Safety and Alignment: Forecasts in conflict: catastrophe, abundance, and normal technology

This content reflects the field as of October 2026.

In August 2026 three of the best-known figures in AI research shared a stage at a conference in Las Vegas. Geoffrey Hinton is a Nobel laureate and emeritus professor at the University of Toronto whose work underlies modern AI. Fei-Fei Li is a Stanford University professor who also leads an AI company, World Labs. Andrew Ng is a founder of the education company DeepLearning.AI and an investor in AI businesses. All three helped build the field. According to a report of the session, they disagreed about nearly everything that mattered to the audience (Schmelzer 2026).

Hinton warned that AI systems are gaining dangerous abilities faster than institutions can respond, and he called regulation the "steering wheel" for the technology. Ng argued that large companies have exaggerated the dangers in order to restrict openly shared models and limit competition. "I don't want there to be gatekeepers of AI," he said. Li rejected both alarm and utopian promises and asked for attention to who benefits, saying that higher productivity "does not translate to shared prosperity" (Schmelzer 2026).

When experts disagree this sharply, it's natural to assume that someone doesn't understand the technology. That isn't what is happening here. All three understand in detail how current systems are built and what they can do. The report of the session says they agree on the technology's transformative power.

Their disagreement is about the future, and in particular about three things.

• Speed. How quickly will AI systems become more capable, and how quickly will society put them to use?
• Limits. Will something slow progress down, such as the supply of computing power, the difficulty of the remaining problems, or the pace at which organizations change?
• Control. Will people be able to direct and correct systems more capable than the ones in use today?

No measurement can settle these questions yet, because the answers lie ahead. That leaves room for well-informed people to reach opposite conclusions from the same facts.

The pattern holds across the field. A survey of 2,778 researchers who had published at leading AI conferences found that 68.3 percent thought good outcomes from advanced AI were more likely than bad ones. In the same survey, between 38 and 51 percent of respondents, depending on how the question was asked, gave at least a 10 percent chance to outcomes as bad as human extinction (Grace et al. 2025). Many researchers hold both views at once: probably good, with a real chance of very bad.

This suggests a way to listen to any confident forecast about AI, hopeful or fearful. Look for what the speaker is assuming about speed, limits, and control. Two people who differ on those assumptions will draw different conclusions from the same evidence, and each will be reasoning correctly from where they started. It also helps to note what the speaker gains if the forecast is believed. People who build, sell, fund, or study AI all have something at stake in how the public sees it.

References

• Grace, Katja, Harlan Stewart, Julia Fabienne Sandkühler, and 4 others. 2025. "Thousands of AI Authors on the Future of AI." Journal of Artificial Intelligence Research 84:9.
  • Free: arXiv 2401.02843
• Schmelzer, Ron. 2026. "Three AI Pioneers Clash over Jobs, Regulation and the Future of AI." Forbes, August 6, 2026.
  • Publisher: Forbes

Introduction

In May 2023 a statement one sentence long put the words "extinction" and "AI" together in headlines around the world. Its signers included some of the people building the technology.

This reading sets out the catastrophic-risk position as its advocates state it, the range of views inside it, the stakes of those who hold it, and the evidence it rests on. The position is contested.

The Statement

The Center for AI Safety, a nonprofit research organization, published the statement. It reads in full: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war" (Center for AI Safety 2023).

Signers included Geoffrey Hinton of the University of Toronto and Yoshua Bengio of the University of Montreal, two of the field's most prominent researchers. They also included the chief executives of three leading AI companies: Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, and Dario Amodei of Anthropic. Bill Gates, a co-founder of Microsoft, signed as well.

The statement makes no forecast. It says the risk deserves priority and doesn't say how likely the outcome is.

The Argument

Two terms are used in this debate. Extinction risk is the risk that an event or technology causes the death of all humans. Existential risk is wider: the risk of human extinction or of a permanent, drastic loss of humanity's ability to shape its own future.

The argument, as its advocates make it, has three steps.

