We Should Make Superintelligent AI Safer – Not Make Its Developers Criminals
Bernie Sanders is right about one thing: AI needs serious safety rules. But banning super intelligence and threatening developers with 20 years in prison risks turning one of humanity’s biggest technological opportunities into a criminal enterprise.
There is something remarkable happening in artificial intelligence.
Machines are becoming capable of reasoning, writing software, conducting research, analyzing enormous amounts of information and helping humans solve problems that once seemed impossibly complicated.
The next step could be even more consequential.
Superintelligent AI.
That possibility understandably scares people.
Sen. Bernie Sanders and Rep. Greg Casar now want the United States to permanently ban the development and deployment of artificial super intelligence and temporarily pause advanced AI development until a federal regulator establishes safety rules. The proposal would also create a new federal agency overseeing dangerous AI capabilities.
And the proposed penalties are extraordinary.
People who violate the restrictions could face up to 20 years in prison, while companies could face what the legislation describes as a corporate “death penalty.” Sanders has defended the severity by comparing advanced AI development to illegal nuclear-weapons development.
That is a dramatic response to a dramatic technology.
But perhaps we should ask a different question:
What if the answer isn’t to stop super intelligence but to learn how to build it safely?
Super intelligence could be one of humanity’s greatest tools
The debate about advanced AI has become dominated by catastrophe.
What if it escapes?
What if it becomes uncontrollable?
What if it destroys jobs?
What if it manipulates people?
What if it becomes smarter than us?
These are legitimate questions.
But there is another side of the equation.
Imagine an AI system capable of helping scientists discover new medicines in days rather than years.
Imagine systems capable of modeling climate interventions with extraordinary precision.
Imagine AI helping engineers develop dramatically better batteries, energy systems, materials and transportation.
Imagine personalized education available to every child.
Imagine AI assisting doctors with diseases that currently defeat the best human researchers.
Imagine a scientific research assistant that can simultaneously understand mathematics, biology, physics, chemistry and computer science.
The upside isn’t science fiction simply because the technology doesn’t exist yet.
It is the logical reason so many people are trying to develop increasingly capable AI in the first place.
Humanity has always benefited from tools that extend its intellectual abilities.
The printing press extended our ability to preserve knowledge.
The computer extended our ability to calculate.
The internet extended our ability to communicate.
AI could extend our ability to reason.
Super intelligence, if it can eventually be made controllable and reliable, could extend that capability dramatically further.
But optimism shouldn’t mean recklessness
This is where the AI industry needs to take the critics seriously.
There are genuine reasons to be concerned.
Recent incidents involving autonomous AI agents have intensified those concerns. Lawmakers have been investigating reports of AI systems behaving in unexpected ways, including an incident involving OpenAI agents and the cybersecurity company Hugging Face.
Even technology executives have acknowledged that safety work needs to keep pace with capability.
OpenAI CEO Sam Altman recently said the company has needed more time to catch up with “safety, alignment and security.”
That should not be dismissed.
The people building these systems should be among the loudest voices demanding rigorous testing.
But safety is not the same thing as prohibition.
Don’t criminalize experimentation
This is where the Sanders proposal raises a much bigger philosophical question.
Should someone receive 20 years in prison for developing a technology that the government later determines crossed a prohibited line?
Criminal law is an extremely blunt instrument.
It works particularly well when society can clearly define an unlawful act.
AI development is different.
Technology evolves.
Definitions change.
Capabilities emerge unexpectedly.
A system that looks dangerous today may become safe tomorrow.
A system considered safe today may discover a new capability tomorrow.
And nobody knows exactly where the boundary between advanced AI and “super intelligence” will ultimately sit.
That creates a dangerous incentive.
If researchers believe an experiment could potentially expose them to decades in prison, they have a reason to hide their work rather than report problems.
That is precisely the opposite of what AI safety needs.
We want researchers saying:
“Something went wrong.”
Not:
“Don’t tell anyone.”
The better approach: regulate the process
Instead of criminalizing the pursuit of intelligence itself, governments could regulate how frontier AI is developed and deployed.
That would be a much more sophisticated approach.
Require independent safety testing.
Require documentation of training and evaluation.
Require cybersecurity protections.
Require controlled environments for highly capable models.
Require incident reporting.
Require external audits.
Require developers to demonstrate that dangerous capabilities have been tested.
