Debaters, Look Up: AGI Is Three to Five Years Out
Why every argument on this year’s topic — and probably the next decade of them — runs through the AI question
If you are a competitive debater right now, you are arguing through the most important transition in human history, and you may not know it.
Last week at Google I/O, Sundar Pichai said he is “less able to predict with certainty whether it’s (AGI) in the three-to-five-year time frame or the five- to 10-year time frame” — but “the rate of progress over the last one to two years has made me feel it’s on the closer side.”
His co-presenter, Demis Hassabis, told the room something even more striking: we are “standing in the foothills of the singularity.” Hassabis, who runs Google DeepMind and won a Nobel Prize for using AI to solve protein folding, is not the type to throw around the s-word for fun. In Korea, he said, we are 3-4 years away from AGI.
He’s one of the most credible voices on the subject.
Three to five years. From the CEOs racing to build it.
Now read what Beth Barnes — founder of METR, the independent evaluation lab that Anthropic, Google, Meta, and OpenAI granted access to their internal models — said the same week, in a public statement about her organization’s first joint report on whether the frontier labs can keep control of their own agents:
“We are likely on track to develop AI systems capable of causing human extinction/permanent disempowerment, quite possibly within the next few years.”
Not a college freshman with a Substack. The person the four largest AI companies in the world picked to test whether their own models could escape their control. She also said this:
“Sometimes people outside the field say things like ‘The AI situation can’t be that bad, there must be experts who are on top of it.’ As ‘an expert,’ I would like to be clear that we are not on top of it.”
Existence. And, yes — the question Barnes is raising. Could we lose control? Eliezer Yudkowsky and Nate Soares — who run the Machine Intelligence Research Institute and who have spent two decades thinking about this problem — published a book this past fall whose thesis is right there in the title: If Anyone Builds It, Everyone Dies. The core claim, in their own words:
“If any company or group, anywhere on the planet, builds an artificial superintelligence using anything remotely like current techniques, based on anything remotely like the present understanding of AI, then everyone, everywhere on Earth, will die. We do not mean that as hyperbole. We are not exaggerating for effect.”
You do not have to believe Yudkowsky and Soares are right. Yoshua Bengio and Geoffrey Hinton — two of the three “godfathers of deep learning,” both Turing Award winners — signed the open letter calling AI extinction risk a global priority on par with pandemics and nuclear war. They aren’t as certain as Yudkowsky. But they are certain enough to sign their names to it. You do not have to share their conclusion to engage seriously with the argument. You do have to know it exists and know what answers to it look like.
That is the actual situation as the people closest to it describe it. Powerful systems arriving fast. Safety work woefully under-resourced. Labs that “regularly violate user intent” and “train on things they meant to avoid.” No serious plan for how to stay in control once the systems become smarter than the people building them.
This intersects all of your topics. And debate has not caught up.
Why every debater needs to understand this
I have been coaching debate for 35 years. I have watched topics come and go. I have watched arguments that seemed peripheral become central and central arguments quietly die. I am telling you, plainly, that AI is now the connective tissue of nearly every policy debate worth having.
Consider what AGI in 3–5 years actually means as a debate variable.
Warfare. Whatever you are reading about war powers, alliance commitments, deterrence, escalation, autonomous weapons — all of it gets rewritten when one side has access to recursively self-improving systems and the other doesn’t. The METR report focuses narrowly on whether current agents could escape human control. It explicitly excludes misuse risks — like AI helping a state actor or a terrorist plan a bioweapons attack. Those risks are not smaller than the loss-of-control risks. They are different.
Inequality. The economists you cite — Acemoglu, Autor, Brynjolfsson — are not arguing about whether AI displaces workers. They are arguing about how much, how fast, and whether the gains concentrate at the top. Kevin Frazier at the University of Texas, writing in Reason, calls this “the coming techlash” — and notes that only 17 percent of Americans believe AI will have a net positive impact on society over the next two decades. The technology that the labs are racing to build is the technology a supermajority of the country does not want. That is a debate. That is a lot of debates.