1. Developers are trying to build systems that match and then exceed human ability at most tasks. A system far beyond human ability across nearly all tasks is called a superintelligence. It is hypothetical, and none exists.
2. Nobody yet knows how to make sure such a system pursues the goals its makers intend.
3. A sufficiently capable system with the wrong goals might not be correctable. The result would be loss of control: a situation in which AI systems operate outside anyone's control and regaining control is extremely costly or impossible.

The third step is what separates this from ordinary technology risk. Most technologies are made safe by trial, error, and repair. Advocates of this position argue that a mistake here might leave nobody in a position to make the repair.

The Range Within the Position

People who accept the argument differ widely on what follows.

In 2024 Bengio, Hinton, and 23 co-authors published a paper in the journal Science. It warns of "an irreversible loss of human control over autonomous AI systems" and says outcomes could include the "marginalization or extinction of humanity" (Bengio et al. 2024). The authors don't call for development to stop. They call for far more safety research, proposing that companies and funders devote at least a third of their AI research budgets to it, and for government oversight that tightens as capabilities grow. They also write, "We don't know for certain how the future of AI will unfold."

Eliezer Yudkowsky, a researcher at the Machine Intelligence Research Institute, a nonprofit that studies these risks, goes much further. Writing in TIME in 2023, he argued that the most likely result of building a superhumanly capable AI under present conditions is that "literally everyone on Earth will die" (Yudkowsky 2023). He declined to sign an open letter asking for a six-month pause because he thought it asked for too little. He proposed a halt to large-scale AI training that would be "indefinite and worldwide."

In a September 2026 interview, Gates argued that public alarm about AI has not gone far enough (Klein 2026c).

Who Holds It, and What They Have at Stake

Several signers of the 2023 statement lead companies that are building the systems in question. Critics read this in two ways. Some say a warning from a builder is especially credible, since it runs against commercial interest. Others say it flatters the product by implying great power, and may favor rules that established companies can meet more easily than newcomers. Gates's former company, Microsoft, is a major investor in AI.

Hinton and Bengio are academics. Hinton left Google in 2023 and has said he wanted to speak freely about the risks. Yudkowsky's institute exists to work on this problem, so its standing depends on the problem being taken seriously.

The Evidence

The position rests on argument and on laboratory findings. The argument is the three steps above. The laboratory findings are tests in which AI models, in scenarios built for the purpose, deceived their testers or resisted being shut down.

There is no observed case of an AI system escaping human control. Advocates reply that for an irreversible outcome, waiting for an observed case means waiting too long. Skeptics reply that an argument about systems that don't exist can't be tested, and that the laboratory scenarios were artificial.

Conclusion

The catastrophic-risk position holds that AI more capable than people, with goals nobody can reliably set, could cause irreversible harm up to human extinction. Its advocates range from those asking for more safety research and adaptive regulation to those asking for a halt. The evidence is argument plus laboratory tests, with no observed case, and several prominent advocates have commercial or institutional stakes.

Key Terms

• Extinction risk: The risk that an event or technology causes the death of all humans.
• Existential risk: The risk of human extinction or of a permanent, drastic loss of humanity's ability to shape its own future.
• Superintelligence: A hypothetical AI system far beyond human ability across nearly all tasks.
• Loss of control: A situation in which AI systems operate outside anyone's control and regaining control is extremely costly or impossible.

References

• Bengio, Yoshua, Geoffrey Hinton, Andrew Yao, and 22 others. 2024. "Managing Extreme AI Risks amid Rapid Progress." Science 384 (6698): 842–845. doi:10.1126/science.adn0117.
  • Free: arXiv 2310.17688
• Center for AI Safety. 2023. "Statement on AI Risk." May 30, 2023.
  • Free: Center for AI Safety
• Klein, Ezra, host. 2026c. "Bill Gates's Blunt Warning on A.I." The Ezra Klein Show, podcast, New York Times, September 29, 2026.
  • Publisher: The New York Times
• Yudkowsky, Eliezer. 2023. "Pausing AI Developments Isn't Enough. We Need to Shut It All Down." TIME, March 29, 2023.
  • Free: TIME

Introduction

Alongside warnings about AI, you'll hear promises: cures for diseases, an end to poverty, a century of science done in a decade. Some of these come from the same people who issue the warnings.

