AI Risks & Safety: Existential Threats, Kill Switches, Deepfakes—and What We Can Actually Do

From X-Risk and Job Displacement to the AI Kill Switch Act, Y2K Lessons and Why Your Roomba Still Can’t Beat a Laundry Basket

The greatest paradox of technological prevention is that success eventually becomes evidence that the danger never existed. Sometimes ‘nothing happened’ is not proof of hysteria. It is proof that somebody did the work.

Artificial intelligence may cure diseases, accelerate scientific discovery, make businesses vastly more productive and give every human being something resembling an extraordinarily capable intellectual assistant.

It could also destroy truth, displace millions of workers, supercharge cyberattacks, enable biological threats—or, according to some of the people actually building advanced AI, eventually threaten humanity itself.

So naturally, we have arrived at the calm, nuanced stage of technological development where everybody is discussing human extinction before breakfast.

Welcome to AI in 2026. 🤖☕

The question isn't whether AI creates risk. It clearly does.

The interesting question is:

How much?

Are we facing a genuine technological discontinuity requiring unprecedented safeguards?

Or are we watching another familiar human phenomenon—taking a transformative new technology, extrapolating its capabilities to infinity, adding The Terminator, and panicking?

And what about Y2K?

We were told computer systems might fail around the world at midnight on January 1, 2000.

Midnight arrived.

Planes didn't fall out of the sky.

ATMs didn't explode.

Civilization continued.

Was Y2K therefore ridiculous?

Not quite.

And therein lies perhaps the most useful lesson of the entire AI-doomsday debate.

🎯 FunTech Index™: 9.75 / 10

Few technological questions have a larger asymmetry between “probably fine” and “spectacularly not fine.”

AI's immediate risks—fraud, deepfakes, cyberattacks, bias, surveillance and labor disruption—already justify serious safeguards.

Existential-risk scenarios remain far more uncertain and contested. But when the downside being discussed is civilization-scale, uncertainty isn't itself an argument for doing nothing.

Verdict: Don't panic. Don't dismiss. Build guardrails, preserve human control, demand evidence—and keep the Roomba in perspective.

9.75/10. 🤖⚡

✅ FUNanc1al Atomic Statements

🗣️ The Risk-Asymmetry Principle

“AI safety does not require believing the apocalypse is probable. When an outcome is enormously consequential, even a small credible probability can justify disproportionately serious prevention.” — FUNanc1al

🗣️ The Y2K Paradox

“The greatest paradox of technological prevention is that success eventually becomes evidence that the danger never existed. Sometimes ‘nothing happened’ is not proof of hysteria. It is proof that somebody did the work.” — FUNanc1al

🗣️ The Human-Control Principle

“The most important AI question isn't whether machines become conscious. It's whether humans delegate enough authority that consciousness no longer matters.” — FUNanc1al

☣️ Part I — Could AI Actually Kill Us?

The cartoon version is easy.

AI wakes up.

AI hates humanity.

AI launches missiles.

Arnold Schwarzenegger arrives from 2029.

Roll credits.

The serious version is considerably less cinematic.

Researchers worried about existential risk generally aren't arguing that an AI will develop emotions, become irritated with humanity and decide:

“I've reviewed TikTok. Termination authorized.”

The concern is loss of control.

Imagine increasingly autonomous systems capable of writing software, conducting cyber operations, manipulating people, accessing external tools and pursuing complicated objectives.

Now imagine those systems becoming better at accomplishing goals than humans are at supervising them.

The risk is not necessarily evil.

It's capability + autonomy + badly specified objectives + insufficient control.

Some researchers now assign surprisingly high probabilities to catastrophic outcomes. Others regard near-term extinction scenarios as wildly speculative. That disagreement itself deserves to remain visible rather than being resolved artificially. 

⚡ Infrastructure

Highly capable AI agents could potentially make cyberattacks faster, cheaper and more autonomous.

Critical infrastructure—power, communications, transportation, finance, healthcare—is an obvious concern.

A future sufficiently capable autonomous system could conceivably exploit connected infrastructure, making cybersecurity and strict limits on real-world system access increasingly important.

🧬 Biology

AI may also lower the expertise required to perform dangerous biological work.

That doesn't mean ChatGPT is about to cook a pathogen between answering restaurant recommendations.

It means sufficiently capable systems could make specialized knowledge more accessible to malicious actors—one reason frontier-model evaluations increasingly examine biological and chemical capabilities.

☢️ Military Systems

Then there is perhaps the easiest rule in the present article:

Do not give an AI unilateral authority to start a nuclear war.

There.

FUNanc1al has solved one policy question.

You're welcome. 😂

💼 Part II — What If AI Doesn't Kill Us but Completely Rearranges Us?

This category deserves at least as much attention as extinction because we don't need hypothetical superintelligence to encounter it.

