My Probability Estimates for the AI Apocalypse
Yann LeCun recently tweeted that most leading AI figures think p(doom) estimates are complete bullshit and that existential risk is essentially zero but stay silent while the doomers attract disproportionate attention. I have been making the analytical case for that position across a hundred Substack pieces. This one does something different. It puts numbers on the scenarios.
Probability estimates force intellectual honesty in a way that qualitative hedging does not. If you believe something is possible, you should be able to say how possible. If you cannot assign a number, you probably do not have a well-formed belief. What follows are my honest estimates for the three scenarios that dominate AI doom discourse, with the reasoning behind each one. They are falsifiable. I will update them as evidence arrives.
Scenario 1: Mass Unemployment. Probability: 5 percent.
By mass unemployment I mean sustained unemployment above 20 percent driven primarily by AI displacement. That is Great Depression territory. It would require not just AI capability improvement but the simultaneous failure of multiple independent absorption mechanisms I have developed across previous pieces.
The Task Value Gradient would have to fail, meaning workers in jobs with steep gradients from routine to complex work would have to be unable to reallocate upward as AI handles the low end. The Keynesian Multiplier would have to fail, meaning domestic productivity gains from AI would not circulate through the economy to create new demand and new work. The Throughput Paradox would have to fail, meaning automating bottlenecks would not generate expanded activity in adjacent layers requiring more human coordination. The Liability Requirement would have to fail, meaning the entire tort and professional licensing infrastructure of common law economies would have to become irrelevant to AI deployment. Natural churn, retirement, and gradual diffusion would all have to be overwhelmed simultaneously.
Each of those mechanisms failing independently has some probability. Their joint failure is much less probable. The 5 percent reflects that.
The honest acknowledgment of genuine uncertainty in this estimate is the robot combination problem. My absorption frameworks were developed primarily for cognitive work displacement. Physical robotics combined with LLM reasoning is a different threat model that those frameworks do not fully address. A warehouse robot that can also reason about logistics, a surgical robot that combines physical precision with diagnostic inference, a construction system that integrates vision, planning, and manipulation, these attack the physical gradient jobs my framework currently treats as structurally safe. That combination faces its own ceiling problems: manufacturing capacity, cost curves, reliability in unstructured physical environments, and a Liability Requirement that operates even more strongly in physical domains than cognitive ones. But it is the scenario where I hold my frameworks most loosely.
The 20 percent unemployment threshold also matters for the definition. Could AI produce 10 percent sustained unemployment? That is more plausible, perhaps 15 to 20 percent likely, and would represent serious disruption even if it falls short of the mass unemployment scenario. The 5 percent is specifically for the Great Depression scale displacement that rewrites the basic structure of the labor market.
And if by p(doom) you mean the chance that 50 percent of all workers will be unemployed, that is essentially 0%. Not even close to happening due to the likelhood that one or more of the above framworks hold.
Scenario 2: Superintelligence. The Ant Analogy. Probability: 0.001 percent.
The ant analogy is the right way to define this scenario because it is specific enough to evaluate. Ants cannot conceptualize the problems humans solve. They cannot understand calculus, democracy, or the causes of their own displacement. The superintelligence scenario requires AI that operates on problems we cannot even conceptualize, that reasons in categories of thought unavailable to human minds, that makes humans as cognitively irrelevant to the new intelligence as ants are to us.
Nothing in current LLM architecture produces that. Next token prediction scaled arbitrarily does not generate new categories of thought. It generates more sophisticated pattern matching within the space of human generated text. The data ceiling is the decisive constraint. Every LLM is bounded by its training corpus. It can interpolate, recombine, and surface non-obvious connections within that corpus with impressive sophistication. It cannot transcend it. The ant level superintelligence scenario requires transcending not just the training data but the entire conceptual framework that training data encodes. There is no mechanism in current architectures by which that happens regardless of scale.
The 0.001 percent is not zero because intellectual honesty requires acknowledging physics we do not yet understand. Transformative surprises have happened before. But this scenario requires so many individually improbable things to happen in sequence, the data ceiling being overcome, reliable autonomous reasoning emerging from probabilistic next token prediction, recursive self-improvement actually working, all of this happening faster than human institutions can respond, that the joint probability is vanishingly small. The people asserting AGI 2027 as a realistic timeline for anything approaching this scenario are not engaging with the structural constraints the evidence actually reveals. If this happens, it is due to amazing, mind-blowing discoveries that are decades away.
Scenario 3: Sentience. Probability: 0 percent.
This is the scenario I am most confident about and the one where the standard epistemic hedging is least warranted.
Four billion years of evolution on the only planet we know of with the right conditions produced exactly one species with anything resembling the kind of self-aware subjective experience we mean when we say sentience. One. In four billion years. On a planet optimized by natural selection specifically for the emergence of complexity.
The Fermi paradox has an implication that is striking here. If sentience were easy to engineer from information processing, if consciousness emerged reliably from sufficiently sophisticated computation, we would expect to see evidence of it elsewhere in a universe containing hundreds of billions of galaxies each containing hundreds of billions of stars, most of which are older than our sun, many of which almost certainly host planets with the right conditions for chemistry. We see no such evidence. The silence of the universe is a very strong prior against the idea that consciousness is a software problem that sufficiently sophisticated code reliably solves.
We cannot out-engineer the universe. Four billion years of evolution operating under selection pressure specifically favoring survival and reproduction, with the entire complexity of biological chemistry available as substrate, produced consciousness exactly once in the observable record. The claim that we can replicate or exceed that outcome in decades by scaling a next-token predictor trained on human text requires solving the hard problem of consciousness, which remains genuinely unsolved, without acknowledging that this is what is being claimed.
The people making sentience claims about current LLMs are smuggling in a solution to the hardest problem in philosophy without showing their work. When someone tells you that Claude or GPT-4 might be sentient, they are making an implicit claim about the nature of consciousness that neither they nor anyone else can currently justify. The 0 percent is not dogmatic. It is the correct response to an extraordinary claim that is unsupported by any mechanism and contradicted by the entire evolutionary and cosmological record.
What the Numbers Add Up To
Mass unemployment: 5 percent. Ant level superintelligence: 0.001 percent. Sentience: 0 percent. The remaining roughly 95 percent is the scenario almost nobody in the doomer discourse is willing to call by its right name: genuinely transformative technology following the same S-curve as every previous general purpose technology, with real but manageable labor market disruption, near full employment maintained through the absorption mechanisms, and a discourse that looks embarrassingly alarmist in retrospect.
That 95 percent scenario is not a consolation prize. It is not a boring outcome. We are living through one of the most significant technological transitions in human history in real time. The productivity gains are real. The labor market disruption for specific populations is real and deserves serious policy attention. The technology will become invisible infrastructure within a generation, the way electricity and the internet did. And the people who spent the transition period warning about paperclip maximizers and ant-level superintelligence will look, in retrospect, like the people who warned that the railroad would drive passengers insane from the speed or that the telephone would allow the devil to enter your home.
I will check back on these numbers as evidence arrives. That is what falsifiable claims require. AGI 2027 is 7 months away. The easiest one to evaluate is already almost upon us.
About the Author
Sean Richey, Ph.D., is a Professor of Political Science at Georgia State University specializing in AI information environments and digital political communication.
Expert Witness & Consulting Services
Dr. Richey provides expert witness testimony, case review and analysis for counsel, survey methodology evaluation, and policy consulting on AI-associated information environments. Visit my website or email consulting@seanrichey.com.