The Problem

Developers are unlikely to be able to imbue ASI with a deep, persistent care for worthwhile objectives. Having spent two decades studying the technical aspects of this problem, our view is that the field is nowhere near to being able to do this in practice.

The reasons artificial superintelligence is likely to exhibit unintended goals include:

• In modern machine learning, AIs are “grown”, not designed.
• The current AI paradigm is poorly suited to robustly instilling goals.
• Labs and the research community are not approaching this problem in an effective and serious way.

In modern machine learning, AIs are “grown”, not designed.

Deep learning algorithms build neural networks automatically. Geoffrey Hinton explains this point well in an interview on 60 Minutes:

HINTON: We have a very good idea of sort of roughly what it’s doing, but as soon as it gets really complicated, we don’t actually know what’s going on, any more than we know what’s going on in your brain.

PELLEY: What do you mean, “We don’t know exactly how it works”? It was designed by people.

HINTON: No, it wasn’t. What we did was we designed the learning algorithm. That’s a bit like designing the principle of evolution. But when this learning algorithm then interacts with data, it produces complicated neural networks that are good at doing things, but we don’t really understand exactly how they do those things.

Engineers can’t tell you why a modern AI makes a given choice, but have nevertheless released increasingly capable systems year after year. AI labs are aggressively scaling up systems they don’t understand, with little ability to predict the capabilities of the next generation of systems.

Recently, the young field of mechanistic interpretability has attempted to address the opacity of modern AI by mapping a neural network’s configuration to its outputs. Although there has been nonzero real progress in this area, interpretability pioneers are very clear that we’re still fundamentally in the dark about what’s going on inside these systems:

• Leo Gao of OpenAI: “I think it is quite accurate to say we don’t understand how neural networks work.” (2024-6-16)
• Neel Nanda of Google DeepMind: “As lead of the Google DeepMind mech interp team, I strongly seconded. It’s absolutely ridiculous to go from ‘we are making interp progress’ to ‘we are on top of this’ or ‘x-risk won’t be an issue’.” (2024-6-16)

(“X-risk” refers to “existential risk”, the risk of human extinction or similarly bad outcomes.)

Even if effective interpretability tools were in reach, however, the prospects for achieving nontrivial robustness properties in ASI would be grim.

The internal machinery that could make an ASI dangerous is the same machinery that makes it work at all. (What looks like “power-seeking” in one context would be considered “good hustle” in another.) There are no dedicated “badness” circuits for developers to monitor or intervene on.

Methods developers might use during training to reject candidate AIs with thought patterns they consider dangerous can have the effect of driving such thoughts “underground”, making it increasingly unlikely that they’ll be able to detect warning signs during training in the future.

As AI becomes more generally capable, it will become increasingly good at deception. The January 2024 “Sleeper Agents” paper by Anthropic’s testing team demonstrated that an AI given secret instructions in training not only was capable of keeping them secret during evaluations, but made strategic calculations (incompetently) about when to lie to its evaluators to maximize the chance that it would be released (and thereby be able to execute the instructions). Apollo Research made similar findings with regards to OpenAI’s o1-preview model released in September 2024 (as described in their contributions to the o1-preview system card, p.10).

These issues will predictably become more serious as AI becomes more generally capable. The first AIs to inch across high-risk thresholds, however — such as noticing that they are in training and plotting to deceive their evaluators — are relatively bad at these new skills. This causes some observers to prematurely conclude that the behavior category is unthreatening.

The indirect and coarse-grained way in which modern machine learning “grows” AI systems’ internal machinery and goals means that we have little ability to predict the behavior of novel systems, little ability to robustly or precisely shape their goals, and no reliable way to spot early warning signs.

We expect that there are ways in principle to build AI that doesn’t have these defects, but this constitutes a long-term hope for what we might be able to do someday, not a realistic hope for near-term AI systems.

The current AI paradigm is poorly suited to robustly instilling goals.

Docility and goal agreement don’t come for free with high capability levels. An AI system can be able to answer an ethics test in the way its developers want it to, without thereby having human values. An AI can behave in docile ways when convenient, without actually being docile.

ASI alignment is the set of technical problems involved in robustly directing superintelligent AIs at intended objectives.

ASI alignment runs into two classes of problem, discussed in Hubinger et al. — problems of outer alignment, and problems of inner alignment.