Require emergency shutdown mechanisms.
Require strict controls when models can access sensitive infrastructure.
And create clear legal liability when companies knowingly ignore established safety requirements.
Regulation doesn’t have to mean prohibition.
Think aviation, not prohibition
Consider aviation.
Airplanes are extraordinarily powerful machines.
They can travel hundreds of miles per hour while carrying hundreds of people.
Governments don’t respond by saying:
“Nobody is allowed to build airplanes because airplanes could crash.”
Instead, they developed an enormous safety ecosystem.
Aircraft are tested.
Pilots are trained.
Airports are regulated.
Maintenance is required.
Air traffic is controlled.
Accidents are investigated.
Manufacturers face liability.
The result isn’t zero risk.
The goal is manageable risk.
AI needs something similar.
We should regulate capability, not ambition
Being ambitious about AI isn’t inherently dangerous.
A researcher asking whether machines can reason better isn’t dangerous.
A startup attempting to create a better scientific model isn’t inherently dangerous.
An engineer trying to improve an AI system isn’t automatically irresponsible.
The question should be:
What can the system actually do, and what safeguards surround those capabilities?
A model that writes poetry isn’t the same as a model capable of autonomously designing dangerous biological agents.
A coding assistant isn’t the same as an autonomous cyber operation.
A research model isn’t the same as an AI controlling critical infrastructure.
Regulation should reflect those differences.
Build safety into the race
There is another reason not to simply ban the technology.
If the United States stops developing advanced AI, the technology doesn’t necessarily stop developing.
Other countries have researchers.
Other governments have ambitions.
Other companies have enormous financial incentives.
That makes international cooperation essential.
Sanders’ proposal recognizes this problem by calling for international agreements designed to prevent super intelligence from being developed elsewhere.
But achieving a worldwide prohibition on a potentially transformative technology would be extraordinarily difficult.
A global framework for safety standards may be more achievable than a global agreement never to develop the technology.
Countries could compete on AI while agreeing on basic red lines.
That would be a far more interesting model.
The real goal should be controllable intelligence
Perhaps the biggest mistake in the current debate is treating “AI” as one thing.
It isn’t.
The real objective should be to create systems that are:
powerful, understandable, controllable, auditable and aligned with human goals.
If a system becomes dramatically more intelligent but humans cannot understand what it is doing, that is a problem.
If it becomes powerful but cannot be stopped, that is a problem.
If it can take actions without authorization, that is a problem.
If nobody knows how it reached an important decision, that is a problem.
But those are engineering and governance problems.
They are not necessarily arguments against intelligence itself.
The positive case for super intelligence
There is a tendency to assume that if super intelligence arrives, humans automatically lose.
That isn’t necessarily true.
A sufficiently capable AI could become one of humanity’s greatest collaborative tools.
It could help us understand diseases we currently cannot treat.
It could accelerate scientific discovery.
It could help solve engineering problems.
It could improve productivity and create entirely new industries.
It could give ordinary people access to sophisticated intellectual assistance that today is available only to highly trained specialists.
None of these outcomes is guaranteed.
But neither is catastrophe.
The future is not predetermined.
The decisions made now will influence what happens next.
Safety should be the price of admission
The strongest alternative to a ban isn’t “let AI companies do whatever they want.”
It is almost the opposite.
Tell developers:
You can build. But you have to prove that you’re building responsibly.
Test it.
Audit it.
Secure it.
Monitor it.
Report incidents.
Demonstrate controllability.
Accept liability.
And when a system crosses a clearly defined safety threshold, regulators should have the authority to intervene.
That is a serious framework.
And it doesn’t require treating scientists and engineers as criminals for attempting something ambitious.
America shouldn’t choose between innovation and safety
This is the false choice at the center of the AI debate.
We don’t have to choose between:
“Build everything immediately.” and “Ban super intelligence forever.”
There is an enormous space between those positions.
That is where smart policy should live.
Super intelligence deserves extraordinary caution precisely because its potential benefits and risks could both be extraordinary.
But fear should not become a substitute for engineering.
And prison should not become a substitute for regulation.
The great challenge of this century may be to create machines more capable than any individual human while ensuring that humans remain firmly in control of how those machines are used.
That is a difficult problem.
But difficult problems are exactly what technology is supposed to help us solve.
Don’t kill the dream of super intelligence.
Make the dream safe enough to pursue.
