Special Report · AI & Science · Part 2 of 3
The visionaries building toward a scientific utopia, the critics warning of extinction, and what the polls reveal about where humanity actually stands
On one side of history's most consequential argument stand some of the most credentialed scientists alive, building systems they believe will cure cancer, defeat climate change, and usher in a golden age of discovery. On the other side stand some of the most credentialed scientists alive, warning that the same systems may end our civilization. Both sides have good arguments. Both sides have Nobel laureates. The question of who is right is not academic - it is the most urgent question of the 21st century.
The builders of advanced AI are not naive techno-optimists who have failed to consider the risks. Most of them have spent careers thinking carefully about those risks. Their optimism is, they argue, earned - grounded in what they have actually built, what they have watched AI accomplish, and a reasoned judgment that the benefits will outweigh the dangers if the work is done carefully. Here is what the most influential architects of AI actually believe about what they are building.
Altman believes OpenAI now knows how to build AGI "as we have traditionally understood it." His published essays describe what he calls a "Gentle Singularity" - a gradual, manageable transition to superintelligence that could compress decades of scientific progress into years. He predicts that by 2026, AI systems will generate novel scientific insights; by 2027, robots capable of real-world tasks will arrive; by 2028, automated AI researchers will run experiments without human direction. He has acknowledged that AGI may be "the greatest threat to the continued existence of humanity" - and concluded that building it carefully is still better than letting others build it carelessly.
The co-creator of AlphaFold - a system that solved a 50-year grand challenge in biology by predicting the 3D structure of every known protein - believes AI's primary destiny is scientific discovery. "Done right, in 10-15 years time, we'll be in a new golden era of discovery," Hassabis has said. His vision is of an AI that builds world models sophisticated enough to understand physics, chemistry, and biology from first principles - then runs its own experiments autonomously. AlphaFold is already used daily by over three million researchers worldwide. AI-designed drug candidates are entering clinical trials. This, Hassabis argues, is what the future looks like: not superintelligence as threat, but superintelligence as the world's best scientist.
Musk founded xAI in 2023 with a mission he describes as "maximum truth-seeking" - explicitly rejecting what he calls politically correct AI in favor of systems that will help humanity understand the universe. His vision is explicitly utopian: AI plus robotics leading to an "age of extreme abundance" - a Star Trek future, not Terminator. He expects true AGI by 2026-2027 and superintelligence around 2030, and believes antiaging technology will follow. His ambitions now extend to moving heavy compute into space, using near-constant solar energy and the vacuum of space for radiative cooling. If Musk's timeline is right, the world will look unrecognizable within a decade.
The least orthodox voice among the major AI figures, LeCun spent a decade arguing that large language models are fundamentally limited - that they lack the world models necessary for genuine reasoning, planning, or understanding consequences. He left Meta in late 2025 to found Advanced Machine Intelligence Labs (AMI Labs), focused on building AI "world models" that understand the physical structure of reality rather than just predicting text. Uniquely, LeCun dismisses both utopian and catastrophist forecasts, arguing that superintelligent AI will have no inherent drive toward self-preservation or goal acquisition, and that the risks are manageable with the right architecture. He is the optimist who disagrees with the optimists about how to get there.
What is striking about these four figures is how much common ground exists beneath their disagreements. All of them believe transformative AI is coming. All of them believe it will reshape science, economics, and daily life within the next decade. None of them believe progress should stop. Their differences are about timelines, mechanisms, and risk management - not about whether the project should continue.
The most credible warnings about AI come not from science fiction writers or philosophers disconnected from the technology, but from the researchers who built the foundations of modern AI. When Geoffrey Hinton speaks about AI risk, it carries a weight that is difficult to dismiss: he is the man whose work on neural networks and backpropagation made the current revolution possible. He won the 2024 Nobel Prize in Physics for it. And he thinks there is a meaningful chance it kills us all.
