Benefits of superintelligence: what its builders promise, and what the evidence shows

The claimed benefits of superintelligence fall into five groups: much faster progress in science and medicine, far greater wealth, cheap clean energy, an expert assistant for every person, and protection against other global dangers. Every one of them is a forecast. As of November 2026 no superintelligent system exists, the evidence so far comes from narrow AI tools, and most of the largest promises come from the companies trying to build it.

What benefits do supporters of superintelligence claim?#

The idea that one machine could solve problems for everyone goes back at least to the mathematician I. J. Good. In 1965 he wrote that the first ultraintelligent machine would be the last invention humanity need ever make, “provided that the machine is docile enough to tell us how to keep it under control.” The benefit and the condition have traveled together ever since.

Today the promises come mostly from AI company leaders, and they are specific. The table sets out the main ones in the claimants’ own words.

One of the five promises is about individuals rather than discoveries: an expert assistant for every person. Suleyman writes that “Everyone who wants one will have a perfect and cheap AI companion helping you learn, act, be productive and feel supported.” Zuckerberg’s 2025 post “Personal Superintelligence” describes “a personal superintelligence that helps you achieve your goals.”

The claims share a shape. Each starts from what a much smarter system could do in principle and then names the problems it would solve. Several give a date: the next decade, the 2030s, before 2040. Few say how anyone would measure whether the gain had arrived. The company definitions guide explains what each firm means by the word, and the timelines guide lists when each expects it.

Could superintelligence speed up science and medicine?#

This is the benefit supporters stress most, and the one with the most evidence from today’s narrower systems.

What is claimed

Amodei’s October 2024 essay “Machines of Loving Grace” is the most detailed version. He predicts that AI-enabled biology and medicine “will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years.” His list includes “Reliable prevention and treatment of nearly all natural infectious disease,” the elimination of most cancer, a probable cure for most mental illness and a human lifespan of 150. He adds that the list “will almost certainly prove wrong in the details.”

Hassabis has made the shorter version of the claim. On 60 Minutes in April 2025 he said that designing one drug takes about “10 years and billions of dollars,” and that AI could “maybe reduce that down from years to maybe months or maybe even weeks.” Nick Bostrom’s 2026 working paper “Optimal Timing for Superintelligence” argues that superintelligence “would be able to enormously accelerate advances in biology and medicine.” Zuckerberg wrote in August 2026 that, with advances in AI, he expects the goal he and Priscilla Chan set for their Biohub, to help scientists cure or prevent all diseases this century, to be possible “much sooner.”

What narrow AI has shown so far

The record from systems far short of superintelligence is real but mixed.

• Protein structures. The 2024 Nobel Prize in Chemistry went half to David Baker “for computational protein design” and half jointly to Hassabis and John Jumper of Google DeepMind “for protein structure prediction,” the work behind AlphaFold.
• Diagnosis. In October 2026 Microsoft researchers reported in Science that their diagnostic system, which simulates a panel of physicians, reached 83.2% accuracy on New England Journal of Medicine case records, against 36.1% for generalist physicians. They note that the results were obtained “in a controlled benchmark setting.” Microsoft’s June 2025 preprint version had reported 85.5%, against 20% for physicians working without colleagues or textbooks, and Microsoft said then that the system was “not yet approved for clinical use.”
• Drugs. Rentosertib, a drug for idiopathic pulmonary fibrosis that Insilico Medicine discovered and designed with generative AI, reported a small 12-week phase 2a trial in Nature Medicine in June 2025. The trial was designed mainly to test safety. The drug is still investigational.
• Materials. In November 2023 Google DeepMind announced “the discovery of 2.2 million new crystals,” of which it called 380,000 the most stable. Two materials chemists, Anthony Cheetham and Ram Seshadri, examined a random sample and found “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility.” DeepMind said it stood by its paper.