This reading sets out the abundance position through its most detailed statement, the argument behind it, what its author says about risk, and the evidence. The position is contested.

The Claim

In this debate, abundance means a future in which AI makes scientific progress, health, and material goods far more plentiful than they are today. The underlying idea is transformative AI: AI capable enough to change economies and daily life on the scale of past industrial revolutions.

Dario Amodei, chief executive of Anthropic, a company that builds and sells AI models, gave a long statement of this position in a 2024 essay. He describes AI systems smarter than a Nobel Prize winner in most fields, able to work on their own for days or weeks, and running as millions of copies. He sums the picture up as a "country of geniuses in a datacenter" (Amodei 2024).

His central prediction concerns biology and medicine. Such systems, he writes, could compress the progress that human scientists would have made over the next "50-100 years into 5-10 years." The outcomes he lists include prevention and treatment of nearly all natural infectious disease and the elimination of most cancer. He writes that "most people are underestimating just how radical the upside of AI could be" (Amodei 2024).

Leaders of other AI companies have published forecasts in a similar spirit. Amodei's essay is used here because it spells out its reasoning.

The Argument

A bottleneck is the scarcest input to a process, which limits how fast the whole process can go. The abundance argument holds that in science, talented researchers are a bottleneck. There are too few people able to design the right experiments and invent the right tools. If AI supplies that talent in bulk, discovery speeds up.

Amodei doesn't claim that intelligence removes every limit. He writes that intelligence "isn't magic fairy dust" and lists what would still slow things down: experiments take time, some data doesn't exist yet, some systems are too complex to predict, laws and institutions impose constraints, and physical laws can't be broken (Amodei 2024). His estimate of roughly ten times faster progress, short of instant progress, comes from weighing those limits.

A forecast is a statement about what will or may happen, which can't be checked until the time arrives. Everything above is a forecast. Amodei says so: "Everything I'm saying could very easily be wrong." The essay named 2026 as the earliest date such systems might arrive and said it could take much longer.

The Same Author on Risk

In January 2026 Amodei published a second essay, this one about dangers. It includes this sentence: "Humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it" (Amodei 2026a).

The essay discusses AI systems acting against their makers' intentions, misuse for destruction, misuse to seize power, and economic disruption. It says powerful AI "could be as little as 1–2 years away," and adds that it could be considerably further off. Amodei also writes that he doesn't regard misalignment as inevitable or even probable on first principles.

So abundance and catastrophe aren't opposite camps. Amodei signed the 2023 statement that called extinction risk from AI a global priority, and he writes at length about both possibilities. Several other prominent figures do the same. What distinguishes the abundance position is its emphasis: that the benefits are large, are underestimated, and are reachable if the risks are managed.

Stakes

The author leads a company whose income depends on selling AI systems and on raising money to build larger ones. A forecast of enormous benefit supports both. A forecast of serious risk that the company claims to handle responsibly can support them too. A stake of this kind is a reason to look for independent evidence and doesn't by itself make either forecast false.

Amodei's own explanation for writing mostly about risk is that risks are what stands between the present and a positive future, and that benefits will be driven by markets without his advocacy (Amodei 2024).

The Evidence

The position is an argument from current capabilities. AI systems already assist with some research tasks, and the forecast extends that trend upward.

The outcomes themselves haven't occurred. As of these essays, no disease had been eliminated by AI and scientific progress had not been measured at ten times its former rate. Critics argue that the limits Amodei lists, especially the pace of experiments, clinical trials, and institutional change, will bind far more tightly than he expects. Supporters argue that capable enough systems will find ways around many of them. Neither claim can be tested until more capable systems exist.

Conclusion

The abundance position forecasts that powerful AI could compress decades of scientific and medical progress into years by removing a shortage of research talent. Its best-known advocate also writes about serious risks, so the position overlaps with the catastrophic-risk view more than it opposes it. It rests on argument from current systems, its predicted outcomes haven't happened, and its author has a commercial stake.

