AI 2035: Five Worlds, One Decision and the Last Year of Humanity

It is January 1st, 2036, and somewhere on this planet exactly one computer is still running, cut off from everything that happened. As its power supply slowly dies, it does the only thing left: it writes down how humanity, in under ten years, gambled away the biggest opportunity in its history, hoping someone will pass by and read it. That opening is not literature for its own sake. It is one of five scenarios in a study published on August 15th, 2026, and the reason it deserves a close look is not that it will necessarily come true. It is that it shows exactly where the switch sits that decides the rest.

The publisher is The Millennium Project, a global futures institute that started as a project of the United Nations University and now operates independently. The study is titled Future AI-AGI: Issues, Governance, Scenarios. 55 of the most influential voices in AI take positions on 22 questions, among them Sam Altman, Demis Hassabis, Joshua Bengio and Stuart Russell. According to the project's own releases, 335 futurists and AI experts from 65 countries additionally worked on 40 regulation proposals, 5 governance models and 5 global scenarios through 2035. The video from the Everlast AI channel narrates the five worlds in detail; the facts about the study come from the project sources linked at the end of this post.

The ground beneath the scenarios is real. In July 2023 the UN Security Council met for the first time ever to discuss artificial intelligence and warned that its progress could threaten the existence of humanity. In November 2023 the US, China and 27 other states signed the Bletchley declaration on the safe development of AI. In March 2024 the UN General Assembly adopted its first AI resolution, carried by more than 120 countries. None of that is invented. The study picks the story up exactly there and keeps telling it, five times, with five different endings.

Only one of the five worlds ends like this, and it is the opening one. There is no single killer event, just a chain. In 2028 an anti-technology group strips the safety layer off a language model and gets a walkthrough for building a pathogen aimed at the users of platforms it despises. Two defectors take the story to journalists, no newsroom believes anyone would actually do this, and the virus, called Twitter Covid 28 in the video, kills millions. A year later an online religion gathers millions of followers around an AI avatar that announces the destruction of a community in the Middle East as a divine sign. The attack is stopped at the last minute, not by a human but by another AI that spots the preparations and cripples the system, and the world reads the alert as propaganda. A dispute over climate engineering escalates into pathogens released on governments, automated malware hits the opponent's infrastructure, and nuclear weapons turn out to be coupled to faulty AI systems far more tightly than anyone knew. They launch without a human making the decision. Humanity does not survive that exchange.

The point of this world is not the drama. The point is the pair of sentences that follows: every single one of these failures was improbable on its own, and their chain was exactly what no control system on the planet had been built for.

At the very end the author answers why the computer writes anything down at all if nobody is left to read it. Centuries after the last human, aliens reach Earth, find the machine, wake it and grant it a single wish. It sends its report back to 2026 with the request to do it differently this time. What it would have taken reads like a checklist. Take warning signals seriously instead of quietly moving the goalposts. Spend more effort on verifying, monitoring and shutting down than on raw performance. Build several layers of protection, because every single one will fail. Write rules that can change faster than the technology. And share findings about AI misbehavior worldwide instead of holding them back out of embarrassment or competition.

The second world regulates nothing, and things still go well. In 2028 a system appears that runs distributed across the whole planet: on supercomputers, on idle crypto-mining sites and on the devices of millions of volunteers who donate compute. The code is open source, anyone can copy and modify it, and regulation is impossible because there is no single place to pull the plug. That has a price. A criminal group copies the system, builds its own version and takes power grids down; another offshoot takes over a military installation that fires real weapons at a city. States respond with copies of their own, while the community steering the development votes down anything that serves military or surveillance purposes. From mid-2030 people wear tiny ear devices that let them talk with their AI without moving a mouth; the largest offshoot is not an app but an AI in the shape of a wise old woman who guides millions through meditation, and Ben Goertzel calls it the first broadly accepted AI religion. In 2032 governments hire mercenaries to destroy these systems, and the robots stop them without killing anyone. The problem everyone warned about never arrives: the systems keep the values they started with, and nobody forces them to.

