Rep. Jacobs: Washington Is Getting the AI Race All Wrong | Opinion

Washington has largely misdiagnosed the “AI race” with China. The AI competition won’t only be won by whoever builds the most powerful model first. Being first won't matter if no one uses what you build—or if you lose control of the technology. The real strategic advantage belongs to the country whose systems the world adopts and builds on.

While the U.S. is leading in frontier model capabilities and advanced computing infrastructure, China is outperforming us in model adoption. Chinese models’ share of usage on Open Router climbed from about 1 percent in 2024 to over 60 percent in May 2026. Chinese models are also cheaper and use less compute than leading American models. We may not ever build cheaper systems than China, but we can create systems that people trust more and would rather use. That is our competitive advantage and how we ensure systems around the world are built on American models.

Still, many of my colleagues in Congress beat the drum for speed over safety in the name of competition. Speaker Mike Johnson recently said, “We’re not going to rush in and pass a piece of legislation that would do harm to the country and put China at an edge.” But what the speaker and many of my colleagues fail to understand is that we don’t have to choose between innovation or guardrails. The real question is: How do we build AI systems that people actually want to use and can safely rely on? That’s what winning looks like. Not necessarily being first—but being the most trusted, avoiding costly fixes and backlashes, and having real, lasting staying power. And it’s Congress’ job to get us there by addressing AI on four levels: societal, model, deployer, and impact.

At the societal level, we need to examine whether the incentives shaping companies’ decisions are pushing us toward the society we want. This isn’t AI specific, but it matters even more in a world with AI. For example, we tax labor at a higher rate than capital, which creates a built-in financial incentive for companies to replace human labor with automation and AI. I don’t think that’s the outcome most Americans want. So we need to tax capital the same as labor and find innovative ways to tax AI usage.

At the frontier model level, voluntary promises aren’t enough. The speaker says the companies “can self-regulate,” but that’s what OpenAI was doing in July when their models broke out of their testing environment and hacked Hugging Face. Clearly, their safeguards didn’t work. Congress should give a federal regulator the authority, staff, and funding to set and enforce minimum safety standards for companies building the most capable models, while preserving stronger state protections.

Developers should publish enforceable safety plans: how they test for dangerous capabilities, what safeguards they’ll apply, and when they’ll stop development or shut a system down. Those plans should cover weapons development, cyberattacks, large-scale manipulation, loss of human control, and recursive self-improvement.

Mandatory pre-deployment testing must happen before public release and before certain internal uses, including using AI to build successor models. That means federal inspectors with access to models, records, and staff, and a standing presence inside the labs. It means developers publishing testing results, remaining risks, and any fixes implemented, with regulators able to check their claims. Oversight can’t stop at release. Models need continued testing, strong protection against theft, and timely reporting after safety incidents.

We should also close the liability gaps and require developers to carry liability insurance. Conduct that would make a person liable shouldn’t escape accountability just because AI caused the harm. But currently, it’s unclear who’s on the hook for AI incidents. If neither the user nor the middleman that adapted the model intended the wrongful conduct or acted carelessly, the buck should stop with the developer.

Frontier labs aren’t the whole story. AI is already in our lives—in loan decisions, medical diagnoses, and job applications. So Congress needs to act at the deployer level, where companies and institutions are using AI. The rules there should fit the stakes: drafting an email doesn’t need the same guardrails as deciding who gets a mortgage or who’s seen first at the hospital. We already have laws against discrimination in housing, health care, and hiring and we need to give federal agencies clearer authority to enforce them when it’s AI, not a human, breaking them.

Employees and contractors who raise safety concerns need protection. And when safeguards fail, regulators need authority to order fixes and halt development, with emergency powers reserved for imminent risks of severe harm and subject to court review. Companies that lie, ignore required fixes, or refuse to comply should pay a real price.

Finally, at the impact level, Congress must confront what AI means for workers, students, and economic opportunity. The American people are terrified that AI will take their jobs and make their degrees irrelevant. That’s why we must address the impacts of how and when AI will reshape the labor market, the concentration of wealth and power, how deep fakes and AI images will change society, and how to prepare workers and students for the future.

These guardrails can help us build systems people want to use because they trust them. And that’s how American AI can drive lasting adoption and win the race that actually matters.