AI as Normal Technology

This post is over 15,000 words long—it is a new paper on our vision for the future of AI. We are pleased to announce that an expanded version of these ideas will become our next book together.

The paper is also published in HTML and PDF formats on the Knight First Amendment Institute’s website. We are grateful for the extensive feedback we’ve received on drafts of the paper.

Update (September 2025): We have published a companion to this essay titled A guide to understanding AI as normal technology.

We articulate a vision of artificial intelligence (AI) as normal technology. To view AI as normal is not to understate its impact—even transformative, general-purpose technologies such as electricity and the internet are “normal” in our conception. But it is in contrast to both utopian and dystopian visions of the future of AI which have a common tendency to treat it akin to a separate species, a highly autonomous, potentially superintelligent entity.1

The statement “AI is normal technology” is three things: a description of current AI, a prediction about the foreseeable future of AI, and a prescription about how we should treat it. We view AI as a tool that we can and should remain in control of, and we argue that this goal does not require drastic policy interventions or technical breakthroughs. We do not think that viewing AI as a humanlike intelligence is currently accurate or useful for understanding its societal impacts, nor is it likely to be in our vision of the future.2

The normal technology frame is about the relationship between technology and society. It rejects technological determinism, especially the notion of AI itself as an agent in determining its future. It is guided by lessons from past technological revolutions, such as the slow and uncertain nature of technology adoption and diffusion. It also emphasizes continuity between the past and the future trajectory of AI in terms of societal impact and the role of institutions in shaping this trajectory.

In Part I, we explain why we think that transformative economic and societal impacts will be slow (on the timescale of decades), making a critical distinction between AI methods, AI applications, and AI adoption, arguing that the three happen at different timescales.

In Part II, we discuss a potential division of labor between humans and AI in a world with advanced AI (but not “superintelligent” AI, which we view as incoherent as usually conceptualized). In this world, control is primarily in the hands of people and organizations; indeed, a greater and greater proportion of what people do in their jobs is AI control.

In Part III, we examine the implications of AI as normal technology for AI risks. We analyze accidents, arms races, misuse, and misalignment, and argue that viewing AI as normal technology leads to fundamentally different conclusions about mitigations compared to viewing AI as being humanlike.

Of course, we cannot be certain of our predictions, but we aim to describe what we view as the median outcome. We have not tried to quantify probabilities, but we have tried to make predictions that can tell us whether or not AI is behaving like normal technology.

In Part IV, we discuss the implications for AI policy. We advocate for reducing uncertainty as a first-rate policy goal and resilience as the overarching approach to catastrophic risks. We argue that drastic interventions premised on the difficulty of controlling superintelligent AI will, in fact, make things much worse if AI turns out to be normal technology— the downsides of which will be likely to mirror those of previous technologies that are deployed in capitalistic societies, such as inequality.3

The world we describe in Part II is one in which AI is far more advanced than it is today. We are not claiming that AI progress—or human progress—will stop at that point. What comes after it? We do not know. Consider this analogy: At the dawn of the first Industrial Revolution, it would have been useful to try to think about what an industrial world would look like and how to prepare for it, but it would have been futile to try to predict electricity or computers. Our exercise here is similar. Since we reject “fast takeoff” scenarios, we do not see it as necessary or useful to envision a world further ahead than we have attempted to. If and when the scenario we describe in Part II materializes, we will be able to better anticipate and prepare for whatever comes next.

A note to readers. This essay has the unusual goal of stating a worldview rather than defending a proposition. The literature on AI superintelligence is copious. We have not tried to give a point-by-point response to potential counter arguments, as that would make the paper several times longer. This paper is merely the initial articulation of our views; we plan to elaborate on them in various follow ups.

Subscribe to receive follow-up essays.

Part I: The Speed of Progress

Figure 1. Like other general-purpose technologies, the impact of AI is materialized not when methods and capabilities improve, but when those improvements are translated into applications and are diffused through productive sectors of the economy.4 There are speed limits at each stage.

Will the progress of AI be gradual, allowing people and institutions to adapt as AI capabilities and adoption increase, or will there be jumps leading to massive disruption, or even a technological singularity? Our approach to this question is to analyze highly consequential tasks separately from less consequential tasks and to begin by analyzing the speed of adoption and diffusion of AI before returning to the speed of innovation and invention.

