[Crossposted from windowsontheory]

The following statements seem to be both important for AI safety and are not widely agreed upon. These are my opinions, not those of my employer or colleagues. As is true for anything involving AI, there is significant uncertainty about everything written below. However, for readability, I present these points in their strongest form, without hedges and caveats. That said, it is essential not to be dogmatic, and I am open to changing my mind based on evidence. None of these points are novel; others have advanced similar arguments. I am sure that for each statement below, there will be people who find it obvious and people who find it obviously false.

AI safety will not be solved on its own.
An “AI scientist” will not solve it either.
Alignment is not about loving humanity; it’s about robust reasonable compliance.
Detection is more important than prevention.
Interpretability is neither sufficient nor necessary for alignment.
Humanity can survive an unaligned superintelligence.

Before going into the points, we need to define what we even mean by “AI safety” beyond the broad sense of “making sure that nothing bad happens as a result of training or deploying an AI model.” Here, I am focusing on technical means for preventing large-scale (sometimes called “catastrophic”) harm as a result of deploying AI. There is more to AI than technical safety. In particular, many potential harms, including AI enabling mass surveillance and empowering authoritarian governments, can not be addressed by technical means alone.

My views on AI safety are colored by my expectations of how AI will progress over the next few years. I believe that we will make fast progress, and in terms of technical capabilities, we will reach “AGI level” within 2-4 years, and in a similar timeframe, AIs will gain superhuman capabilities across a growing number of dimensions. (Though I suspect translating these technical capabilities to the economic and societal impact we associate with AGI will take significantly longer.) This assumes that our future AGIs and ASIs will be, to a significant extent, scaled-up versions of our current models. On the one hand, this is good news, since it means our learnings from current models are relevant for more powerful ones, and we can develop and evaluate safety techniques using them. On the other hand, this makes me doubt that safety approaches that do not show signs of working for our current models will be successful for future AIs.

1. AI safety will not be solved on its own.

Safety and alignment are AI capabilities, whether one views these as preventing harm, following human intent, or promoting human well-being. The higher the stakes in which AI systems are used, the more critical these capabilities become. Thus, one could think that AI safety will “take care of itself”: As with other AI capabilities, safety will improve when we scale up resources, and as AI systems get deployed in more critical applications, people will be willing to spend resources to achieve the required safety level.

There is some truth to this view. As we move to more agentic models, safety will become critical for utility. However, markets and other human systems take time to adapt (as Keynes said, “Markets can remain irrational longer than you can remain solvent”), and the technical work to improve safety takes time as well. The mismatch in timescales between the technical progress in AI—along with the speed at which AI systems can communicate and make decisions—and the slower pace of human institutions that need to adapt means that we cannot rely on existing systems alone to ensure that AI turns out safe by default.

Consider aviation safety: It is literally a matter of life and death, so there is a strong interest in making commercial flights as safe as possible.

Indeed, the rate of fatal accidents in modern commercial aviation is extremely low: about one fatal accident in 16 million flights (i.e., about 99.99999%—seven “nines”—of flights do not suffer a fatal accident). But it took decades to get to this point. In particular, it took 50 years for the rate of fatalities per trillion passenger kilometers to drop by roughly two orders of magnitude.

There are multiple ways in which AI safety is harder than aviation safety.

Timeline: We don’t have 50 years. If we compare AI progress to transportation, we might go from bicycles (GPT-2) to spaceships (AGI) in a decade.
Metrics and generality: We can’t just track a single outcome (like “landed safely”). The G in AGI means that the number of ways that AGI can go wrong is as large as the number of ways that applications of human intelligence can go wrong, which include direct physical harm from misuse, societal impacts through misinformation, social upheaval from too fast changes, AIs autonomously causing harm and more.

Of course, we have a lot of experience dealing with the ways in which human intelligence goes wrong. However, we have gained this experience through thousands of years, and we still haven’t figured out how to get along with each other. The different skill profiles of AI, the ways it can be integrated with other systems, and the speed at which it evolves all pose unique challenges.

2. An “AI scientist” will not solve it either.

One approach for AI safety is to follow the following plan:

Develop an AI that is superhuman (or at least expert level) in AI research
Use this AI to do research and solve the alignment problem.
Deploy safe AGI.
Profit.