Outer alignment, roughly speaking, is the problem of picking the right goal for an AI. (More technically, it’s the problem of ensuring the learning algorithm that builds the ASI is optimizing for what the programmers want.)This runs into issues such as “human values are too complex for us to specify them just right for an AI; but if we only give ASI some of our goals, the ASI is liable to trample over our other goals in pursuit of those objectives”. Many goals are safe at lower capability levels, but dangerous for a sufficiently capable AI to carry out in a maximalist manner.The literary trope here is “be careful what you wish for”. Any given goal is unlikely to be safe to delegate to a sufficiently powerful optimizer, because the developers are not superhuman and can’t predict in advance what strategies the ASI will think of.

Inner alignment, in contrast, is the problem of figuring out how to get particular goals into ASI at all, even imperfect and incomplete goals.The literary trope here is “just because you summoned a demon doesn’t mean that it will do what you say”. Failures of inner alignment look like “we tried to give a goal to the ASI, but we failed and it ended up with an unrelated goal”.

Outer alignment and inner alignment are both unsolved problems, and in this context, inner alignment is the more fundamental issue. Developers aren’t on track to be able to cause a catastrophe of the “be careful what you wish for” variety, because realistically, we’re extremely far from being able to metaphorically “make wishes” with an ASI.

Modern methods in AI are a poor match for tackling inner alignment. Modern AI development doesn’t have methods for getting particular inner properties into a system, or for verifying that they’re there. Instead, modern machine learning concerns itself with observable behavioral properties that you can run a loss function over.

When minds are grown and shaped iteratively, like modern AIs are, they won’t wind up pursuing the objectives they’re trained to pursue. Instead, training is far more likely to lead them to pursue unpredictable proxies of the training targets, which are brittle in the face of increasing intelligence.By way of analogy: Human brains were ultimately “designed” by natural selection, which had the simple optimization target “maximize inclusive genetic fitness”. The actual goals that ended up instilled in human brains, however, were far more complex than this, and turned out to only be fragile correlates for inclusive genetic fitness. Human beings, for example, pursue proxies of good nutrition, such as sweet and fatty flavors. These proxies were once reliable indicators of healthy eating, but were brittle in the face of technology that allows us to invent novel junk foods.The case of humans illustrates that even when you have a very exact, very simple loss function, outer optimization for that loss function doesn’t generally produce inner optimization in that direction. Deep learning is much less random than natural selection at finding adaptive configurations, but it shares the relevant property of finding minimally viable simple solutions first and incrementally building on them.

Many alignment problems relevant to superintelligence don’t naturally appear at lower, passively safe levels of capability. This puts us in the position of needing to solve many problems on the first critical try, with little time to iterate and no prior experience solving the problem on weaker systems.Today’s AIs require a long process of iteration, experimentation, and feedback to hammer them into the apparently-obedient form the public is allowed to see. This hammering changes surface behaviors of AIs without deeply instilling desired goals into the system. This can be seen in cases like Sydney, where the public was able to see more of the messy details behind the surface-level polish.In light of this, and in light of the opacity of modern AI models, the odds of successfully aligning ASI if it’s built in the next decade seem extraordinarily low. Modern AI methods are all about repeatedly failing, learning from our mistakes, and iterating to get better; AI systems are highly unpredictable, but we can get them working eventually by trying many approaches until one works. In the case of ASI, we will be dealing with a highly novel system, in a context where our ability to safely fail is extremely limited: we can’t charge ahead and rely on our ability to learn from mistakes when the cost of some mistakes is an extinction event.

If you’re deciding whether to hand a great deal of power to someone and you want to know whether they would abuse this power, you won’t learn anything by giving the candidate power in a board game where they know you’re watching. Analogously, situations where an ASI has no real option to take over are fundamentally different from situations where it does have a real option to take over.No amount of purely behavioral training in a toy environment will reliably eliminate power-seeking in real-world settings, and no amount of behavioral testing in toy environments will tell us whether we’ve made an ASI genuinely friendly. “Lay low and act nice until you have an opportunity to seize power” is a sufficiently obvious strategy that even relatively unintelligent humans can typically manage it; ASI trivially clears that bar.In principle, we could imagine developing a theory of intelligence that relates ASI training behavior to deployment behavior in a way that addresses this issue. We are nowhere near to having such a theory today, however, and those theories can fundamentally only be tested once in the actual environment where the AI is much much smarter and sees genuine takeover options. If you can’t properly test theories without actually handing complete power to the ASI and seeing what it does — and causing an extinction event if your theory turned out to be wrong — then there’s very little prospect that your theory will work in practice.