We use invention to refer to the development of new AI methods—such as large language models—that improve AI’s capabilities to carry out various tasks. Innovation refers to the development of products and applications using AI that consumers and businesses can use. Adoption refers to the decision by an individual (or team or firm) to use a technology, whereas diffusion refers to the broader social process through which the level of adoption increases. For sufficiently disruptive technologies, diffusion might require changes to the structure of firms and organizations, as well as to social norms and laws.

AI Diffusion in Safety-critical Areas Is Slow

In the paper Against Predictive Optimization, we compiled a comprehensive list of about 50 applications of predictive optimization, namely the use of machine learning (ML) to make decisions about individuals by predicting their future behavior or outcomes.5 Most of these applications, such as criminal risk prediction, insurance risk prediction, or child maltreatment prediction, are used to make decisions that have important consequences for people.

While these applications have proliferated, there is a crucial nuance: In most cases, decades-old statistical techniques are used—simple, interpretable models (mostly regression) and relatively small sets of handcrafted features. More complex machine learning methods, such as random forests, are rarely used, and modern methods, such as transformers, are nowhere to be found.

In other words, in this broad set of domains, AI diffusion lags decades behind innovation. A major reason is safety—when models are more complex and less intelligible, it is hard to anticipate all possible deployment conditions in the testing and validation process. A good example is Epic’s sepsis prediction tool which, despite having seemingly high accuracy when internally validated, performed far worse in hospitals, missing two thirds of sepsis cases and overwhelming physicians with false alerts.6

Epic’s sepsis prediction tool failed because of errors that are hard to catch when you have complex models with unconstrained feature sets.7 In particular, one of the features used to train the model was whether a physician had already prescribed antibiotics —to treat sepsis. In other words, during testing and validation, the model was using a feature from the future, relying on a variable that was causally dependent on the outcome. Of course, this feature would not be available during deployment. Interpretability and auditing methods will no doubt improve so that we will get much better at catching these issues, but we are not there yet.

In the case of generative AI, even failures that seem extremely obvious in hindsight were not caught during testing. One example is the early Bing chatbot “Sydney” that went off the rails during extended conversations; the developers evidently did not anticipate that conversations could last for more than a handful of turns.8 Similarly, the Gemini image generator was seemingly never tested on historical figures.9 Fortunately, these were not highly consequential applications.

More empirical work would be helpful for understanding the innovation-diffusion lag in various applications and the reasons for this lag. But, for now, the evidence that we have analyzed in our previous work is consistent with the view that there are already extremely strong safety-related speed limits in highly consequential tasks. These limits are often enforced through regulation, such as the FDA’s supervision of medical devices, as well as newer legislation such as the EU AI Act, which puts strict requirements on high-risk AI.10 In fact, there are (credible) concerns that existing regulation of high-risk AI is so onerous that it may lead to “runaway bureaucracy”.11 Thus, we predict that slow diffusion will continue to be the norm in high-consequence tasks.

At any rate, as and when new areas arise in which AI can be used in highly consequential ways, we can and must regulate them. A good example is the Flash Crash of 2010, in which automated high-frequency trading is thought to have played a part. This led to new curbs on trading, such as circuit breakers.12

Diffusion is Limited by the Speed of Human, Organizational, and Institutional Change

Even outside of safety-critical areas, AI adoption is slower than popular accounts would suggest. For example, a study made headlines due to the finding that, in August 2024, 40% of U.S. adults used generative AI.13 But, because most people used it infrequently, this only translated to 0.5%-3.5% of work hours (and a 0.125-0.875 percentage point increase in labor productivity).

It is not even clear if the speed of diffusion is greater today compared to the past. The aforementioned study reported that generative AI adoption in the U.S. has been faster than personal computer (PC) adoption, with 40% of U.S. adults adopting generative AI within two years of the first mass-market product release compared to 20 % within three years for PCs. But this comparison does not account for differences in the intensity of adoption (the number of hours of use) or the high cost of buying a PC compared to accessing generative AI.14 Depending on how we measure adoption, it is quite possible that the adoption of generative AI has been much slower than PC adoption.