Since doing AI research is a narrower task than AGI and has less of a potential attack surface, the hope is that this will be an easier lift than solving the general problem of AI alignment.

We absolutely should be using AI to help AI safety research, and with time, AIs will take an ever-growing role in it. That said, given the current trajectory of AI progress, I don’t think the plan above will work out as stated, and we can’t just hope that the AI scientist will come and solve all our problems for us like a “deus ex machina.”

The reasons are twofold:

No temporal gap: We will not have a discrete “AGI moment” and generally a clear temporal gap between the point where we have AI that is capable of expert-level autonomous AI/ML research and the point at which powerful AI agents are deployed at scale.
No “magic” insight: AI safety will not be solved via one brilliant idea that we need an AI scientist to discover. It will require defense in depth (i.e., the “Swiss cheese” approach) and significant work on safety infrastructure both pre and post-deployment.

No temporal gap. Frontier AI systems exhibit an uneven mixture of capabilities. Depending on the domain and distance from their training distribution, they can range from middle-school to graduate-level knowledge. Economic integration of AI will also be uneven. The skill requirements, error tolerance, regulatory barriers, and other factors vary widely across industries.

Furthermore, our current competitive landscape means AI is being developed and deployed by multiple actors across many countries, accelerating both capabilities and integration. Given these factors, I suspect we will not see a discrete “AGI moment.” Instead, AGI may only be recognized in retrospect, just like official declarations of recessions.

All of this means that we will not have a “buffer” between the time we have an expert AI scientist (Step 1 in the plan above) and the time agentic AIs are widely deployed in high-stakes settings (Step 3). We are likely not to have discrete “steps” at all but more of a continuous improvement of a jagged frontier of both capabilities and economic impact. I expect AI systems to provide significant productivity boosts for AI safety researchers, as they would for many other professions, including AI capability researchers (to the extent these can be distinguished). However, given long tails of tasks and incremental deployment, by the time we have a true superhuman safety researcher, AI will already be deeply integrated into our society, including applications with huge potential for harm.

No “magic” insight. Like computer security, AI safety will not be solved by one magic insight; it will require defense in depth. Some of this work can’t wait till AGI and has to be done now and built into our AI infrastructure so we can use it later. Without doing work on safety now, including collecting data and evaluation, and building in safety at all stages, including training, inference, and monitoring, we won’t be able to take advantage of any of the ideas discovered by AI scientists. For instance, advances in discovering software vulnerabilities wouldn’t help us if we didn’t already have the infrastructure for over-the-air digitally signed patches. Adopting best practices also takes time. For example, although MD5 weaknesses were known in 1996, and practical attacks emerged in 2004, many systems still used it as late as 2020.

It is also unclear how “narrow” the problem of AI alignment is. Governing intelligence, whether natural or artificial, is not just a technical problem. There are undoubtedly technical hurdles we need to overcome, but they are not the only ones. Furthermore, like human AI researchers, an automated AI scientist will need to browse the web, use training data from various sources, import external packages, and more, all of which provide openings for adversarial attacks. We can’t ignore adversarial robustness in the hope that an AI scientist will solve it for us.

Constant instead of temporal allocation. I do agree that as capabilities grow, we should be shifting resources to safety. But rather than temporal allocation (i.e., using AI for safety before using it for productivity), I believe we need constant compute allocation: ensuring a fixed and sufficiently high fraction of compute is always spent on safety research, monitoring, and mitigations. See points 4 (monitoring) and 6 (dealing with unaligned ASI) below.

3. Alignment is not about loving humanity; it’s about robust reasonable compliance.

One way of phrasing the AI alignment task is to get AIs to “love humanity” or to have human welfare as their primary objective (sometimes called “value alignment”). One could hope to encode these via simple principles like Asimov’s three laws or Stuart Russel’s three principles, with all other rules derived from these.

There is something very clean and appealing about deriving all decisions from one or few “axioms,” whether it is Kant’s categorical imperative or Bentham’s principle of utility. But when we try to align complex human systems, whether it’s countries, institutions, or corporations, we take a very different approach. The U.S. Constitution, laws, and regulations take hundreds of thousands of pages, and we have set up mechanisms (i.e., courts) to arbitrate actual or potential contradictions between them. A 100k-page collection of rules makes a boring science fiction story, but it is needed for the functioning of complex systems. Similarly, ensuring the safety of complex computer programs requires writing detailed specifications or contracts that each component must satisfy so that other pieces can rely on it.

