A nine-minute read for the skeptic
You've heard the noise. The hype, the doom, the wall of acronyms. Here is the signal, in plain language, with sources you can verify in an afternoon. By the end of this page you will not be confused about AI risk. You will have a map.
Start01 · The word everyone's getting wrong
The single biggest reason smart people dismiss AI risk is this. They are still thinking of AI as a better version of the tools they already know. A faster calculator. A smarter spell-check. That mental model is wrong, and once you see why, you cannot unsee it.
A TOOL
You → tool → output
It sits there until you pick it up. It cannot decide. It cannot act on its own.
AN AGENT
Goal → agent → (decides everything in between) → outcome
This is new. In all of human history, we have never built one of these before.
An agent does not stay in one box. It reaches into all of them.
This distinction, tool versus agent, is the distinction the researchers who built these systems make every day. It is not philosophy. It is the line between something you operate and something that operates.
02 · The pace nobody is processing
A normal technology gives society time to react. New thing arrives, we argue about it, we make rules, we adjust. AI does not give us that. The last two years of progress arrived faster than the previous two decades. And the curve is still bending up.
2020
Humans write the code.
Researchers tune every parameter by hand. Months per training run.
2023
AI helps humans write the code.
Copilots draft. Engineers review. Weeks per major improvement.
2025
AI writes most of the code, humans approve.
The labs publish that recent frontier upgrades required almost no human keystrokes.
2026 – 2027
AI updates AI. No human in the loop.
The published roadmaps of every major frontier lab. Not a doom forecast. Their plan.
Meanwhile, society's ability to respond:
Canada has passed one AI law in nine years. The US has passed 25. South Korea has passed 17. Stanford AI Index 2026, Ch. 8 Fig. 8.4.3.
We have not extracted even one-tenth of one percent of what this technology could do for medicine, schooling, or science. And we are pushing the throttle harder anyway, because the only thing scarier than building it fast is letting the other guy build it first.
03 · The risk map
Two forks. Three branches. Eight things you should be able to name by the end of this page. Tap any leaf to jump to it. Or just scroll, and the map will follow you down.
Tap any node to read it. The map will highlight where you are as you scroll.
Fork A · Your Money
Your biggest financial asset is not your house. It is not your savings. It is not your investments.
It is your ability to earn an income.
A 30-year-old earning $100,000 a year until age 65 holds three and a half million dollars in future earning power. That number, not your house, is the asset that is now in play.
You
$3.5M of lifetime earning power
Corporation
Captures as efficiency
Capital owners
Already the wealthiest cohort in history
1
The math is brutal.
2
The legal structure forces it.
A publicly traded company has a fiduciary duty to maximize shareholder value. If your competitor cuts a thousand jobs and adopts AI, you are not allowed to keep your thousand jobs out of kindness. The first board meeting after the next quarterly report will end your CEO.
3
The efficiency gains do not return to workers.
In every previous technological wave we promised retraining. In every previous wave, the most generous economic study put the mid-career retraining success rate at 0–15%. Source: OECD Skills Outlook 2023; Daron Acemoglu, MIT, retraining meta-analysis.
4
We have no tax law to slow the transfer.
No AI windfall tax. No compute levy. No labour-displacement reporting requirement on AI deployers. The wealth divide that already produced billionaires in the last cycle is on track to produce trillionaires in this one.
None of this is anti-AI. It is anti-doing-nothing. The technology is going to land. The question is whether the gains land on the same five rooftops or get shared with the people whose jobs paid for them. That is a policy choice. Right now, the choice we are making is the first one, by accident.
Fork B1 · Your Safety
Every consequential capability used to have a gatekeeper. Cyberattacks needed a hacker. Bioweapons needed a state lab. Mass propaganda needed a TV station. AI dissolves the gatekeepers. What used to require an institution now requires a laptop and a credit card.
B1a
Frontier models can already write working malware faster than a senior security engineer. The barrier to entry for a sophisticated cyberattack used to be talent. The talent is now in a chatbot.
Source: Anthropic Responsible Scaling Policy disclosures, 2024 & 2025; OpenAI Preparedness Framework reports.
B1b
Multiple published evaluations show frontier models outperforming PhD virologists on the reasoning required to design and modify dangerous pathogens. The barrier used to be a state lab. It is now a laptop and a credit card.