Why more intelligence may not mean proportionally faster cures

Amodei’s own essay explains the main limit. He argues that “we should be talking about the marginal returns to intelligence” and names the factors that cap them: the speed of the outside world, the need for data, intrinsic complexity, constraints from humans and physical laws. The time a major project such as a cancer cure takes, he writes, “may have an irreducible minimum” however smart the researcher.

Emilia Javorsky of the Future of Life Institute made the same point more bluntly in an April 2026 podcast on why superintelligence will not cure cancer by itself: “We can’t scale the number of patients in a clinical trial,” she said, and “You can’t scale time.”

Would superintelligence make everyone richer?#

Supporters predict abundance. Published estimates of the gain range from modest to explosive, and most of the gap comes from what each one assumes AI can do.

Altman’s February 2025 post “Three Observations” says that “The cost to use a given level of AI falls about 10x every 12 months” and that “The price of many goods will eventually fall dramatically.” Amodei’s essay describes, as “a dream scenario,” 20% annual GDP growth in the developing world. He also writes that he is less confident about growth and inequality than about invention, because the economy “involves a lot of constraints from humans.”

Two published estimates show why.

The two are not really in conflict. Acemoglu’s paper is about the coming decade of AI, not about superintelligence. The Epoch paper asks what happens if AI can broadly substitute for human labor, which is close to what superintelligence would mean. The large economic promises depend on that second assumption.

A third view says even a capable technology spreads slowly. The Princeton computer scientists Arvind Narayanan and Sayash Kapoor argue in “AI as Normal Technology” (April 2025) that “Diffusion occurs over decades, not years,” and that sudden improvements in AI methods “do not directly translate to economic impacts.”

Could superintelligence help with energy and climate?#

Suleyman wrote in November 2025, “I predict we will have cheap and abundant renewable generation and storage before 2040.” Altman’s June 2025 essay also predicts abundant energy in the 2030s.

The near-term evidence points both ways. The International Energy Agency’s 2025 report “Energy and AI” projects that data centre electricity use will “more than double to around 945 TWh by 2030,” from 415 TWh in 2024. The same report finds that wider use of existing AI tools to run grids, factories and buildings more efficiently could cut energy use and emissions, though how much depends on whether those tools are adopted at scale.

Both findings are about today’s AI. The case for superintelligence on energy is that it would invent better technology, such as cheaper storage or new reactor designs, rather than run existing systems more efficiently. None of the sources cited here puts a number on that.

Could superintelligence make the world safer?#

Some supporters argue that the biggest benefit is protection. Bostrom’s 2026 paper says an aligned superintelligence “could also do much to enhance humanity’s collective safety against global threats,” for example by helping to avoid war and to counter new bioweapons.

The same paper turns the benefit into an argument about timing. Its abstract says developing superintelligence is “not like playing Russian roulette; it is more like undergoing risky surgery for a condition that will otherwise prove fatal.” Without it, Bostrom writes, “170,000 people die every day of disease, aging, and other tragedies.” He concludes that for many settings of his models the best course is to move quickly to AGI capability and then pause briefly before full deployment, and he warns that “poorly implemented pauses could do more harm than good.” He limits the paper to people alive today and sets aside more speculative considerations.

His earlier work pointed the other way on speed. Bostrom’s 2003 paper “Astronomical Waste” argued that each year of delayed technological development has a vast opportunity cost, but concluded that the lesson is “not that we ought to maximize the pace of technological development, but rather that we ought to maximize its safety.” Much of the difference is whose lives are counted. The 2003 paper’s main argument weighs the potential lives of future generations, and on that accounting safety comes first. The paper also said views focused on people who already exist should accept “a modified version of this conclusion.” The 2026 paper counts only people alive today, whose life expectancy, Bostrom argues, is higher if superintelligence arrives reasonably soon. The debate guide covers the risks these benefits are weighed against.

Who would get the benefits?#

Nearly every builder promises broad sharing. Few have said how it would work.