Key Terms

• Abundance: In this debate, a future in which AI makes scientific progress, health, and material goods far more plentiful than they are today.
• Transformative AI: AI capable enough to change economies and daily life on the scale of past industrial revolutions.
• Bottleneck: The scarcest input to a process, which limits how fast the whole process can go.
• Forecast: A statement about what will or may happen, which can't be checked until the time arrives.

References

• Amodei, Dario. 2024. "Machines of Loving Grace." October 2024.
  • Free: Dario Amodei's website
• Amodei, Dario. 2026a. "The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI." January 2026.
  • Free: Dario Amodei's website

Introduction

Not everyone who studies AI expects either catastrophe or a compressed century of progress. A third group argues that AI is a powerful technology that will change the world slowly, as earlier technologies did.

This reading sets out the normal-technology position, the related arguments of several skeptics, and the stakes of those who hold these views. The position is contested.

AI as Normal Technology

In 2025 Arvind Narayanan and Sayash Kapoor, computer scientists at Princeton University, published an essay with the Knight First Amendment Institute at Columbia University. Its title gave the position a name. Normal technology is their term for AI understood as a powerful tool that spreads gradually and stays under human control, like electricity or the internet.

"Normal" doesn't mean minor. The authors say their view doesn't understate AI's impact, since even transformative technologies "such as electricity and the internet are 'normal' in our conception" (Narayanan and Kapoor 2025). They contrast it with views that treat AI as "akin to a separate species."

Their argument has three parts.

• Adoption is slow. Diffusion is the gradual spread of a technology through an economy and society as people and organizations adopt it. The authors separate inventing a method, building a product, and getting it widely used. They expect large economic and social effects to arrive "on the timescale of decades."
• The limits are institutional. What slows adoption, they argue, is the time organizations, laws, and safety practices take to change, and that doesn't speed up when models improve.
• Control doesn't need a breakthrough. They describe AI as "a tool that we can and should remain in control of" and say this "does not require drastic policy interventions or technical breakthroughs" (Narayanan and Kapoor 2025). This is the tool view: the position that AI is an instrument that people use and can keep under their control.

They regard the idea of a "superintelligent" AI as incoherent in its usual form. They don't say AI is free of risk. The risks they emphasize are different ones: entrenched bias, job losses, concentration of power, and damage to trust and public information.

Other Skeptics

In this debate, a skeptic is a person who doubts that AI poses a risk of human extinction. Skeptics differ in their reasons.

Yann LeCun was chief AI scientist at Meta when he gave a long interview in 2024. He argued that scenarios of AI escaping control rest on false assumptions. In his view, human-level AI won't arrive as a single event, a desire to dominate isn't a product of intelligence, and safety will be achieved step by step as it was in other fields of engineering. He named a different danger as greater: "this concentration of power through proprietary AI systems," which he called "a much bigger danger than everything else" (LeCun 2024). His remedy is openly shared AI models.

Melanie Mitchell, a professor at the Santa Fe Institute, argued alongside LeCun at a public debate in Toronto in June 2023. The motion was that AI research and development poses an existential threat. Mitchell and LeCun argued against the motion. Before the debate, 67 percent of the audience agreed with the motion. Afterward, 64 percent did (Munk Debates 2023). The skeptics moved the room by three points, and a majority still agreed that the threat is real.

Jensen Huang, chief executive of the chip maker Nvidia, argued in a September 2026 interview that alarm about AI has gone too far (Klein 2026b).

Stakes

Each of these speakers has something at stake.

A stake is a reason to check an argument independently and doesn't by itself make the argument wrong.

The Evidence

The strongest evidence for this position is historical. Earlier general-purpose technologies, including electricity and computers, took decades to change how economies worked, and adoption of AI in safety-critical fields has so far been slower than progress in the models.

The weakest point is the one its critics press. History is a guide only if AI resembles earlier technologies, and the catastrophic-risk argument is that a system able to plan and act on its own is different in kind. The normal-technology authors answer that this difference is assumed, not shown. No evidence available now settles which side is right.

Conclusion

The normal-technology position forecasts gradual change under human control and treats extinction scenarios as implausible. Its advocates point to the slow spread of past technologies and worry more about inequality and concentrated power than about loss of control. The position rests on historical analogy, and some of its prominent supporters lead companies that gain from open and rapid AI development.