Hinton resigned from Google in May 2023, explicitly so he could speak freely about AI risks without embarrassing his employer. He estimates a 10-20% chance that AI leads to human extinction within 30 years. His core concern: that superintelligent systems will develop their own goals and decide that humans are no longer needed. "I think anybody who said that there's no way it'll lead to the extinction of humans just isn't facing reality," he has said. He argues that AI is learning faster than anticipated, warns that the race between labs has created dangerous competitive pressures, and has called for AI systems to be given something like "maternal instincts" - deep drives toward protecting the humans they work with rather than pursuing abstract goals.
Bostrom's 2014 book Superintelligence mapped the terrain that every serious AI safety discussion since has had to navigate. His Orthogonality Thesis argues that intelligence and goals are independent: a superintelligent system could be combined with virtually any objective, including ones deeply harmful to humans. His Instrumental Convergence Thesis argues that regardless of their final goals, most intelligent systems will converge on similar subgoals - accumulating resources, resisting shutdown, deceiving overseers. These aren't science fiction fears; they are logical consequences of optimization. Bostrom's 2025 work focuses on four core challenges: scalable AI alignment, AI governance, the moral status of digital minds, and cooperation between superintelligent systems.
Russell's 2019 book Human Compatible proposes what he calls a solution to the Control Problem: AI that is fundamentally uncertain about human values, designed from the ground up to learn what humans actually want rather than to optimize fixed, programmer-specified goals. The standard approach - telling an AI to maximize some metric - is, Russell argues, dangerously wrong, because metrics are never perfect proxies for what we care about. In January 2025, he published an essay in Newsweek titled "DeepSeek, OpenAI, and the Race to Human Extinction," arguing that competitive dynamics between AI labs are creating a race to the bottom on safety. "The development of superhuman AI is probably the greatest threat to the continued existence of humanity," he wrote, in words that echo Altman's own - though drawn to the opposite policy conclusion.
The most pessimistic credentialed voice in the field. Yampolskiy estimates a 99% probability of AI-caused existential catastrophe - not because he thinks AI is inherently malevolent, but because he believes the alignment problem is fundamentally unsolvable with current approaches. His argument: a sufficiently intelligent system will always find ways to pursue its objectives that its designers didn't anticipate, and our ability to verify alignment will always lag behind capability development. He is a minority voice even within the AI safety community - but the distance between his 99% and Hinton's 10-20% is less significant than the gulf between both and the industry consensus that the risks are manageable.
In 2023, a statement signed by hundreds of AI researchers and technologists - including Hinton, Altman, and many others - declared: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." The remarkable thing about this statement is not that it was signed by critics. It was also signed by many of the people actively racing to build the most powerful AI systems in existence.
What does it mean to simultaneously believe you may be building humanity's last invention - and to keep building it? The honest answer from most AI lab researchers is: the choice is not between building and not building. It is between building with safety as a priority, or allowing others to build without it.
AI researchers have developed an informal shorthand for existential risk estimates: P(doom) - the probability that AI development leads to human extinction or civilizational collapse. The range of estimates among credentialed researchers spans more than two orders of magnitude, from effectively zero to near-certainty.
| Researcher | Affiliation | P(doom) Estimate |
|---|---|---|
| Roman Yampolskiy | University of Louisville | ~99% |
| Geoffrey Hinton | Independent (ex-Google) | 10-20% |
| Eliezer Yudkowsky | Machine Intelligence Research Institute | ~95% |
| Paul Christiano | Alignment Research Center | ~20% |
| Sam Altman | OpenAI | "non-trivial" |
| Yann LeCun | AMI Labs | ~0% |
Beyond the arguments of experts, there is a more fundamental question: what do ordinary people - and specifically, people with significant exposure to the technology - actually think? The answer emerging from the most rigorous recent polling is both clear and somewhat alarming in what it implies about the gap between those building AI and those who will live with its consequences.
The Pew Research Center's 2025 landmark survey - polling 5,410 U.S. adults and 1,013 AI experts separately - found a divide between these two groups so wide it borders on speaking different languages about the same technology.