• OpenAI. Its Charter says the mission is to ensure that AGI “benefits all of humanity,” and one of its sections is headed “Broadly distributed benefits.” When OpenAI restructured its for-profit arm as a public benefit corporation in October 2025, its nonprofit, the OpenAI Foundation, announced “a $25B commitment” across health and curing diseases and AI resilience. In March 2026 it said it expected to invest at least $1 billion over the following year. These are pledges, not money already spent.
• Meta. Zuckerberg’s August 2026 essay “The Future is for Everyone” argues that “There is no such thing as a singular benevolent superintelligence” and that the answer is to ensure superintelligence “is broadly distributed to empower people.” Meta says it will offer free versions “accessible to billions of people.” See personal superintelligence.
• Bostrom. His 2014 book proposes a “common good principle”: superintelligence “should be developed only for the benefit of all of humanity and in the service of widely shared ethical ideals.”

Specific mechanisms are rarer. In 2021 Altman proposed an “American Equity Fund” paid for by a 2.5% annual tax on the value of large companies and of privately held land. A 2020 proposal from Oxford researchers, the Windfall Clause, would have AI firms commit in advance to donate a rising share of profits above 0.1% of gross world product, reaching 50% of profits above 10%. No company is known to have signed one. In June 2026 Senator Bernie Sanders released the text of a bill for an AI sovereign wealth fund. TechCrunch reported in July 2026 that it had not advanced to committee.

One sharing idea has been tested. An OpenResearch study gave 1,000 low-income people $1,000 a month for three years, with 2,000 people receiving $50 a month as a control. In the August 2026 revision of the paper, participants worked 1 to 2 fewer hours a week and their income excluding the transfers fell by about $1,900 a year relative to the control group. Gains in subjective well-being in the first year faded to control-group levels. The study shows how people respond to a guaranteed income. It says nothing about whether AI could pay for one.

Altman has himself written that “increasing equality does not seem technologically determined.” Whether the gains are shared is a policy question, not a property of the technology.

What do critics say about the promised benefits?#

The critics rarely deny that advanced AI could do good. They dispute whether the good requires superintelligence, and whether the promises are reliable.

• Benefits without superintelligence. The Statement on Superintelligence of October 2025, which calls for a prohibition on developing it, opens its context note with “Innovative AI tools may bring unprecedented health and prosperity.” Max Tegmark, who leads the Future of Life Institute, said that “What most people instead want are AI tools that — by definition — we can control.” Yoshua Bengio and colleagues proposed in February 2025 a non-agentic “Scientist AI,” arguing that this approach “may enable the benefits of AI innovation while avoiding the risks associated with the current trajectory.” See tool AI.
• Utopian promises as a pattern. Timnit Gebru and Émile P. Torres argued in First Monday in April 2024 that the language of “safety” and “benefiting humanity” is used “to evade accountability,” and that “undefined systems like ‘AGI’ cannot be appropriately tested for safety.”
• Slow diffusion. Narayanan and Kapoor’s view, above, is that capability and impact are separated by decades of adoption. They also describe superintelligence as “incoherent as usually conceptualized.”

The public is far less optimistic than AI experts. In an April 2025 Pew Research Center report, 56% of US AI experts expected AI to have a positive impact on the United States over the next 20 years, against 17% of US adults. In June 2026 polling, Pew found that 52% of Americans were more concerned than excited about the increased use of AI in daily life, and 9% were more excited than concerned. The concerned share was up from 37% in 2021 and close to the level it has held since 2023.

How should you judge a claimed benefit?#

Four questions do most of the work.

• Who is making the claim? Most of the largest promises come from the companies building the systems. That does not make them wrong, but it is a reason to look for an outside check.
• Is it a result or a forecast? AlphaFold’s structures are a result. A lifespan of 150 is a forecast.
• Who measured it? Microsoft’s diagnostic results and DeepMind’s materials count came from the companies’ own researchers. Microsoft’s work later passed peer review as a benchmark study. In the materials case, outside chemists disputed the claim.
• What else has to change? A new drug still needs trials, regulators and manufacturing. A cheaper good still needs to reach the people who need it.