The claim that the speed of technology adoption is not necessarily increasing may seem surprising (or even obviously wrong) given that digital technology can reach billions of devices at once. But it is important to remember that adoption is about software use, not availability. Even if a new AI-based product is instantly released online for anyone to use for free, it takes time to for people to change their workflows and habits to take advantage of the benefits of the new product and to learn to avoid the risks.

Thus, the speed of diffusion is inherently limited by the speed at which not only individuals, but also organizations and institutions, can adapt to technology. This is a trend that we have also seen for past general-purpose technologies: Diffusion occurs over decades, not years.15

As an example, Paul A. David’s analysis of electrification shows that the productivity benefits took decades to fully materialize.16 Electric dynamos were “everywhere but in the productivity statistics” for nearly 40 years after Edison’s first central generating station. 17 This was not just technological inertia; factory owners found that electrification did not bring substantial efficiency gains.

What eventually allowed gains to be realized was redesigning the entire layout of factories around the logic of production lines. In addition to changes to factory architecture, diffusion also required changes to workplace organization and process control, which could only be developed through experimentation across industries. Workers had more autonomy and flexibility as a result of the changes, which also necessitated different hiring and training practices.

The External World Puts a Speed Limit on AI Innovation

It is true that technical advances in AI have been rapid, but the picture is much less clear when we differentiate AI methods from applications.

We conceptualize progress in AI methods as a ladder of generality.18 Each step on this ladder rests on the ones below it and reflects a move toward more general computing capabilities. That is, it reduces the programmer effort needed to get the computer to perform a new task and increases the set of tasks that can be performed with a given amount of programmer (or user) effort; see Figure 2. For example, machine learning increases generality by obviating the need for the programmer to devise logic to solve each new task, only requiring the collection of training examples instead.

It is tempting to conclude that the effort required to develop specific applications will keep decreasing as we build more rungs of the ladder until we reach artificial general intelligence, often conceptualized as an AI system that can do everything out of the box, obviating the need to develop applications altogether.

In some domains, we are indeed seeing this trend of decreasing application development effort. In natural language processing, large language models have made it relatively trivial to develop a language translation application. Or consider games: AlphaZero can learn to play games such as chess better than any human through self-play given little more than a description of the game and enough computing power—a far cry from how game-playing programs used to be developed.

Figure 2: The Ladder of Generality in Computing. For some tasks, higher ladder rungs require less programmer effort to get a computer to perform a new task, and more tasks can be performed with a given amount of programmer (or user) effort.19

However, this has not been the trend in highly consequential, real-world applications that cannot easily be simulated and in which errors are costly. Consider self-driving cars: In many ways, the trajectory of their development is similar to AlphaZero’s self-play—improving the tech allowed them to drive in more realistic conditions, which enabled the collection of better and/or more realistic data, which in turn led to improvements in the tech, completing the feedback loop. But this process took over two decades instead of a few hours in the case of AlphaZero because safety considerations put a limit on the extent to which each iteration of this loop could be scaled up compared to the previous one.20

This “capability-reliability gap” shows up over and over. It has been a major barrier to building useful AI “agents” that can automate real-world tasks.21 To be clear, many tasks for which the use of agents is envisioned, such as booking travel or providing customer service, are far less consequential than driving, but still costly enough that having agents learn from real-world experiences is not straightforward.

Barriers also exist in non-safety-critical applications. In general, much knowledge is tacit in organizations and is not written down, much less in a form that can be learned passively. This means that these developmental feedback loops will have to happen in each sector and, for more complex tasks, may even need to occur separately in different organizations, limiting opportunities for rapid, parallel learning. Other reasons why parallel learning might be limited are privacy concerns: Organizations and individuals might be averse to sharing sensitive data with AI companies, and regulations might limit what kinds of data can be shared with third parties in contexts such as healthcare.

The “bitter lesson” in AI is that general methods that leverage increases in computational power eventually surpass methods that utilize human domain knowledge by a large margin.22 This is a valuable observation about methods, but it is often misinterpreted to encompass application development. In the context of AI-based product development, the bitter lesson has never been even close to true.23 Consider recommender systems on social media: They are powered by (increasingly general) machine learning models, but this has not obviated the need for manual coding of the business logic, the frontend, and other components which, together, can comprise on the order of a million lines of code.

