by Justin TRUGMAN, Cofounder & Head of Technology, BetterFutureLabs
Presented at the LMGlobal.Org‘s panel on “Towards a Public Discussion for Shaping the Future of A.I. Regulations in the U.S. & Beyond“, which virtually presented in collaboration with Bons-Ai.Org on November 15, 2025.

Presentation (PDF)

Introduction by Adam Bricker:
[LMGlobal.Org] I met Justin about a year and a half ago. He gave me some personal advice on AI and how to get started, and that advice has been immensely helpful to me. I hope you’ll also look at his bio, see the startups he’s built, and the accomplishments he’s delivered.
I’m really grateful for the chance to listen and learn from him. He’s going to share his perspective on the state of AI regulation in the United States. Justin, over to you, sir.

Justin Trugman “Thanks, Adam, and thanks, Bircan and Victor, for the opportunity. Let me share my screen and we’ll get started.

Okay, you should be able to see my screen.

Today I’ll talk about balancing AI regulation and innovation.

A bit about me: I’m sharing these thoughts from the trenches of AI research. I’m the co-founder of Better Future Labs, a venture studio working specifically on multi-agent AI. That’s our niche inside the broader AI landscape.

We’ve been doing a lot of cutting-edge research, and I’ll mention that briefly. But I want to focus mainly on why it’s important to regulate AI in the right way, and what lessons the U.S. and other countries can learn from each other.

Why this matters now

AI capabilities are advancing very rapidly. You already heard from earlier speakers how fast this is moving. I think Charles mentioned that his teams change their assessments every two weeks. That is the pace we are dealing with.

Because of that, first-mover advantage is critical. The first countries and regions that adopt and release models under clear rules will shape global standards.

The window for thoughtful policy is closing. If regulation comes too late, governments often regulate in hindsight. That leads to rushed, poorly designed laws that can hurt industry instead of protecting people.

There are also misaligned incentives. Bad regulation can stifle innovation, as we are already seeing to some extent in Europe with the EU AI Act and its impact on companies like Mistral.

So we have to protect people and still allow innovation. It is a very delicate balance.

Three global approaches

We can roughly see three different approaches around the world:

1. European Union – EU AI Act

Very strict, with four risk tiers.

Strong penalties for violations.

Phased enforcement running through 2027, to give time to adapt because the rules are so strict.

2. China – Multi-layered framework

Content and labeling rules. Data protection rules.

Algorithm filing requirements – companies must file their algorithms with the government; more than 1,400 have already been filed.

Specific deepfake regulations as well.

3. United States – “Evolving”

Initially regulated mainly by Executive Orders, not by laws passed through Congress.

The early approach focused on compute capacity and the amount of “math” used to train models (floating-point operations).

That approach was unpopular in the research community and was later repealed.

We now have the AI Action Plan, which I’ll touch on.

U.S. federal level: two recent approaches

Under the previous administration, an Executive Order required safety testing for models above 10²⁶ floating-point operations used in training.

For those models, it required:

Red-teaming (adversarial testing).

Stronger cybersecurity.

A focus on safety, equity, and civil rights.

The controversial piece was that 10²⁶ number. No one really knew why that threshold was chosen. It felt arbitrary to many researchers.

The newer approach removed that compute threshold and deregulated some of those oversight measures. It shifted toward competitiveness and innovation, but in doing so, it also dropped some of the useful safety requirements from the first Executive Order.

In terms of actual laws passed by Congress and signed by the President, there is really just one major federal AI law at the moment:

The Take It Down Act, which regulates the use of AI to create deepfake intimate images.

It imposes severe penalties on people who use AI to create and share that kind of non-consensual content.

U.S. state laws: a growing patchwork

At the state level, the picture is very uneven.

A small set of states (shown in red on the map) have enacted AI laws.

Some states have inactive or stalled bills (darker gray).

Many states have no comprehensive AI bills at all (lighter gray).

So it really is a kind of “Wild West” across the states.

The good news is that, because there are not that many state laws yet, it is still possible to read and analyze them all. When you do, you see some common pillars:

I. Risk-based governance

Policies and risk assessments.

Training requirements.

Clear assignment of accountability – who is responsible when something goes wrong?

II. Transparency

Notice to users and buyers.

Labeling requirements.

Explanations after incidents or high-risk uses.

III. Assurances

Third-party reviews for high-risk cases.

Examples of high-risk cases: healthcare, self-driving vehicles, and similar critical areas.

IV. User protections

Right to opt out of certain AI uses.

Right to appeal AI-driven decisions.

Duties around anti-discrimination in AI systems.

There are also important differences between states:

Some laws apply only to frontier models – the largest, most advanced models (large language models, multimodal models, etc.).

Some focus on labeling AI-generated content.

Some are broad consumer protection laws. For example, Texas passed a broad AI-related consumer protection law.

States also differ in how they assign duties:

Colorado splits responsibilities between developers and deployers.

Other states lean more heavily on developer transparency, like California.

They also differ in the depth of obligations:

How often is external review required?

How strong are the appeal mechanisms for affected users?

Lessons from the EU AI Act and Mistral

A key case is Mistral, the French AI company known for its mixture-of-experts models.

Mistral’s CEO has been strongly critical of the EU AI Act. His main argument:

You should not regulate the engine itself, you should regulate how it is used.

He compared regulating foundational models to banning the C programming language because it can be used to write malware.

C is a fundamental language used for many critical systems. We do not ban C. We regulate malware and harmful uses instead.

He called the EU AI Act “premature but workable”, and France itself has pushed back on some parts of the Act to protect its AI ecosystem.

The end result is that the EU has added some flexibilities for open-source models, but the tension remains.

Suggested approach from the research community

From much of the U.S. research community, one clear message is emerging:

Regulate the application of the technology, not the foundational technology itself.

A couple of important quotes:

Andrew Ng (co-founder of Google Brain, co-founder of Coursera, Stanford professor) said:

Regulating big AI models makes about as much sense as regulating high-horsepower motors.

Yann LeCun (Chief AI Scientist at Meta, former Bell Labs) added:

Regulating the deployment of products is fine, and in life-critical applications it is necessary.

But Meta is opposed to regulating AI research and development with arbitrary compute limits.

This criticism points directly back to the earlier U.S. Executive Order that tried to regulate models above 10²⁶ FLOPs.

From my vantage point in AI research, I agree with the core idea:

We do need AI regulation.

But it should focus on high-risk applications and deployments, not on basic research or on arbitrary total compute numbers.

Let researchers innovate on the underlying models and systems, and then put strong guardrails around how those models are used in products and services.

That way we can:

Protect society from harmful or abusive uses.

Avoid crushing innovation with rules that don’t match how the technology actually works.

That is my perspective from the trenches.

Thank you again for the time and the opportunity. If you’re interested in our research or products at Better Future Labs, there’s a QR code on the slide you can scan.

_ . _

Presentation (PDF)

On the A.I. Policy: Federal Approach and Current Landscape

Nations that invest in a human-centered infrastructure will be the true winners of the AI revolution.

Invitation: Towards a Public Discussion for Shaping the Future of A.I. Regulations in the U.S. & Beyond ~ on Nov. 15 (Virtual)