Work. Jensen Huang of Nvidia gave a long interview earlier this month making the case that AI will create more jobs than it eliminates — that radiologists are now busier, not gone; that small businesses will get a CFO they could never afford. Maybe. But that is a contested empirical claim, not a settled one. Activision laid off thousands of game developers last year as AI absorbed parts of their pipeline. Customer service, paralegal work, entry-level coding, marketing copy — those displacements are already happening. The question is whether the new work scales fast enough to absorb the displaced, and whether the workers being displaced have the runway to retrain. Those are debate questions.
The unemployed and the poor will need health care, though it could be provided very cheaply by AI.
Democratic governance. Pope Leo XIV, in his first encyclical, framed the problem this way: technological power is now “predominantly private” in a way it has never been before. The main drivers of AI development are transnational companies with resources that surpass most governments. Regulation matters — but regulation alone, the Pope warns, is not enough when the power being regulated outruns the regulators. That is a frame worth carrying into every governance debate this year.
AI overwhelms any other link to democracy.
What slows it down and speeds it up — and why every one of these is a link
Here is the part most debaters miss. AI is not a force of nature. It runs on chips, energy, capital, peace, and political license — and every one of those is contestable. Which means every one of those is a link.
Let me be explicit about what I mean by “link,” because this is exactly where most AI debate goes wrong.
If you are running an AI advantage — your aff makes AI better, faster, safer, more available — you need a story for what your plan does to one of those five inputs. You don’t just claim “AI good.” You claim, “the plan prevents X, X is what would have slowed AI down, therefore the plan speeds AI up.” That’s the link.
And here is the part that catches debaters off guard: avoiding bad things is also a link to AI bad. If your aff prevents an economic crash, you are protecting the capital that funds AI capex. If your aff prevents a war over Taiwan, you are protecting the fabs that produce two-thirds of the world’s GPUs. If your aff prevents a nuclear exchange, you are — per Yudkowsky and Soares’ own argument — preserving the substrate on which AI development continues. Negative-impact-avoidance arguments are AI-link arguments, whether or not the aff names them as such.
Here are the actual lever arguments and the evidence behind them.
Energy. A 2024 Goldman Sachs analysis projects U.S. data center power demand will grow at a 15% CAGR through 2030, with data centers’ share of total U.S. power demand rising from ~3% to ~8%. Brian Janous, former VP of Energy at Microsoft, says U.S. utilities — which haven’t seen consumption growth in two decades — “aren’t prepared.” Any aff that expands generation capacity (nuclear, natural gas, renewables, transmission) is an AI link. Any aff that constrains it (carbon tax, data center moratorium, renewables-only mandate) is also an AI link, in the opposite direction. Microsoft’s deal to restart Three Mile Island is the canonical example of how tightly energy and AI are now coupled.
Chips. Goldman’s semiconductor team expects HBM undersupply through 2026 and CoWoS packaging tightness for at least as long. Two-thirds of all GPUs in the world still flow through Taiwan. Dave Blundin put it bluntly this month: if anything happens to Taiwan, “everything we’re talking about just grinds to a halt.” TSMC has reportedly committed to destroying its fabs rather than let them fall to China. This is the cleanest AI link in the entire topic literature. Any aff that reduces the probability of a Taiwan contingency speeds AI up. Any neg argument that the plan increases Taiwan tensions slows it down. The CHIPS Act, export controls, alliance commitments — all of it routes through this link.
War. The Iran-Israel conflict is already complicating the $300+ billion that Gulf states had committed to AI infrastructure, according to The Information. Wars in the wrong places at the wrong times slow the buildout, fragment supply chains, redirect capital. Wars in the right places — from a “slow AI” perspective — could stop it entirely. Yudkowsky and Soares argue that the only realistic intervention left at this point may be coordinated international action that looks more like nuclear arms control than like tech regulation. Any aff that strengthens or weakens that international order is an AI link.
Regulation. The EU’s GDPR cut investment in European startups by 36% in the short run, according to NBER research, and the EU’s new AI Act layers on top of that. The U.S. has been more permissive. The argument that regulation “kills innovation” is real. The argument that no regulation lets a handful of private actors reshape society without democratic input is also real. Debaters should be fluent in both. Any aff that expands federal regulatory authority is potentially an AI link — slowing it. Any aff that preempts state regulation or limits agency authority is potentially an AI link the other way — speeding it up.
