Picture the scene. Jensen Huang is on stage at a panel discussion, mid-thought, wearing the leather jacket he wears everywhere. His phone rings. It’s the President of the United States. Instead of stepping offstage like a normal person, Huang puts Donald Trump on speaker in front of a live audience, and the two of them proceed to tell everyone in the room that the dangers of artificial intelligence are a hoax.
That happened. In September 2026, on a Monday, in public.
I review AI tools for a living. I spend my days watching agents fail at multi-step tasks, hallucinate function calls, and confidently tell users the wrong thing in polished prose. So when the man who sells the chips and the man who signs the executive orders agree that concern about this technology is a “hoax,” my first reaction isn’t political. It’s professional confusion. Which AI are they using?
The word “hoax” is doing a lot of work here
A hoax is a deliberate fabrication. Calling AI risk a hoax is a claim that people worried about this stuff are either lying or being lied to. That’s a strong position, and it flattens a genuinely messy set of concerns into one dismissible lump.
The concerns aren’t one lump. They’re at least three, and they don’t have much to do with each other:
- The science-fiction tier — superintelligence, loss of control, the stuff that fuels conference panels and Twitter threads.
- The boring-but-real tier — models that leak data, agents with overly broad permissions, automated decisions nobody can audit or appeal.
- The economic tier — what happens to the people whose jobs get partially automated, badly, by a tool that’s 80% good enough.
You can think tier one is overblown and still care intensely about tiers two and three. Most working engineers I talk to land exactly there. “Hoax” wipes out that distinction, which is convenient if you’d rather not discuss tiers two and three at all.
Follow the incentives, not the vibes
Huang’s position isn’t a mystery. Nvidia sells the hardware that every AI lab, hyperscaler, and startup needs. Faster development means more chips. Regulatory friction means slower deployment cycles and more time spent on compliance instead of capacity. This isn’t a scandal — it’s just how a business works. But it means his read on AI risk carries the same weight as an oil executive’s read on drilling policy. Informed, yes. Disinterested, no.
What makes this moment notable is that his position now has the loudest possible amplifier. Other tech leaders have called for regulation. Huang has gone the other direction, and he’s the one who got the phone call. That asymmetry matters more than anything either man actually said on that stage.
What this means if you actually build things
Here’s where I get practical, because this is a site about tools, not politics.
If the federal posture is “no meaningful guardrails,” the guardrails become your job. That’s not a hypothetical. It changes what you should demand from every vendor you evaluate:
- Audit logs you can actually read. Not a dashboard with a satisfaction score. Raw records of what the agent did, what it called, and what it received back.
- Permission scoping that defaults to nothing. Any agent framework that ships with broad filesystem or network access enabled by default is telling you what it thinks of your security posture.
- A documented failure mode. Vendors love talking about accuracy. Ask what happens when the model is wrong and nobody notices for three weeks.
- Data handling in writing. Where does your input go, how long does it live, and who else’s model gets to learn from it.
None of that is regulation. It’s procurement hygiene. But in an environment where the top of the government has decided the risks are imaginary, procurement hygiene is the only layer between your users and someone else’s optimism.
My honest read
I don’t think Huang is lying. I think he’s describing a version of AI risk — the robot-apocalypse version — that he finds silly, and he’s not wrong that it gets more airtime than it earns. The problem is that the dismissal scales up to cover everything, and the person handing him the microphone has the power to turn that dismissal into policy.
The tools I test aren’t dangerous because they’re too smart. They’re dangerous because they’re confidently mediocre and people wire them into systems that matter. That failure mode doesn’t require superintelligence. It requires a Tuesday and a bad integration.
Two men on a speakerphone won’t change what your agent does at 3 a.m. when the API returns something unexpected. Build accordingly.
🕒 Published: