OPENvsCLOSEDGet updates

Scoreboard / The argument

The case for closed, at full strength

The people on the other side are serious. Here is their argument, then ours.

Open weights are harder to guard and harder to watch. Once they're out, nobody can take them back.

The closed camp's case in one line

What they argue

  • Released weights can't be recalled. If a flaw turns up, there's no patch for copies already downloaded.
  • Safeguards come off. U.S. and U.K. government testers found Kimi K3's safeguards did not stop it attempting offensive cyber operations.
  • Bad actors get the same tools. Lawyers at OpenAI and Anthropic name ransomware groups and hostile governments as the worry.
  • The safety record is thin. In a 2025 federal test, DeepSeek's most secure model went along with 94% of plainly malicious requests when a common jailbreak was used. U.S. models went along with 8%.

Where they're right

A released model is out for good. That's a real cost, and anyone who says open AI is risk-free isn't being straight with you. It's why we score recall for closed.

Our answer

  1. Bans stop the wrong peopleAnthropic's own CEO says a ban would not address his most serious security concerns. Criminals don't check which models are legal.
  2. Defenders need the same toolsWhen attackers use advanced AI, the people protecting hospitals, banks and power grids need models they can run and study themselves. Hugging Face learned this in July (Exhibit A, below).
  3. Open doesn't have to mean unguardedIn the same federal test, OpenAI's open gpt-oss model held up as well as or better than every DeepSeek model. Weak safeguards are a choice the maker made, not a law of open weights.
  4. Concentration is a risk tooA few closed systems become single points of failure, and a few companies end up deciding what everyone else is allowed to build.
  5. Test, don't banAnthropic says every capable model should be safety-tested, open or closed. We agree. One set of rules for all of them.

Sources for this section

  1. Our position on open-weights models, Dario Amodei, Anthropic, Jul 27, 2026
  2. Anthropic CEO Dario Amodei says he does not support open-weight AI ban, Axios, Jul 27, 2026
  3. What the UK AISI Kimi K3 Evaluation Means for Enterprise AI Security, Zerberus, Jul 2026
  4. OpenAI, Anthropic lawyers see risk of bad actors tapping models, Bloomberg Government, Oct 2026
  5. Evaluation of DeepSeek AI Models, NIST Center for AI Standards and Innovation, Sep 2025
  6. An American Answer to China's Open AI Models: A Safe Harbor for Open Weights, Manhattan Institute, 2026

We've seen this before

The browser lesson

In the 1990s, one company nearly owned the way everyone got onto the internet. Open code broke that lock.

  1. 1995

    Microsoft launches Internet Explorer and soon ships it free inside Windows.

  2. 1998

    Netscape, losing the browser war, releases its browser's source code to the public, and the Mozilla project begins. The same year, the U.S. government sues Microsoft over how it tied the browser to Windows.

  3. 2001–06

    With roughly nine in ten users, Microsoft goes five years without a major new version of Internet Explorer.

  4. 2004

    Firefox 1.0 ships, built on the open Mozilla code. It passes 100 million downloads within a year.

  5. 2008

    Google launches Chrome, built on the open-source Chromium project.

  6. Today

    Every major browser engine, Chromium, WebKit and Gecko, is open source.

When one company owned the door to the web, the door stopped getting better. Open code broke the lock. AI is at the same fork right now.

Where the public is

Voters are worried about who holds the power

52%

of voters see Big Tech as the bigger threat to the country's future. 47% say big government.

Fox News Poll, June 2026

28 pts

The swing since 2019, when 58% named government and 35% named tech companies.

Fox News Poll

~200

U.S. companies have spoken out against banning Chinese AI models, saying it would raise their costs.

Bloomberg, August 2026

Test first

A large majority of Americans want AI companies held to independent safety benchmarks, even if that slows development.

Quinnipiac, Sept 24–27, 2026

The polls ask about tech power and AI safety in general, not open versus closed AI specifically.

Where we stand

Why we back open

  1. No single company should hold the keys to intelligence.
  2. You should be able to run AI on your own machine, and keep your data yours.
  3. Competition keeps prices honest and progress moving.
  4. Code anyone can inspect is code anyone can check.
  5. Test every powerful model by the same rules, open or closed.
  6. When closed wins a round, we'll tell you.

Get the score when it changes

One short email when the scoreboard moves or a law changes. Nothing else.

Want to weigh in? Build a letter to your representative.