We don’t want an AI philosophizing about abstract principles and deciding that a benevolent AI dictatorship (or alternatively, reversion to pre-industrial times) is best for humanity. We want an AI to comply with a given specification that tells it precisely what constraints it must satisfy while optimizing whatever objective it is given. For high-stakes applications, we should be able to ensure this compliance to an arbitrary number of “nines,” possibly by dedicating resources to safety that scale with the required level of reliability. (We don’t know how to do this today. We also don’t have ways to automatically discover and arbitrate edge cases or contradictions in specifications. Achieving both is a research effort that I am very excited about.)

Perfect compliance does not mean literal compliance. We don’t want systems that, like the genies in fables (or some of the robots in Asimov’s stories), follow the letter of the specifications while violating their spirit. In both common law and civil law systems, there is significant room for interpretation when applying a specification in a new situation not foreseen by its authors. Charles Marshall said, “Integrity is doing the right thing, even when no one is watching.” For AIs, we could say that “alignment is doing the right thing, even when you are out of the training distribution.”

What we want is reasonable compliance in the sense of:

Following the specification precisely when it is clearly defined.
Following the spirit of the specification in a way that humans would find reasonable in other cases.

One way to define “reasonable” is to think what a “jury of one’s peers” or “lay judges” - random humans from the pool relevant to the situation - would consider in such a case. As in jury trials, we rely on the common sense and moral intuitions of the typical community member. Current LLMs are already pretty good at simulating humans; with more data and research, they can become better still.

One could argue that when the specification is not well defined, AIs should fall back to general ethical principles and analyze them, and so we should be training AIs to be expert ethicists. But I prefer the random human interpretation of “reasonable.” William Buckley famously said that he’d rather be governed by the first 2,000 people in the Boston phone book than the faculty of Harvard University. As a Harvard faculty member, I can see his point. In fact, in my time at Harvard, I have seen no evidence that philosophers or ethicists have any advantage over (for example) computer scientists in matters of governance or morality. I’d rather AIs simulate normal people than ethics experts.

As another example, logician Kurt Gödel famously claimed to find an internal contradiction in the U.S. Constitution that could enable a dictatorship. We want our AIs to be smart enough to recognize such interpretations but reasonable enough not to follow them.

While I advocate for detailed specifications over abstract values, moral principles obviously inform us when writing specifications. Also, specifications can and likely will include guiding principles (e.g., the precautionary principle) for dealing with underspecified cases.

Specifications as expert code of conduct. Another reason compliance with specifications is a better alignment objective than following some higher-level objectives is that superhuman AIs may be considered analogous to human experts. Throughout history, people have trusted various experts, including priests, lawyers, doctors, and scientists, partially because these experts followed explicit or implicit codes of conduct.

We expect these professionals to follow these codes of conduct even when they contradict their perception of the “greater good.” For example, during the June 2020 protests following the murder of George Floyd, over 1,000 healthcare professionals signed a letter saying that since “white supremacy is a lethal public health issue,” protests against systemic racism “must be supported.” They also said their stance “should not be confused with a permissive stance on all gatherings, particularly protests against stay-home orders.” In other words, these healthcare professionals argued that COVID-19 regulations should be applied to a protest based on its message. Undoubtedly, these professionals believed that they were promoting the greater good. Perhaps they were even correct. However, they also abused their status as healthcare professionals and injected political considerations into their recommendations. Unsurprisingly, trust in physicians and hospitals decreased substantially over the last four years.

Trust in AI systems will require legibility in their decisions and ensuring that they comply with our policies and specifications rather than making some 4D chess ethical calculations. Indeed, some of the most egregious current examples of misalignment are models faking alignment to pursue some higher values.

Robust compliance. AI safety will require robust compliance, which means that AIs should comply with their specifications even if adversaries provide parts of their input. Current chat models are a two-party interaction— between model and user— and even if the user is adversarial, at worst, they will get some information, such as how to cook meth, that can also be found online. But we are rapidly moving away from this model. Agentic systems will interact with several different parties with conflicting goals and incentives and ingest inputs from multiple sources. Adversarial attacks will shift from being the province of academic papers and fun tweets into real-world attacks with significant negative consequences. (Once again, we don’t know how to achieve robust compliance today, and this is a research effort I’m very excited about.)