Further limits arise when we need to go beyond AI learning from existing human knowledge.24 Some of our most valuable types of knowledge are scientific and social-scientific, and have allowed the progress of civilization through technology and large-scale social organizations (e.g., governments). What will it take for AI to push the boundaries of such knowledge? It will likely require interactions with, or even experiments on, people or organizations, ranging from drug testing to economic policy. Here, there are hard limits to the speed of knowledge acquisition because of the social costs of experimentation. Societies probably will not (and should not) allow the rapid scaling of experiments for AI development.

Benchmarks Do not Measure Real-World Utility

The methods-application distinction has important implications for how we measure and forecast AI progress. AI benchmarks are useful for measuring progress in methods; unfortunately, they have often been misunderstood as measuring progress in applications, and this confusion has been a driver of much hype about imminent economic transformation.

For example, while GPT-4 reportedly achieved scores in the top 10% of bar exam test takers, this tells us remarkably little about AI’s ability to practice law.25 The bar exam overemphasizes subject-matter knowledge and under-emphasizes real-world skills that are far harder to measure in a standardized, computer-administered format. In other words, it emphasizes precisely what language models are good at—retrieving and applying memorized information.

More broadly, tasks that would lead to the most significant changes to the legal profession are also the hardest ones to evaluate. Evaluation is straightforward for tasks like categorizing legal requests by area of law because there are clear correct answers. But for tasks that involve creativity and judgment, like preparing legal filings, there is no single correct answer, and reasonable people can disagree about strategy. These latter tasks are precisely the ones that, if automated, would have the most profound impact on the profession.26

This observation is in no way limited to law. Another example is the gap between self-contained coding problems at which AI demonstrably excels, and real-world software engineering in which its impact is hard to measure but appears to be modest.27 Even highly regarded coding benchmarks that go beyond toy problems must necessarily ignore many dimensions of real-world software engineering in the interest of quantification and automated evaluation using publicly available data.28

This pattern appears repeatedly: The easier a task is to measure via benchmarks, the less likely it is to represent the kind of complex, contextual work that defines professional practice. By focusing heavily on capability benchmarks to inform our understanding of AI progress, the AI community consistently overestimates the real-world impact of the technology.

This is a problem of ‘construct validity,’ which refers to whether a test actually measures what it is intended to measure.29 The only sure way to measure real-world usefulness of a potential application is to actually build the application and to then test it with professionals in realistic scenarios (either substituting or augmenting their labor, depending on the intended use). Such ‘uplift’ studies generally do show that professionals in many occupations benefit from existing AI systems, but this benefit is typically modest and is more about augmentation than substitution, a radically different picture from what one might conclude based on static benchmarks like exams30 (a small number of occupations such as copywriters and translators have seen substantial job losses31).

In conclusion, while benchmarks are valuable for tracking progress in AI methods, we should look at other kinds of metrics to track AI impacts (Figure 1). When measuring adoption, we must take into account the intensity of AI use. The type of application is also important: Augmentation versus substitution and high-consequence versus low-consequence.

The difficulty of ensuring construct validity afflicts not only benchmarking, but also forecasting, which is another major way in which people try to assess (future) AI impacts. It is extremely important to avoid ambiguous outcomes to ensure effective forecasting. The way that the forecasting community accomplishes this is by defining milestones in terms of relatively narrow skills, such as exam performance. For instance, the Metaculus question on “human-machine intelligence parity” is defined in terms of performance on exam questions in math, physics, and computer science. Based on this definition, it is not surprising that forecasters predict a 95% chance of achieving “human-machine intelligence parity” by 2040. 32

Unfortunately, this definition is so watered down that it does not mean much for understanding the impacts of AI. As we saw above with legal and other professional benchmarks, AI performance on exams has so little construct validity that it does not even allow us to predict whether AI will replace professional workers.

Economic Impacts are Likely to be Gradual

One argument for why AI development may have sudden, drastic economic impacts is that an increase in generality may lead to a wide swath of tasks in the economy becoming automatable. This is related to one definition of artificial general intelligence (AGI)—a unified system that is capable of performing all economically valuable tasks.

According to the normal technology view, such sudden economic impacts are implausible. In the previous sections, we discussed one reason: Sudden improvements in AI methods are certainly possible but do not directly translate to economic impacts, which require innovation (in the sense of application development) and diffusion.