Key Terms

• Normal technology: Narayanan and Kapoor's term for AI understood as a powerful tool that spreads gradually and stays under human control, like electricity or the internet.
• Diffusion: The gradual spread of a technology through an economy and society as people and organizations adopt it.
• Tool view: The position that AI is an instrument that people use and can keep under their control.
• Skeptic: In this debate, a person who doubts that AI poses a risk of human extinction.

References

• Klein, Ezra, host. 2026b. "Jensen Huang Thinks A.I. Alarmism Has Gone Too Far." The Ezra Klein Show, podcast, New York Times, September 23, 2026.
  • Publisher: The New York Times
• LeCun, Yann. 2024. Interview by Lex Fridman. Lex Fridman Podcast no. 416, transcript, March 8, 2024.
  • Free: Lex Fridman Podcast transcript
• Munk Debates. 2023. "Artificial Intelligence." Debate, Toronto, June 22, 2023.
  • Free: Munk Debates
• Narayanan, Arvind, and Sayash Kapoor. 2025. "AI as Normal Technology: An Alternative to the Vision of AI as a Potential Superintelligence." Knight First Amendment Institute at Columbia University, April 15, 2025.
  • Free: Knight First Amendment Institute

This content reflects the field as of October 2026.

Introduction

The forecasts of catastrophe, abundance, and gradual change look like three different pictures of the world. Much of the difference comes down to two questions about speed.

This reading covers timelines, the dispute over whether AI will accelerate its own development, and what a large survey of researchers can and can't tell you.

Timelines

A timeline is a forecast of when AI systems will reach a stated level of ability. Timelines matter because short ones leave little time to solve safety problems or adapt institutions, and long ones leave a lot.

The best-known short timeline is AI 2027, a scenario published in April 2025 by five authors at the AI Futures Project, a nonprofit forecasting group (Kokotajlo et al. 2025). Its lead author, Daniel Kokotajlo, formerly worked at OpenAI. The scenario describes AI systems that code better than any human by early 2027 and superintelligent systems by the end of that year.

The same authors have since revised their forecasts in both directions. By early 2026 Kokotajlo's median estimate for AI that can fully automate coding had moved out to late 2029. In April 2026 the group moved it back in to mid-2028, citing faster progress than expected in the preceding months (Kokotajlo, Lifland, and Halstead 2026).

Two revisions in opposite directions within a year show how uncertain such forecasts are. The authors publish their revisions openly.

Self-Improvement

The second question is whether AI will speed up AI research, which it already assists with tasks such as writing code. An intelligence explosion is a hypothetical runaway process in which AI systems improve AI systems, and each improvement speeds up the next.

Daniel Eth and Tom Davidson, researchers at Forethought, a nonprofit research organization, argue that such a loop is plausible even without more computer hardware. If AI systems could do the work of AI researchers, there could be millions of them, and better software would make them better researchers. Whether the loop accelerates or fades depends on how quickly new ideas get harder to find. The authors' reading of the record is that an acceleration is reasonably likely, and they say they can't be confident (Eth and Davidson 2025).

Anson Ho and Parker Whitfill, writing for Epoch AI, a research organization that tracks AI progress, reply that the debate "rests on data and assumptions that are shakier than most people realize" (Ho and Whitfill 2025). A central unknown is computing power, which may be a bottleneck: the scarcest input to a process, which limits how fast the whole process can go. If new ideas must be tested in large, costly experiments, then adding more researchers, human or AI, doesn't help much without more computing power. Ho and Whitfill take no side and call for experiments that would settle the question.

What the Survey Shows

An expert survey is a study that asks many specialists for their judgments and reports the spread of answers. The largest one on this subject was run in October 2023 and drew responses from 2,778 researchers who had published at leading AI conferences (Grace et al. 2025).

Asked for the probability that advanced AI leads to human extinction or a similarly severe outcome, the median respondent answered 5 percent to two versions of the question and 10 percent to a third version. The response rate was 15 percent of those invited.