But what about higher values? Wojciech Zaremba views alignment as “lovingly reasonable robust compliance” in the sense that the AI should have a bias for human welfare and not (for example) blindly help a user if they are harming themselves. Since humans have basic empathy and care, I believe that we may not need “lovingly” since “reasonable” does encompass some basic human intuitions. There is a reason courts and juries are staffed by humans. Courts also sometimes reach for moral principles or “natural law” in interpretations and decisions (though that is controversial). The primary way I view the compliance approach as different from “value alignment” is the order of priorities. In “value alignment,” the higher-order principles determine the lower-level rules. In the compliance-based approach, like in the justice system, an AI should appeal to higher values only in cases where there are gaps or ambiguities in the specifications. This is not always a good thing. This approach rules out an AI Stanislav Petrov that would overrule its chain of command in the name of higher-order moral principles. The approach of using a “typical human” as the measuring yard for morality also rules out an AI John Brown, and other abolitionists who are recognized as morally right today but were a minority at the time. I believe it is a tradeoff worth taking: let AIs follow the rules and leave it to humans to write the rules, as well as decide when to update them or break them.

All of the above leaves open the question of who writes the specification and whether someone could write a specification of the form “do maximum harm.” The answer to the latter question is yes, but I believe that the existence of a maximally evil superintelligence does not have to spell doom; see point 6 below.

4. Detection is more important than prevention.

When I worked at Microsoft Research, my colleague Butler Lampson used to say, “The reason your house is not burglarized is not because you have a lock, but because you have an alarm.” Much of current AI safety focuses on prevention—getting the model to refuse harmful requests. Prevention is essential, but detection is even more critical for truly dangerous settings.

If someone is serious about creating a CBRN threat and AI refuses to help, they won’t just give up. They may combine open-source information, human assistance, or partial AI help to pursue this goal. It is not enough to simply refuse such a person; we want to ensure that we stop them before they cause mass harm.

Another reason refusal is problematic is that many queries to AI systems have dual uses. For example, it can be that 90% of the people asking a particular question on biology are doing so for a beneficial purpose, while 10% of them could be doing it for a nefarious one. Given the input query, the model may not have the context to determine what is its intended usage.

For this reason, simple refusal is not enough. Measures such as “know your customer” and the ability to detect and investigate potentially dangerous uses would be crucial. Detection also shifts the balance from the attacker to the defender. In the “refusal game,” the attacker only needs to win once and get an answer to their question. In the “detection game,” they must avoid detection in every query.

In general, detection allows us to set lower thresholds for raising a flag (as there is no performance degradation, it is only a matter of the amount of resources allocated for investigation) and enables us to learn from real-world deployment and potentially detect novel risks and vulnerabilities before they cause large-scale damage.

Detection does not mean that model-based work on compliance and robustness is irrelevant. We will need to write specifications on the conditions for flagging and build monitoring models (or monitoring capabilities for generative/agentic models) that are robust to adversarial attacks and can reasonably interpret specifications. It may turn out that safety requires spending more resources (e.g., inference-time compute) for monitoring than we do for generation/action.

Finally, there is an inherent tension between monitoring and preserving privacy. One of the potential risks of AIs is that in a world where AIs are deeply embedded in all human interactions, it will be much easier for governments to surveil and control the population. Figuring out how to protect both privacy and security, which may require tools such as on-device models, is an urgent research challenge.

5. Interpretability is neither sufficient nor necessary for alignment.

Mechanistic interpretability is a fascinating field. I enjoy reading interpretability papers and learn a lot from them. I think it can be useful for AI safety, but it’s not a “silver bullet” for it and I don’t believe it lies on the critical path for AI safety.

The standard argument is that we cannot align or guarantee the safety of systems we don’t understand. But we already align complex systems, whether it’s corporations or software applications, without complete “understanding,” and do so by ensuring they meet certain technical specifications, regulations, or contractual obligations.

More concretely, interpretability is about discovering the underlying algorithms and internal representations of AI systems. This can allow for both monitoring and steering. However, I suspect that the actual algorithms and concepts of AI systems are inherently “messy.” Hence, I believe that there would be an inherent tradeoff between reliability (having a concept or algorithm that describes the system in 99.999% of cases) and interpretability. For safety and control, reliability is more critical than interpretability.