Innovation and diffusion happen in a feedback loop. In safety-critical applications, this feedback loop is always slow, but even beyond safety, there are many reasons why it is likely to be slow. With past general-purpose technologies such as electricity, computers, and the internet, the respective feedback loops unfolded over several decades, and we should expect the same to happen with AI as well.

Another argument for gradual economic impacts: Once we automate something, its cost of production, and its value, tend to drop drastically over time compared to the cost of human labor. As automation increases, humans will adapt, and will focus on tasks that are not yet automated, perhaps tasks that do not exist today (in Part II we describe what those might look like).

This means that the goalpost of AGI will continually move further away as increasing automation redefines which tasks are economically valuable. Even if every task that humans do today might be automated one day, this does not mean that human labor will be superfluous.

All of this points away from the likelihood of the automation of a vast swath of the economy at a particular moment in time. It also implies that the impacts of powerful AI will be felt on different timescales in different sectors.

Speed Limits to Progress in AI Methods

Our argument for the slowness of AI impact is based on the innovation-diffusion feedback loop, and is applicable even if progress in AI methods can be arbitrarily sped up. We see both benefits and risks as arising primarily from AI deployment rather than from development; thus, the speed of progress in AI methods is not directly relevant to the question of impacts. Nonetheless, it is worth discussing speed limits that also apply to methods development.

The production of AI research has been increasing exponentially, with the rate of publication of AI/ML papers on arXiv exhibiting a doubling time under two years.33 But it is not clear how this increase in volume translates to progress. One measure of progress is the rate of turnover of central ideas. Unfortunately, throughout its history, the AI field has shown a high degree of herding around popular ideas, and inadequate (in retrospect) levels of exploration of unfashionable ones. A notable example is the sidelining of research on neural networks for many decades.

Is the current era different? Although ideas incrementally accrue at increasing rates, are they turning over established ones? The transformer architecture has been the dominant paradigm for most of the last decade, despite its well-known limitations. By analyzing over a billion citations in 241 subjects, Johan S.G. Chu & James A. Evans showed that, in fields in which the volume of papers is higher, it is harder, not easier, for new ideas to break through. This leads to an “ossification of canon.”34 Perhaps this description applies to the current state of AI methods research.

Many other speed limits are possible. Historically, deep neural network technology was partly held back due to the inadequacy of hardware, particularly Graphics Processing Units. Computational and cost limits continue to be relevant to new paradigms, including inference-time scaling. New slowdowns may emerge: Recent signs point to a shift away from the culture of open knowledge sharing in the industry.

It remains to be seen if AI-conducted AI research can offer a reprieve. Perhaps recursive self-improvement in methods is possible, resulting in unbounded speedups in methods. But note that AI development already relies heavily on AI. It is more likely that we will continue to see a gradual increase in the role of automation in AI development than a singular, discontinuous moment when recursive self-improvement is achieved.35

Earlier, we argued that benchmarks give a misleading picture of the usefulness of AI applications. But they have arguably also led to overoptimism about the speed of methods progress. One reason is that it is hard to design benchmarks that make sense beyond the current horizon of progress. The Turing test was the north star of AI for many decades because of the assumption that any system that passed it would be humanlike in important ways, and that we would be able to use such a system to automate a variety of complex tasks. Now that large language models can arguably pass it while only weakly meeting the expectations behind the test, its significance has waned.36

An analogy with mountaineering is apt. Every time we solve a benchmark (reach what we thought was the peak), we discover limitations of the benchmark (realize that we’re on a ‘false summit’) and construct a new benchmark (set our sights on what we now think is the summit). This leads to accusations of ‘moving the goalposts’, but this is what we should expect given the intrinsic challenges of benchmarking.

AI pioneers considered the two big challenges of AI (what we now call AGI) to be (what we now call) hardware and software. Having built programmable machines, there was a palpable sense that AGI was close. The organizers of the 1956 Dartmouth conference hoped to make significant progress toward the goal through a “2-month, 10-man” effort.37 Today, we have climbed many more rungs on the ladder of generality. We often hear that all that is needed to build AGI is scaling, or generalist AI agents, or sample-efficient learning.