Economy. This is the one debaters miss most often. A real recession — not a soft landing, an actual downturn — would cut AI capex hard. So would a bursting of what some analysts believe is an AI investment bubble. So would a sustained techlash that makes the public refuse to adopt the products and makes Congress willing to regulate them. Kevin Frazier’s Reason piece this summer argued that “the coming techlash could kill AI innovation before it helps anyone” — and he notes that only 17 percent of Americans believe AI will have a net positive impact on society over the next two decades. Any aff that prevents economic collapse is an AI-acceleration link. Any aff that addresses the inequalities driving the techlash may reduce the techlash and therefore speed AI up. These are non-obvious links, but they are real.
Open source. The Bengio-led 2025 International AI Safety Report notes that state-of-the-art open-weight models like Llama-3.1-405B can fit on a USB stick. Once released, they cannot be recalled. That changes the regulatory math entirely. It also changes the misuse math. Any aff or neg argument about open-weight policy, export controls on model weights, or compute thresholds is an AI link.
Every single one of these is on the topic — whatever your topic is. They are the levers. If you don’t know them, you are arguing about AI without knowing how AI actually gets built. And the most important thing to internalize is this: debaters habitually under-link AI advantages and disadvantages because they think of AI as a separate scenario rather than as a downstream consequence of almost every other scenario. AI is not a separate impact. AI is the impact that almost every other impact eventually feeds into.
All of these links are in our evidence files.
What this asks of you
I am not asking you to become an AI doomer. I am not asking you to become an AI accelerationist. I am asking you to take the question seriously, the way the people building this technology and the people testing it take it seriously.
That means three things, practically.
First, read past the headlines. The METR report I quoted is online. Read Barnes’s thread. Read the report itself. Read Pope Leo’s encyclical. Watch the Pichai and Hassabis keynotes. Read the Bengio report. All of them are written for you.
You can find a lot of evidence here.
Second, learn the internal links. You cannot run a credible AI advantage if you don’t know what would slow AI down. You cannot run a credible AI disadvantage if you don’t know what would speed it up. The energy story, the chips story, the Taiwan story, the regulation story, the techlash story — these are the bones of the topic.
Third, ask the hard question. Beth Barnes asked it. Pope Leo asked it. Hassabis hinted at it. Where are we going? Toward what goal do we wish to orient ourselves? What direction should we choose as a people and as a human community? That is the deepest version of every debate question. If your case doesn’t have an answer to it, it isn’t finished.
Debate the question itself: Join the GlobalAI Debates
The CX, LD, and PF topics will only get you part of the way there. The question of where AI is going — and what we as a species should do about it — is too big to fit inside any single resolution.
That is why I co-founded the GlobalAI Debates with the Modus Ponens Institute. Last spring, students from Canada, China, Japan, South Korea, the UK, the US, and South Africa debated resolutions like “We call for a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in.” That is not a resolution you will see at NSDA Nationals. It is a resolution the AI safety field is actually debating right now, and high school students debated it at a high level.
The Fall 2026 GlobalAI Debates are open for registration. Thanks to support from the Future of Life Institute, this round has a $3,500 prize pool, with more financial support available for participants who need it. Judges have included researchers from MIT, Harvard, Perplexity, and OpenAI. The format is flexible: schools can submit written essays, video speeches, or compete in 2-vs-2 online debates. You can do it from anywhere. You can do it alongside your regular tournament season.
If the argument of this post is right — that AGI is three to five years out and that we are not ready — then the debaters who become fluent in this conversation now are going to be the most important people in the room in the rooms that matter most over the next decade. The GlobalAI Debates are the cleanest on-ramp to that fluency I know how to build. Register your team. Bring your students. Let them argue about the thing the world is actually about to argue about.
Resources
If you want to start digging, DebateUS has hundreds of free cards.
Subscribers get the full extension files, including the AGI-timeline evidence, the METR and Bengio report excerpts, and the war-good and war-bad blocks with full warrants.
But honestly? Start with the primary sources. Read the people building this thing. Read the people testing it. Decide for yourself.
The CEOs say three to five years. The safety researchers say we are not on top of it. The Pope says we cannot rely on regulation alone. The economists say the displacement is already happening.
You are competitive debaters. You are trained to weigh evidence under pressure, to argue both sides, to make judgments when the stakes are real. The world is about to need people who can do that more than it has needed almost anything else.