This is not to say interpretability is useless! While I think we may not be able to get the level of reliability needed for steering or monitoring, interpretability can still be helpful as a diagnostic tool. For example, checking whether training method A or training method B leads to more deception. Also, even if we don’t use it directly, interpretability can provide many insights to accelerate safety research and the discovery of other methods. (And we certainly need all the safety acceleration we can get!.) Interpretability can also serve as a sanity check and a way to increase public trust in AI models. The above considerations refer to “classical” weights/activations interpretability; CoT interpretation may well be significantly more robust. Finally, as I said above, it is essential not to be dogmatic. It may turn out that I’m wrong, and interpretability is necessary for alignment.

6. Humanity can survive an unaligned superintelligence.

Kim Jong Un is probably one of the most “misaligned” individuals who ever lived. North Korea's nuclear arsenal is estimated to include more than 50 bombs at least as powerful as the Hiroshima bomb. North Korea is also believed to have biological and chemical weapons. Given its technological and military strength, if North Korea had been transported 200 years ago, it might well have ruled over the world. But in our current world, it is a pariah state, ranking 178th in the world in GDP per capita.

The lesson is that the damage an unaligned agent can cause depends on its relative power, not its absolute power. If there were only one superintelligence and it wanted to destroy humanity, we’d be doomed. But in a world where many actors have ASI, the balance between aligned and unaligned intelligence matters.

To make things more concrete (and with some oversimplification), imagine that “intelligence” is measured to a first approximation in units of compute. Just like material resources are currently spent, compute can be used for:

Actively harmful causes (attackers).
Ensuring safety (defenders).
Neutral, profit-driven causes.

Currently, the world order is kept by ensuring that (2)--- resources spent for defense, policing, safety, and other ways to promote peace and welfare– dominate (1)--- resources spent by criminals, terrorist groups, and rogue states.

While intelligence can amplify the utility for a given amount of resources, it can do so for both the “attackers” and the “defenders.” So, as long as defender intelligence dominates attacker intelligence, we should be able to maintain the same equilibrium we currently have.

Of course, the precise “safe ratios” could change, as intelligence is not guaranteed to have the same amplification factor for attackers and defenders. However, the amplification factor is not infinite. Moreover, for very large-scale attacks, the costs to the attacker may well be superlinear. For instance, killing a thousand people in a single terror attack is much harder than killing the same number in many smaller attacks.

Moreover, it is not clear that intelligence is the limiting factor for attackers. Considering examples of successful large-scale attacks, it is often the defender that could have been most helped by more intelligence, both in the military sense as well as in the standard one. (In fact, paradoxically, it seems that many such attacks, from Pearl Harbor through 9/11 to Oct 7th, would have been prevented if the attackers had been better at calculating the outcome, which more often than not did not advance their objectives.)

I expect that, in general, the effort required for an attacker to cause damage would look somewhat like the sigmoid graph above. The balance of attacker and defender advantages rescales the X axis of effort required for damage, but it would still be extremely hard to extinguish humanity completely.

Another way to say this is that I do not accept Bostrom’s vulnerable world hypothesis, which states that at some level of technological development, civilization will be devastated by default. I believe that as long as aligned superintelligent AIs dominate unaligned ASIs, any dangerous technological development (e.g., cheaper techniques such as SILEX to create nuclear weapons) would be discovered by aligned ASIs first, allowing time to prepare. A key assumption of Bostrom’s 2019 paper is the limited capacity of governments for preventative measures. However, since then, we have seen in the COVID-19 pandemic the ability of governments to mobilize quickly and enforce harsh restrictive measures on their citizens.

The bottom line is not that we are guaranteed safety, nor that unaligned or misaligned superintelligence could not cause massive harm— on the contrary. It is that there is no single absolute level of intelligence above which the existence of a misaligned intelligence with this level spells doom. Instead, it is all about the world in which this superintelligence will operate, the goals to which other superintelligent systems are applied, and our mechanisms to ensure that they are indeed working towards their specified goals.

Acknowledgements. I am grateful to Sam Altman, Alec Radford, and Wojciech Zaremba for comments on this blog post, though they do not necessarily endorse any of its views and are not responsible for any errors or omissions in it.