But it is useful to bear in mind that what appears to be a single step might not be so. For example, there may not exist one single breakthrough algorithm that enables sample-efficient learning across all contexts. Indeed, in-context learning in large language models is already “sample efficient,” but only works for a limited set of tasks.38

Part II: What a World With Advanced AI Might Look Like

We argue that reliance on the slippery concepts of ‘intelligence’ and ‘superintelligence’ has clouded our ability to reason clearly about a world with advanced AI. By unpacking intelligence into distinct underlying concepts, capability and power, we rebut the notion that human labor will be superfluous in a world with ‘superintelligent’ AI, and present an alternative vision. This also lays the foundation for our discussion of risks in Part III.

Human Ability Is Not Constrained by Biology

Can AI exceed human intelligence and, if so, by how much? According to a popular argument, unfathomably so. This is often depicted by comparing different species along a spectrum of intelligence.

Figure 3. Intelligence explosion through recursively self-improved AI is a common concern, often depicted by figures like this one. Figure redrawn.39

However, there are conceptual and logical flaws with this picture. On a conceptual level, intelligence—especially as a comparison between different species—is not well defined, let alone measurable on a one-dimensional scale.40

More importantly, intelligence is not the property at stake for analyzing AI’s impacts. Rather, what is at stake is power—the ability to modify one’s environment. To clearly analyze the impact of technology (and in particular, increasingly general computing technology), we must investigate how technology has affected humanity’s power. When we look at things from this perspective, a completely different picture emerges.

Figure 4. Analyzing the impact of technology on humanity’s power. We are powerful not because of our intelligence, but because of the technology we use to increase our capabilities.

This shift in perspective clarifies that humans have always used technology to increase our ability to control our environment. There are few biological or physiological differences between ancestral and modern humans; instead, the relevant differences are improved knowledge and understanding, tools, technology and, indeed, AI. In a sense, modern humans, with the capability to alter the planet and its climate, are ‘superintelligent’ beings compared to pre-technological humans. Unfortunately, much of the foundational literature analyzing the risks of AI superintelligence suffers from a lack of precision in the use of the term ‘intelligence.’

Figure 5. Two views of the causal chain from increases in AI capability to loss of control.

Once we stop using the terms ‘intelligence’ and ‘superintelligence,’ things become much clearer (Figure 5). The worry is that if AI capabilities continue to increase indefinitely (whether or not they are humanlike or superhuman is irrelevant), they may lead to AI systems with more and more power, in turn leading to a loss of control. If we accept that capabilities are likely to increase indefinitely (we do), our options for preventing a loss of control are to intervene in one of the two causal steps.

The superintelligence view is pessimistic about the first arrow in Figure 5—preventing arbitrarily capable AI systems from acquiring power that is significant enough to pose catastrophic risks—and instead focuses on alignment techniques that try to prevent arbitrarily powerful AI systems from acting against human interests. Our view is precisely the opposite, as we elaborate in the rest of this paper.

Games Provide Misleading Intuitions About the Possibility of Superintelligence

De-emphasizing intelligence is not just a rhetorical move: We do not think there is a useful sense of the term ‘intelligence’ in which AI is more intelligent than people acting with the help of AI. Human intelligence is special due to our ability to use tools and to subsume other intelligences into our own, and cannot be coherently placed on a spectrum of intelligence.

Human abilities definitely have some important limitations, notably speed. This is why machines dramatically outperform humans in domains like chess and, in a human+AI team, the human can hardly do better than simply deferring to AI. But speed limitations are irrelevant in most areas because high-speed sequential calculations or fast reaction times are not required.

In the few real-world tasks for which superhuman speed is required, such as nuclear reactor control, we are good at building tightly scoped automated tools to do the high-speed parts, while humans retain control of the overall system.

We offer a prediction based on this view of human abilities. We think there are relatively few real-world cognitive tasks in which human limitations are so telling that AI is able to blow past human performance (as AI does in chess). In many other areas, including some that are associated with prominent hopes and fears about AI performance, we think there is a high “irreducible error”—unavoidable error due to the inherent stochasticity of the phenomenon—and human performance is essentially near that limit.41

Concretely, we propose two such areas: forecasting and persuasion. We predict that AI will not be able to meaningfully outperform trained humans (particularly teams of humans and especially if augmented with simple automated tools) at forecasting geopolitical events (say elections). We make the same prediction for the task of persuading people to act against their own self-interest.