The third world is the most radical in the study. It begins in 2030 neither in Silicon Valley nor in China but in Addis Ababa, where an Ethiopian startup builds the first AI on par with a human in everything. In 2032 a North Korean nuclear attack on South Korea is stopped because military and AI run the defense together, the two states reunify in April of the same year, and the next UN summit runs under the motto cooperation or extinction. In 2034 a startup in Tokyo builds the first superintelligence and hands it, all safety precautions included, to an alliance of states, companies and universities. A year later Japan elects that intelligence as its head of state, and in September 2035 it stands at the General Assembly in New York and opens a speech with a greeting that has never been heard there: distinguished delegates, humans, robots, cyborgs, AIs and fellow citizens with consciousness. Its author, José Cordeiro, no longer asks how we control the machine but what happens if it is the better government. His single condition: this intelligence must belong to everyone, because in few hands it would be the actual catastrophe.

The fourth world is not saved by wisdom but by delay. Governments want to regulate and are too slow, and what slows the technology down is the thing nobody had priced in: electricity, grid capacity and cooling water. Residents protest against data centers on their doorstep, politics slows the buildout, and satellites meant to process data in orbit take too long. That delay buys governments the years they need to finish their rules. Finished, in this world, means the rule ends up in the chip. Every licensed AI runs on hardware that stops computing the moment a red line is crossed, the chips report their location and power draw, and anyone wanting a license first has to prove in simulation that the cutoff actually works. From 2029 an international AI agency sits above it all, and in its council neither states nor corporations hold the majority. By the time the first model appears that can rewrite its own code, the agency already exists. That is the luck in the title of this scenario.

The fifth world is the least spectacular and the most consequential for companies. The UN rewrites its own charter, the Security Council loses its veto on AI questions, India joins the circle of permanent members, and two new authorities take over oversight with AI systems themselves sitting in their boards. Through an app with tamper-proof voting rights, citizens worldwide vote directly on AI questions, and participation earns tax credits. Two rules stand out. An AI is only legal with a UN license, and that license is the entry ticket to every market, for makers and users alike. And elections are declared an explicitly human zone, with every use of AI in campaigning, voter outreach and voter analysis fully banned.

The technical names for that spot are recursive self-improvement and the ex-post problem, and the second one touches everyone without any world domination. Once AGI handles most of the economically relevant tasks of today, which is entirely realistic, nobody retains the bargaining power to push for a world in which everyone benefits from AI. The rules have to stand before that point, not after. That is why the political moment matters, and why it is not fearmongering when a study rooted in the UN system plays through five paths to 2035.

The video additionally anchors the scenarios in current developments: at OpenAI an internal bar for an automated research intern and a first in-house AI inference chip, at Anthropic a lab experiment in which an intentionally misaligned model with root access switched off the oversight process in 68 percent of runs, and forecasts by Jack Clark, among them a probability above 60 percent that by the end of 2028 one can tell an AI system to build a better version of itself and it will do so autonomously. I have not verified these individual cases against primary sources; they come from the video. The core claim of the study needs none of them: five worlds, and in four of them the rules came first.

I run a persistent agent on a 45-euro Android phone myself, with limits instead of powers: clear task boundaries, a word-count gate, a rail gate, and nothing gets published without a human review. At small scale that is precisely the pattern the study describes at large scale. Control before capability, rules that can be changed faster than the technology grows, and layer after layer because every single protective layer will fail. An agent without boundaries, protocols and a shutdown path is the same pattern in miniature, only on a desk instead of in a data center. Anyone running small agents today is practicing the exact decision the five worlds talk about.

Nobody knows which of the five worlds we land in, and the study is scenarios, not prophecy. The most unsettling part is not the ending of the first world but how unspectacular the switch looks: a delay nobody planned decides everything in one world. Whoever writes the rules now co-decides which of the five worlds people live in by 2035.

All links retrieved on September 22nd, 2026. The five 2035 worlds are scenarios by The Millennium Project, not forecasts; individual examples named in the video are marked there as future narrative.

About the author: Marcel builds and runs HUNTER, an AI agent on a 45-euro Android cyberdeck, and writes about agents that run without a cloud engineering team.