5. Third-party auditing and evaluation
6. International work and coordination

We also have specific articles on working in US AI policy and China-related AI safety and governance paths.

And you can learn more about where specifically you might work in this career path in our career review.

If you’re new to the topic and interested in learning more broadly about AI governance, our top recommendation is to take a look at the governance curriculum from AGI safety fundamentals.

Complementary (yet crucial) roles

Even in a research organisation, around half of the staff will be doing other tasks essential for the organisation to perform at its best and have an impact. Having high-performing people in these roles is crucial.

We think the importance of these roles is often underrated because the work is less visible. So we’ve written several career reviews on these areas to help more people enter these careers and succeed, including:

• Operations management to help impactful organisations grow and function as effectively as possible.
• Research management at an AI safety research organisation.
• Being an executive assistant to someone who’s doing really important work on safety and governance.

Other ways to help

AI safety is a big problem and it needs help from people doing a lot of different kinds of work.

One major way to help is to work in a role that directs funding or people towards AI risk, rather than working on the problem directly. We’ve reviewed a few career paths along these lines, including:

• Founding new projects — in this case, starting new initiatives aimed at reducing risks from advanced AI.
• Being a grantmaker to fund promising projects focused on reducing catastrophic AI risk.
• Working in communication roles.
• Helping to build communities of people working on this problem. The most relevant community is the AI safety community itself, but it could also be impactful to help build the community of people working on the world’s most pressing problems (including risks from AI).

There are ways all of these could go wrong, so the first step is to become well-informed about the issue.

There are also other technical roles besides safety research that could help contribute, like:

• Working in information security to protect AI (or the results of key experiments) from misuse, theft, or tampering.
• Becoming an expert in AI hardware as a way of steering AI progress in safer directions.

You can read about all these careers — why we think they’re helpful, how to enter them, and how you can predict whether they’re a good fit for you — on our career reviews page.

Find vacancies on our job board

Our job board features opportunities in AI technical safety and governance:

View all opportunities

Top resources to learn more

We've hit you with a lot of further reading throughout this article — here are a few of our favourites:

• AI could defeat all of us combined and the "most important century" blog post series by Holden Karnofsky, co-CEO of Open Philanthropy, argues that the 21st century could be the most important century ever for humanity as a result of AI.
• Why AI alignment could be hard with modern deep learning by Open Philanthropy researcher Cotra is a gentle introduction to how risks from power-seeking AI could play out with current machine learning methods. Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover, also by Cotra, provides a much more detailed description of how risks could play out (which we'd recommend for people familiar with ML).
• AGI safety from first principles provides OpenAI governance researcher Richard Ngo's perspective on how to think about risks from artificial general intelligence.
• Is power-seeking AI an existential risk? by Open Philanthropy researcher Joseph Carlsmith is an in-depth look covering exactly how and why AI could cause the disempowerment of humanity (but watch out — it's even longer than this article!). It's also available as an audio narration. For a shorter summary, see Carlsmith's talk on the same topic.
• Distinguishing AI takeover scenarios by Sam Clarke and Sammy Martin summarises various ways in which AI could go wrong.
• AI governance: Opportunity and theory of impact by DeepMind governance lead Allan Dafoe explores ways in which research into AI governance could effect change.
• A bird's-eye view of the AI alignment landscape by Neel Nanda summarises the different ways in which technical alignment research could reduce the risk from AI.
• An overview of 11 proposals for building safe advanced AI by Evan Hubinger discusses and evaluates plausible techniques for AI alignment.
• Podcasts: The AI X-risk Research Podcast, particularly episode 12 with Paul Christiano and episode 13 with Richard Ngo — both of which serve as excellent introductions to AI risk.

On The 80,000 Hours Podcast, we have a number of in-depth interviews with people actively working to positively shape the development of artificial intelligence:

• Paul Christiano on how OpenAI is developing real solutions to the 'AI alignment problem', and his vision of how humanity will progressively hand over decision-making to AI systems
• Allan Dafoe on trying to prepare the world for the possibility that AI will destabilise global politics
• Carl Shulman on the economy and national security and government and society after AGI
• Richard Ngo on large language models, OpenAI, and striving to make the future go well
• Ajeya Cotra on accidentally teaching AI models to deceive us
• Jan Leike on how to become a machine learning alignment researcher (from 2018) and OpenAI's massive push to make superintelligence safe in 4 years or less (from 2023)
• Nathan Labenz on the final push for AGI, understanding OpenAI's leadership drama, and red-teaming frontier models
• Rohin Shah on DeepMind and trying to fairly hear out both AI doomers and doubters
• Tom Davidson on how quickly AI could transform the world
• Dario Amodei on OpenAI and how AI will change the world for good and ill
• Miles Brundage on the world's desperate need for AI strategists and policy experts
• Holden Karnofsky, cofounder of GiveWell and Open Philanthropy, has been on three of our podcasts, explaining:
  • How AIs might take over even if they're no smarter than humans, and his four-part playbook for AI risk
  • How philanthropy can have maximum impact by taking big risks (including a discussion of his work in positively shaping the development of AI)
  • Why this might be the most important century
• PhD or programming? Fast paths into aligning AI as a machine learning engineer, according to ML engineers Catherine Olsson & Daniel Ziegler

If you want to go into much more depth, the AGI safety fundamentals course is a good starting point. There are two tracks to choose from: technical alignment or AI governance. If you have a more technical background, you could try Intro to ML Safety, a course from the Center for AI Safety.

And finally, here are a few general sources (rather than specific articles) that you might want to explore:

• The AI Alignment Forum, which is aimed at researchers working in technical AI safety.
• AI Impacts, a project that aims to improve society's understanding of the likely impacts of human-level artificial intelligence.
• The Alignment Newsletter, a weekly publication with recent content relevant to AI alignment with thousands of subscribers.
• Import AI, a weekly newsletter about artificial intelligence by Jack Clark (cofounder of Anthropic), read by more than 10,000 experts.
• Jeff Ding's ChinAI Newsletter, weekly translations of writings from Chinese thinkers on China's AI landscape.

Read next:  Explore other pressing world problems

Want to learn more about global issues we think are especially pressing? See our list of issues that are large in scale, solvable, and neglected, according to our research.

Continue →

Acknowledgements

Huge thanks to Joel Becker, Tamay Besiroglu, Jungwon Byun, Joseph Carlsmith, Jesse Clifton, Emery Cooper, Ajeya Cotra, Andrew Critch, Anthony DiGiovanni, Noemi Dreksler, Ben Edelman, Lukas Finnveden, Emily Frizell, Ben Garfinkel, Katja Grace, Lewis Hammond, Jacob Hilton, Samuel Hilton, Michelle Hutchinson, Caroline Jeanmaire, Kuhan Jeyapragasan, Arden Koehler, Daniel Kokotajlo, Victoria Krakovna, Alex Lawsen, Howie Lempel, Eli Lifland, Katy Moore, Luke Muehlhauser, Neel Nanda, Linh Chi Nguyen, Luisa Rodriguez, Caspar Oesterheld, Ethan Perez, Charlie Rogers-Smith, Jack Ryan, Rohin Shah, Buck Shlegeris, Marlene Staib, Andreas Stuhlmüller, Luke Stebbing, Nate Thomas, Benjamin Todd, Stefan Torges, Michael Townsend, Chris van Merwijk, Hjalmar Wijk, and Mark Xu for either reviewing this article or their extremely thoughtful and helpful comments and conversations. (This isn’t to say that they would all agree with everything we’ve said here — in fact, we’ve had many spirited disagreements in the comments on this article!)