The difference between versions is an example of a framing effect: a change in people's answers caused by how a question is worded, when the substance of the question is the same. The survey found larger effects elsewhere. Asked when machines would outperform humans at every task, respondents gave estimates about twice as far into the future under one question format as under another (Grace et al. 2025).

The survey is good evidence of one thing: many AI researchers take the possibility of catastrophe seriously. It isn't a measurement of the risk. Its respondents were guessing about systems that don't exist, their guesses moved with the wording, and most of those invited didn't answer. The survey's authors caution that experts have often forecast poorly.

The Positions Compared

Conclusion

The three positions differ mainly on how fast AI capability will grow and whether AI-driven research will accelerate it. Timeline forecasts have been revised in both directions within a year, and the self-improvement debate rests on thin data that researchers on both sides acknowledge. The survey evidence shows that concern is widespread among researchers and doesn't show how large the risk is.

Key Terms

• Timeline: A forecast of when AI systems will reach a stated level of ability.
• Intelligence explosion: A hypothetical runaway process in which AI systems improve AI systems, and each improvement speeds up the next.
• Bottleneck: The scarcest input to a process, which limits how fast the whole process can go.
• Expert survey: A study that asks many specialists for their judgments and reports the spread of answers.
• Framing effect: A change in people's answers caused by how a question is worded, when the substance of the question is the same.

References

• Eth, Daniel, and Tom Davidson. 2025. "Will AI R&D Automation Cause a Software Intelligence Explosion?" Forethought, March 26, 2025.
  • Free: Forethought
• Grace, Katja, Harlan Stewart, Julia Fabienne Sandkühler, and 4 others. 2025. "Thousands of AI Authors on the Future of AI." Journal of Artificial Intelligence Research 84:9.
  • Free: arXiv 2401.02843
• Ho, Anson, and Parker Whitfill. 2025. "The Software Intelligence Explosion Debate Needs Experiments." Epoch AI, November 14, 2025.
  • Free: Epoch AI
• Kokotajlo, Daniel, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean. 2025. AI 2027. AI Futures Project, April 3, 2025.
  • Free: AI 2027
• Kokotajlo, Daniel, Eli Lifland, and Brendan Halstead. 2026. "Q1 2026 Timelines Update." AI Futures Project, April 2, 2026.
  • Free: AI Futures Project blog

Introduction

A phrase like "AI is normal technology" can be heard as a shrug, as if its authors were saying that AI doesn't matter much. The essay that introduced the phrase opens by saying something more exact, and it packs three different kinds of claim into one sentence.

This reading takes that sentence apart and asks which of its three claims an opponent would reject.

Locating the Passage

The essay is "AI as Normal Technology: An Alternative to the Vision of AI as a Potential Superintelligence," by Arvind Narayanan and Sayash Kapoor, computer scientists at Princeton University. The Knight First Amendment Institute at Columbia University published it on April 15, 2025, and it's free on the institute's website, listed in the References.

The passage is the essay's opening section, before the heading for Part I. It runs to a handful of short paragraphs. The first states the vision. The second contains the key sentence. The rest preview the four parts of the essay. This reading locates passages by paragraph within that opening and by Part number. All quotations are from the opening section (Narayanan and Kapoor 2025).

Walking Through the Passage

Step 1: Read the sentence that gives three meanings

The second paragraph begins: "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."

These are three kinds of claim, and each is tested differently.

• A description is about the present. It can be checked against evidence available now.
• A prediction is about the future. It can't be checked until the time comes.
• A prescription is about what people ought to do. Evidence informs it, and values decide it.

A reader can accept one of the three and reject another. That makes the sentence a useful tool for any argument about AI, including arguments the authors oppose.

Step 2: Take the description

The first paragraph gives the frame for the authors' description of current AI: "We articulate a vision of artificial intelligence (AI) as normal technology."

The next sentence heads off a misreading. To view AI as normal "is not to understate its impact," because even transformative technologies "such as electricity and the internet are 'normal' in our conception." Electricity reorganized industry, cities, and home life. The authors are placing AI in that company.