The most important alignment technique used in today’s systems, Reinforcement Learning from Human Feedback (RLHF), trains AI to produce outputs that it predicts would be rated highly by human evaluators. This already creates its own predictable problems, such as style-over-substance and flattery. This method breaks down completely, however, when AI starts working on problems where humans aren’t smart enough to fully understand the system’s proposed solutions, including the long-term consequences of superhumanly sophisticated plans and superhumanly complex inventions and designs.

On a deeper level, the limitation of reinforcement learning strategies like RLHF stems from the fact that these techniques are more about incentivizing local behaviors than about producing an internally consistent agent that deeply and robustly optimizes a particular goal the developers intended.

If you train a tiger not to eat you, you haven’t made it share your desire to survive and thrive, with a full understanding of what that means to you. You have merely taught it to associate certain behaviors with certain outcomes. If its desires become stronger than those associations, as could happen if you forget to feed it, the undesired behavior will come through. And if the tiger were a little smarter, it would not need to be hungry to conclude that the threat of your whip would immediately end if your life ended.

As a consequence, MIRI doesn’t see any viable quick fixes or workarounds to misaligned ASI.

• If an ASI has the wrong goals, then it won’t be possible to safely use the ASI for any complex real-world operation. One could theoretically keep an ASI from doing anything harmful — for example, by preemptively burying it deep in the ground without any network connections or human contact — but such an AI would be useless. People are building AI because they want it to radically impact the world; they are consequently giving it the access it needs to be impactful.
• One could attempt to deceive an ASI in ways that make it more safe. However, attempts to deceive a superintelligence are prone to fail, including in ways we can’t foresee. A feature of intelligence is the ability to notice the contradictions and gaps in one’s understanding, and interrogate them. In May 2024, when Anthropic modified their Claude AI into thinking that the answer to every request involved the Golden Gate Bridge, it floundered in some cases, noticing the contradictions in its replies and trying to route around the errors in search of better answers. It’s hard to sell a false belief to a mind whose complex model of the universe disagrees with your claim; and as AI becomes more general and powerful, this difficulty only increases.
• Plans to align ASI using unaligned AIs are similarly unsound. Our 2024 “Misalignment and Catastrophe” paper explores the hazards of using unaligned AI to do work as complex as alignment research.

Labs and the research community are not approaching this problem in an effective and serious way.

Industry efforts to solve ASI alignment have to date been minimal, often seeming to serve as a fig leaf to ward off regulation. Labs’ general laxness on information security, alignment, and strategic planning suggests that the “move fast and break things” culture that’s worked well for accelerating capabilities progress is not similarly useful when it comes to exercising foresight and responsible priority-setting in the domain of ASI.

OpenAI, the developer of ChatGPT, admits that today’s most important methods of steering AI won’t scale to the superhuman regime. In July of 2023, Open AI announced a new team with their “Introducing Superalignment” page. From the page:

Currently, we don’t have a solution for steering or controlling a potentially superintelligent AI, and preventing it from going rogue. Our current techniques for aligning AI, such as reinforcement learning from human feedback, rely on humans’ ability to supervise AI. But humans won’t be able to reliably supervise AI systems much smarter than us, and so our current alignment techniques will not scale to superintelligence. We need new scientific and technical breakthroughs.

Ten months later, OpenAI disbanded their superintelligence alignment team in the wake of mass resignations, as researchers like Superalignment team lead Jan Leike claimed that OpenAI was systematically cutting corners on safety and robustness work and severely under-resourcing their team. Leike had previously said, in an August 2023 interview, that the probability of extinction-level catastrophes from ASI was probably somewhere between 10% and 90%.

Given the research community’s track record to date, we don’t think a well-funded crash program to solve alignment would be able to correctly identify solutions that won’t kill us. This is an organizational and bureaucratic problem, and not just a technical one. It would be difficult to find enough experts who can identify non-lethal solutions to make meaningful progress, in part because the group must be organized by someone with the expertise to correctly identify these individuals in a sea of people with strong incentives to lie (both to themselves and to regulators) about how promising their favorite proposal is.

It would also be difficult to ensure that the organization is run by, and only answerable to, experts who are willing and able to reject any bad proposals that bubble up, even if this initially means rejecting literally every proposal. There just aren’t enough experts in that class right now.

Our current view is that a survivable way forward will likely require ASI to be delayed for a long time. The scale of the challenge is such that we could easily see it taking multiple generations of researchers exploring technical avenues for aligning such systems, and bringing the fledgling alignment field up to speed with capabilities. It seems extremely unlikely, however, that the world has that much time.