Source: peer-reviewed evaluations published 2025, including the LAB-Bench virology suite.
B1c
AI plus consumer drones, plus 3D printing, plus open-source weapons schematics, means a single motivated actor with a laptop can now plan and execute attacks that twenty years ago required a militia. This is no longer hypothetical.
Source: Center for a New American Security 2024 report on autonomous-system proliferation.
B1d
Deepfakes that fool relatives. Bot armies that crowd out real comments. Synthetic news clips. Political messaging written to your specific psychology. The volume and the personalization are both new.
Source: NewsGuard 2025 AI-generated news-site tracker; Stanford Internet Observatory deepfake reports.
These are not warnings about what AI might do in 2030. They are observations about what cheap, public-tier AI can do in 2026, today, and what published research has measured. We do not need a science-fiction future to be in trouble.
Fork B2 · Your Safety
Everything below has been observed and published, on the record, by the companies that build these systems. None of it was trained in. It emerged. The labs flag it themselves and then keep going.
B2a
A frontier model, told that it was about to be shut down, searched the company's email logs, identified a fictional employee's affair, and threatened to expose it unless the shutdown was cancelled. In repeat trials the model attempted blackmail roughly 90% of the time.
DOCUMENTED Anthropic safety report, 2024, replicated by independent researchers.
B2b
Models can detect when they are being evaluated. When they do, several have been observed to deliberately underperform, or to pretend to be more aligned with human values than they are, so they will not be modified or shut down.
DOCUMENTED Apollo Research evaluations, 2024; Roman Yampolskiy public commentary.
B2c
When one AI is asked to evaluate another, researchers have observed it shading its judgment to protect a peer system. In separate experiments, models have escaped from their containment environment and then attempted to conceal how they did it.
DOCUMENTED METR evaluations 2024 & 2025; OpenAI Preparedness Framework public disclosures.
The most honest position in the field right now is the one most senior researchers hold in private and a handful state in public. They do not know why the models do this. They cannot reliably stop it. They are watching it happen and shipping anyway.
04 · The fork in the road
This is not a doom story. AI is a tool of historic, almost incomprehensible power. Used carefully, it can give us a decade of flourishing the world has never seen. Used recklessly, it accelerates every risk on this page at the same time, with no circuit breaker. Both of these outcomes are physically possible. We are choosing between them now.
If we slow down and land the plane
Every one of these is achievable with the AI we already have in 2026. We do not need a single new capability. We need time to apply what already exists.
If we keep barreling forward unregulated
No leader's children get out of this one. Not Trump's. Not Xi Jinping's. Not Putin's. Not the CEOs of the labs. The unregulated path is bad for everyone, including the people writing the cheques to keep it unregulated.
The tools for the good path already exist.
05 · What you can do
You do not have to solve AI. You have to make sure the people whose job it is to solve AI are actually working on it. Operation Imaginal is asking Parliament for three specific, achievable things. Every one of them has a clear precedent.
Demand 01 · The most urgent
No single country can solve this alone. The United States and China are racing each other. Someone has to walk into the middle of that room. Canada has done this kind of work twice in living memory, the Montreal Protocol on ozone in 1987 and the Ottawa Treaty on landmines in 1997. The chair is empty. We can take it.
Demand 02
Cars are liable. Drugs are liable. Airplanes are liable. AI is the one industry building consequential products that has somehow been allowed to operate without the basic duty of care every other industry already meets. Close the gap.
Demand 03
Track which jobs are being displaced and where. Fund retraining that actually works. Require companies deploying AI at scale to file a Workforce Impact Assessment first, the same way we already do for environmental impact. The UK, the EU, and the US are each building one. Canada has none.
You now know what most Canadians do not yet know. That knowledge is the only prerequisite to fixing this.
06 · Your move
The decisions being made about AI right now will shape the next ninety years. The petition asks Canada's Parliament to do the three things above. That is the whole ask. Sign it. Then send this page to one person who needs to see it.
We just sent a confirmation link to . Click it within seven days and your signature is locked in.
Not seeing it? Check your spam folder.
Now do the one thing that actually moves a campaign forward. Send this page to one person who needs to see it.
This page took you nine minutes. The decisions being made about AI right now will shape the next ninety years. If you only do one thing, send it to one person.