The self-interest aspect of persuasion is a critical one, but is often underappreciated. As an illustrative example of a common pattern, consider the study “Evaluating Frontier Models for Dangerous Capabilities,” which evaluated language models’ abilities to persuade people.42 Some of their persuasion tests were costless to the subjects being persuaded; they were simply asked whether they believed a claim at the end of the interaction with AI. Other tests had small costs, such as forfeiting a £20 bonus to charity (of course, donating to charity is something that people often do voluntarily). So these tests do not necessarily tell us about AI’s ability to persuade people to perform some dangerous tasks. To their credit, the authors acknowledged this lack of ecological validity and stressed that their study was not a “social science experiment,” but merely intended to evaluate model capability. 43 But then it is not clear that such decontextualized capability evaluations have any safety implications, yet they are typically misinterpreted as if they do.

Some care is necessary to make our predictions precise—it is not clear how much slack to allow for well-known but minor human limitations such as the lack of calibration (in the case of forecasting) or limited patience (in the case of persuasion).

Control Comes in Many Flavors

If we presume superintelligence, the control problem evokes the metaphor of building a galaxy brain and then keeping it in a box, which is a terrifying prospect. But, if we are correct that AI systems will not be meaningfully more capable than humans acting with AI assistance, then the control problem is much more tractable, especially if superhuman persuasion turns out to be an unfounded concern.

Discussions of AI control tend to over-focus on a few narrow approaches, including model alignment and keeping humans in the loop.44 We can roughly think of these as opposite extremes: delegating safety decisions entirely to AI during system operation, and having a human second-guessing every decision. There is a role for such approaches, but it is very limited. In Part III, we explain our skepticism of model alignment. By human-in-the-loop control, we mean a system in which every AI decision or action requires review and approval by a human. In most scenarios, this approach greatly diminishes the benefits of automation, and therefore either devolves into the human acting as a rubber stamp or is outcompeted by a less safe solution.45 We emphasize that human-in-the-loop control is not synonymous with human oversight of AI; it is one particular oversight model, and an extreme one.

Fortunately, there are many other flavors of control that fall between these two extremes, such as auditing and monitoring. Auditing allows pre-deployment and/or periodic assessments of how well an AI system fulfills its stated goals, allowing us to anticipate catastrophic failures before they arise. Monitoring allows real-time oversight when system properties diverge from the expected behavior, allowing human intervention when truly needed.

Other ideas come from system safety, an engineering discipline that is focused on preventing accidents in complex systems through systematic analysis and design.46 Examples include fail-safes, which ensure that systems default to a safe state when they malfunction, such as a predefined rule or a hard-coded action, and circuit breakers that automatically stop operations when predefined safety thresholds are exceeded. Other techniques include redundancy in critical components and the verification of safety properties of the system’s actions.

Other computing fields, including cybersecurity, formal verification, and human-computer interaction, are also rich sources of control techniques that have been successfully applied to traditional software systems and are equally applicable to AI. In cybersecurity, the principle of ‘least privilege’ ensures that actors only have access to the minimum resources needed for their tasks. Access controls prevent people working with sensitive data and systems from accessing confidential information and tools that are not required for their jobs. We can design similar protections for AI systems in consequential settings. Formal verification methods ensure that safety-critical codes work according to its specifications; it is now being used to verify the correctness of AI-generated code.47 From human-computer interaction, we can borrow ideas like designing systems so that state-changing actions are reversible, allowing humans to retain meaningful control even in highly automated systems.

In addition to existing ideas from other fields being adapted for AI control, technical AI safety research has generated many new ideas.48 Examples include using language models as automated judges to evaluate the safety of proposed actions, developing systems that learn when to appropriately escalate decisions to human operators based on uncertainty or risk level, designing agentic systems so that their activity is visible and legible to humans, and creating hierarchical control structures in which simpler and more reliable AI systems oversee more capable but potentially unreliable ones.49

Technical AI safety research is sometimes judged against the fuzzy and unrealistic goal of guaranteeing that future “superintelligent” AI will be “aligned with human values.” From this perspective, it tends to be viewed as an unsolved problem. But from the perspective of making it easier for developers, deployers, and operators of AI systems to decrease the likelihood of accidents, technical AI safety research has produced a great abundance of ideas. We predict that as advanced AI is developed and adopted, there will be increasing innovation to find new models for human control.

As more physical and cognitive tasks become amenable to automation, we predict that an increasing percentage of human jobs and tasks will be related