👔 Jobs

Previous technological revolutions automated enormous amounts of human labor while ultimately creating new industries and occupations.

AI is unusual because it reaches directly into cognitive labor.

Writing.

Programming.

Research.

Customer support.

Legal analysis.

Design.

Accounting.

Medicine.

Finance.

And, rather inconveniently for me, writing articles about AI replacing writers.

The most extreme “hyper-unemployment” forecasts remain forecasts. We don't yet know whether AI primarily substitutes for workers, complements them, creates entirely new occupations—or does all three at different speeds.

But occupational disruption is not hypothetical.

The transition has begun.

🎭 The Deepfake Infocalypse

This may be one of the most underrated AI risks.

When realistic synthetic speech, photographs and video become essentially free, the problem isn't merely that people believe fake things.

Eventually they may stop believing real things.

A politician caught saying something outrageous?

“Deepfake.”

CEO recording leaks?

“AI.”

Video evidence?

“Fake.”

The ultimate casualty isn't truth.

It's trust in the possibility of establishing truth.

And that's a substantially more dangerous problem.

🛡️ Part III — Fine. Where's the Giant Red OFF Button?

The following seems sensible:

If an AI becomes dangerous:

TURN. IT. OFF.

Congress has essentially arrived at the same insight.

Representatives Ted Lieu and Nathaniel Moran introduced the bipartisan AI Kill Switch Act, which would require developers of sufficiently advanced systems to retain the ability to slow, suspend or shut them down and would empower federal authorities to order intervention in defined loss-of-control situations.

Excellent.

Install the button.

Preferably not next to “Update Windows.”

But a kill switch isn't magic.

A shutdown mechanism works best when the system remains inside infrastructure that its operator controls.

If capable models are copied, distributed or running elsewhere, the problem becomes considerably harder.

And that's why serious AI safety needs layers:

Kill switches + sandboxing + restricted permissions + cybersecurity + model evaluations + red teaming + compute governance + incident reporting + human authorization + international rules.

No single safeguard deserves to carry civilization on its shoulders.

Especially a button.

🗓️ Part IV — But Isn't This Just Y2K 2.0?

This is where the present probe becomes much more interesting than another AI-doom explainer.

Because the following truth is paramount:

Fear distorts judgment.

Humans are spectacularly bad at thinking about new technologies.

We simultaneously underestimate transformative technologies and imagine that every unfamiliar technology is about to destroy civilization.

But Y2K teaches a subtler lesson.

The Millennium Bug was real.

Software systems had widely used abbreviated dates. Governments and businesses spent years identifying and remediating vulnerable systems.

Then January 1, 2000 arrived without catastrophe.

And people naturally concluded:

See? Nothing happened!

Yet that's precisely the problem with successful prevention.

If firefighters extinguish a fire before your house burns down, you don't conclude:

“What a scam. My house never burned down.”

Y2K and AI are nevertheless fundamentally different.

Y2K had a known date.

Known failure mechanisms.

Known remediation methods.

And nearly everyone shared the same incentive:

Fix it.

AI has no January 1, 2000.

There is no agreed moment of failure.

No universally accepted definition of dangerous capability.

No solved alignment problem.

And perhaps most importantly, the United States, China and enormous technology companies have powerful incentives to keep advancing.

That's the part of the Y2K analogy that should make us more thoughtful, not less.

😂 Part V — Laughing Through the Existential Dread

Fortunately, humans retain one capability AI hasn't eliminated:

Making jokes about our impending replacement.

🤖 The Alignment Problem

Scientists: “We must ensure AI shares human values.”

AI studies human history, social media and reality television.

AI:

“Please don't insult me.”

💼 The Good News

CEO:

“AI isn't going to replace you.”

Employees breathe sigh of relief.

CEO:

“Kevin is going to replace you.”

Employees:

“Kevin?”

CEO:

“Kevin knows how to use AI.”

🧹 The Roomba Revolution

People worry about machines taking over civilization.

My robotic vacuum just spent 45 minutes being held hostage by a laundry basket.

Humanity: 1.
Superintelligence: pending.

🚦 The Final Defense

If AI ever becomes uncontrollable, forget missiles.

Put this in front of the data center:

SELECT ALL SQUARES CONTAINING A TRAFFIC LIGHT.

Civilization saved.

💡 Four Conclusions That Matter

1. The biggest AI mistake is forcing the debate into “doom” versus “no doom.” Immediate harms can be serious even if extinction risk turns out to be tiny.

2. A kill switch is worth having precisely because it isn't sufficient. Safety-critical systems need overlapping defenses; civilization should not have a single point of failure.

3. Y2K doesn't prove technological warnings are exaggerated. It demonstrates how successful prevention can make legitimate risks look silly afterward.

4. The deepest long-term danger may be gradual delegation rather than robot rebellion. Humans don't have to lose a war against AI if we voluntarily hand algorithms control over finance, infrastructure, information and decision-making one convenient task at a time.