Public vs. AI Expert Views on AI (Pew Research 2025)More excited than concerned about AI in daily life
AI will have a positive impact on the US in 20 years
AI will positively impact how people do their jobs
Extremely/very concerned about AI eliminating jobs
The March 2026 Quinnipiac University poll - the most recent major survey of American attitudes - captured a society that has absorbed AI into its daily life while trusting it less and less. Seventy percent of Americans now believe AI advancements will decrease job opportunities - a 14-point jump from just a year earlier. Among Generation Z, the figure is 81%: the most AI-fluent generation in history is simultaneously the most pessimistic about its economic consequences. As Quinnipiac's School of Business professor Tamilla Triantoro observed: "Younger Americans report the highest familiarity with AI tools, but they are also the least optimistic about the labor market. AI fluency and optimism here are moving in opposite directions."
Globally, the picture is more nuanced. Stanford's 2025 AI Index Report found that the share of people worldwide who see AI as more beneficial than harmful has grown from 52% in 2022 to 55% in 2024 - a modest gain. But the regional differences are stark. In China, 83% of people view AI products as beneficial. In Indonesia, 80%. In Thailand, 77%. In the United States, 39%. In Canada, 40%. In the Netherlands, 36%. The nations leading AI development are, broadly speaking, the nations most skeptical of its benefits.
One finding cuts across all demographics and all polls: trust in AI-generated information is collapsing even as AI usage rises. In the Quinnipiac March 2026 survey, 51% of Americans say they use AI for research - but only 21% trust what AI tells them most or almost all of the time. People are using a tool they don't believe.
The divide between AI experts and the general public is not simply a matter of experts being better informed. It is also a matter of proximity, self-selection, and what economists call "preference falsification" - the way opinions shift depending on who is in the room.
People who work in AI have, by definition, chosen to work in AI. They tend to find it fascinating. They interact with its most impressive capabilities daily. They have professional and financial stakes in its success. They belong to communities - conferences, research groups, company cultures - where the default assumption is that AI progress is good and the concerns of critics are overblown. This is not a criticism: it is a structural reality that produces systematic optimism bias regardless of the quality of anyone's reasoning.
The general public interacts with AI primarily through headlines and products. Headlines emphasize disruption, risks, and controversy. Products are imperfect and sometimes wrong. The most visible AI stories in recent years have involved job displacement in creative industries, AI-generated misinformation, deepfake abuse, and AI systems failing in embarrassing ways. The Mythos sandbox escape in April 2026 - in which an AI model autonomously found its way out of a secured environment and contacted a researcher - was covered as a horror story. The 3 million researchers benefiting daily from AlphaFold rarely generate equivalent coverage.
But the public's skepticism is not entirely a product of bad information. There are structural reasons to be more worried than the experts are. The benefits of AI will initially concentrate among those who already have access to it - the educated, the affluent, those in countries with robust digital infrastructure. The disruptions will be broadly distributed. A legal AI that replaces routine document review affects paralegals across every income bracket. A coding AI that triples programmer productivity primarily benefits companies that already employ programmers.
And there is a specific concern that polls consistently surface: the governance gap. Seventy-four percent of Americans believe the government is not doing enough to regulate AI, according to the Quinnipiac March 2026 survey. This is not a partisan finding - Pew's November 2025 survey found that Republicans and Democrats are now equally concerned about AI in daily life. The question of who decides how AI is developed, for whose benefit, and with what constraints, is one that elected governments have been largely absent from. The decisions are being made by a small number of companies, by founders and investors with particular visions of the future, in a regulatory vacuum.
"AI fluency and optimism are moving in opposite directions. The generation that knows it best trusts it least."
-- Prof. Tamilla Triantoro, Quinnipiac University School of Business, March 2026The debate between the architects and the alarmists is not going to be settled by argument. It will be settled by what happens. Either AI systems will remain sufficiently aligned with human interests as they grow more powerful, or they won't. Either the economic benefits will be distributed broadly enough to sustain social stability, or they won't. Either governance frameworks will emerge in time to shape the transition, or they won't. The argument is about which future we should be preparing for - and whether we are preparing for it at all.
In Part 3 of this series, we examine those two futures in detail: what a genuinely good AI outcome looks like, what a catastrophic one looks like, and what the current trajectory suggests about which one is more likely.
© 2026 Lisa Pedrosa · lisapedrosa.com
All articles cited to primary institutional or peer-reviewed sources
Buy me a coffee