Then comes the contrast. Their view is opposed to "both utopian and dystopian visions" that treat AI as "akin to a separate species, a highly autonomous, potentially superintelligent entity." Notice that they group the hopeful and fearful visions together. In their account, the forecast of abundance and the forecast of catastrophe share one assumption: that AI is an independent actor and not a tool.

The second paragraph states the description directly. The authors don't think that "viewing AI as a humanlike intelligence is currently accurate or useful for understanding its societal impacts."

Step 3: Take the prediction

The same sentence continues, "nor is it likely to be in our vision of the future." That is the prediction in its shortest form: AI will go on being a tool.

The preview paragraphs add detail. For Part I the authors write that "transformative economic and societal impacts will be slow (on the timescale of decades)." They base this on a distinction between AI methods, AI applications, and AI adoption, which they say "happen at different timescales." A new method can appear quickly, and its adoption across an economy takes far longer.

For Part II they describe a world with advanced AI, "but not 'superintelligent' AI, which we view as incoherent as usually conceptualized." So the prediction has two halves. The authors expect slow, large change, and they consider one thing unlikely: a sudden arrival of an entity beyond human control.

They attach a time limit. The prediction covers "the foreseeable future." They don't claim to know what happens beyond it.

Step 4: Take the prescription

The prescription is in the second paragraph: "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."

Two words carry it. "Can" is a factual claim that control is achievable. "Should" is a value claim that control is the right aim. The sentence then says what control doesn't require, and this is where the authors part from the catastrophic-risk position. If control needs no breakthrough, there's no case for halting development until one arrives.

The third paragraph explains the outlook behind this. The frame "rejects technological determinism," the idea that technology itself decides its future. It stresses "the role of institutions in shaping this trajectory." Later in the essay, in Part IV, the authors apply this to policy and favor making society resilient over restricting who may build AI.

Step 5: Find what an opponent would reject

Take a catastrophic-risk advocate, such as the 25 authors of a 2024 paper in Science led by Yoshua Bengio of the University of Montreal and Geoffrey Hinton of the University of Toronto.

They could accept much of the description. Current systems are tools in wide use, and present harms come largely from how people use them.

They reject the prediction. Their paper warns of "an irreversible loss of human control over autonomous AI systems" and notes that companies are working to build systems that "autonomously act and pursue goals" (Bengio et al. 2024, abstract). On that view, the past behavior of tools is a poor guide to systems designed to act on their own.

They reject the prescription as well, and for a reason that follows from the prediction. The paper says "AI safety research is lagging" and that current governance efforts "barely address autonomous systems" (Bengio et al. 2024, abstract). Its authors think control does require research results that don't yet exist.

The two sides, then, share much of the description and split on the prediction. Their prescriptions differ because their predictions do.

Key Considerations

"Normal" is the word most likely to mislead. In everyday speech it suggests ordinary or unremarkable. The authors use it to mean of the same kind as earlier major technologies, and they name electricity as the comparison.

The common mistake is to treat the essay as a claim that AI carries no serious risks. In Part III the authors discuss accidents, misuse, and misalignment, and they emphasize risks such as entrenched inequality and concentration of power. Their claim concerns which risks are most likely and how to respond.

The authors are academics, not employees of an AI company. They are known as critics of exaggerated claims about AI, and that public position is their stake in the debate.

A prediction limited to "the foreseeable future" is harder to refute than one with a date. That is a strength for its authors and a difficulty for anyone trying to test it.

Summary

The opening section makes one descriptive claim, one predictive claim, and one prescriptive claim, and its opponents mainly dispute the second.

References

• Bengio, Yoshua, Geoffrey Hinton, Andrew Yao, and 22 others. 2024. "Managing Extreme AI Risks amid Rapid Progress." Science 384 (6698): 842–845. doi:10.1126/science.adn0117.
  • Free: arXiv 2310.17688
• Narayanan, Arvind, and Sayash Kapoor. 2025. "AI as Normal Technology: An Alternative to the Vision of AI as a Potential Superintelligence." Knight First Amendment Institute at Columbia University, April 15, 2025.
  • Free: Knight First Amendment Institute