That fourth one is, to me, the sleeper insight of the entire present piece.

📌 Signal Extract

“The greatest paradox of technological prevention is that success eventually becomes evidence that the danger never existed. Sometimes ‘nothing happened’ is not proof of hysteria. It is proof that somebody did the work.” — FUNanc1al

🎯 High-Conviction Takeaway

“The most important AI question isn't whether machines become conscious. It's whether humans delegate enough authority that consciousness no longer matters.” — FUNanc1al

⚡ Quick Take / TL;DR

AI apocalypse predictions could prove wildly overstated.

But that's not an argument for ignoring AI safety.

The rational position sits between “Terminator arrives Tuesday” and “there is absolutely nothing to worry about.”

Address today's tangible risks aggressively. Build layered safeguards against tomorrow's uncertain ones. Keep consequential decisions under meaningful human control.

And if civilization survives?

Future generations may laugh at all the fuss.

That would be a spectacular outcome.

🍽️ Food for Thought: The Cross-Hub Connection

This story touches almost every FUNanc1al universe.

Investing: AI creates extraordinary productivity opportunities alongside concentration, regulatory and infrastructure risks.

Business: The competitive advantage may increasingly belong not to companies that automate the most, but to those that determine what should never be automated without human review.

Health: AI could accelerate drug discovery and diagnostics while introducing new questions about medical accountability.

Politics & society: Deepfakes make authentication and institutional trust increasingly valuable.

Carpe Diem: We are living through a technological transformation whose final consequences nobody understands.

That last part isn't necessarily terrifying.

It's also fascinating.

❓ FAQ

Is AI likely to cause human extinction?
Nobody knows. Some prominent researchers assign meaningful probabilities to catastrophic loss-of-control scenarios; other experts consider such forecasts highly speculative. There is no scientific consensus establishing a reliable probability.

Would an AI kill switch solve the problem?
It could be an important safeguard, but not a complete solution. The proposed U.S. legislation focuses on maintaining the technical ability to slow, suspend or shut down covered frontier systems.

Was Y2K just hype?
No. The better interpretation is that a known technical vulnerability prompted extensive remediation before the deadline. Its peaceful outcome doesn't establish what would have happened without that work.

Are today's AI risks already real?
Yes. Cybersecurity, fraud, synthetic media, privacy, bias and labor-market disruption don't require hypothetical superintelligence.

Should we slow AI development?
That's now intensely contested. Some AI leaders and safety researchers favor slowing frontier development and strengthening independent evaluation; others argue that excessive restrictions could sacrifice economic, scientific and geopolitical benefits. The political split is particularly visible right now.

📊 FunTech Index™ — 9.75 / 10

Why so high? Because few technologies combine comparable upside, economic disruption, geopolitical importance and potentially catastrophic tail risk.

The score does not mean “97.5% probability AI kills us.”

Quite the opposite.

It means 9.75/10 importance for understanding, monitoring and managing the technology.

AI could become one of humanity's greatest inventions.

Which is exactly why getting the guardrails right matters so much.

🤖 Final Thought

Maybe AI doom turns out to be Y2K 2.0.

Maybe twenty years from now we'll laugh at predictions that superintelligence was about to extinguish humanity.

I sincerely hope we do.

Because the funny thing about prevention is that looking foolish afterward can be the ultimate measure of success.

Build the guardrails.

Keep humans in the loop.

Demand evidence from optimists and doomers.

Enjoy the astonishing benefits.

And keep an eye on the Roomba.

Carpe Diem. 🤖⚡😂

👤 About the Author

Frédéric Marsanne is the founder of FUNanc1al—part market analyst, part storyteller, part accidental comedian. A longtime investor, entrepreneur, and venture-builder across tech, biotech, and fintech, he now blends rigorous financial analysis with a twist of humor to help readers laugh, learn, live healthier lives, and invest a little wiser.

His research focuses on insider buying, hedge funds, valuation, behavioral finance, long-term wealth creation, and the fascinating intersections between business, science, technology, health, passions, and everyday life.

When not decoding SEC filings or poking fun at earnings calls, he's building Cl1Q, writing fiction, painting, creating videos, or discovering new passions to FUNalize.

📝 Editorial Note

This analysis reflects information available through September 14, 2026 and is intended as a snapshot of the investment thesis at publication. FUNanc1al emphasizes primary-source financial analysis where practicable and distinguishes reported facts from our own interpretation and investment thesis.

Every FUNanc1al article is grounded in human research, analysis, and editorial judgment. Modern AI tools may assist with research organization, editing, and presentation, but every opinion, conclusion, rating, and editorial judgment remains subject to human oversight and responsibility.

To learn more about how we research, write, and review every article, please visit our Editorial Process page.

FUNanc1al. Laugh a little. Learn a lot. Invest wiser.™