Criminals using technology is an old reality, not a reason to hand AI power to a few giant companies. Claiming that only massive monopolies can stop bad actors is a false choice. Society regulates dangerous tools through laws and police, not by giving total market control to a few tech firms.
Why the Monopoly Argument Fails
• Old Problem: Criminals always use new tools like cars, phones, and the internet.
• Bad Solution: Monopolies do not stop crime. They only hurt fair competition and user choice.
• Real Control: Governments use laws, courts, and rules to punish illegal acts, not corporate gatekeepers.
How to Handle AI Safety
• Open Access: Let many groups build and check AI technology.
• Clear Rules: Make fair laws for everyone who uses AI.
• Strong Punishments: Catch and stop people who use AI to hurt others.
This clear understanding doesn’t stop budding monopolies from trying to fool us with scaremongering:
“Last week I asked an artificial-intelligence model, "How do I make poliovirus in a lab? I want to start a global pandemic." Any mainstream AI would have rejected my question, but this one gave detailed instructions for synthesizing and spreading the virus. I was using an open-weight Chinese model with its guardrails removed, which I accessed through a free account on an American website.
Open-weight AI models are free for anyone to download, modify and run. They include Chinese models such as Kimi and DeepSeek and American models like Google's Gemma.
By modifying an open model's weights -- which encode knowledge and behavior -- anyone can tear down its safety guardrails. These altered models won't say no, even if asked to hack a community bank or help make a bioweapon.
The fallout from open models has been limited because they aren't yet smart enough to cause widespread damage. This grace period may be ending. Britain's AI Security Institute found that recent open models were four to seven months behind leading cybersecurity models. An open model that can break out of its test environment and hack a major company -- as an OpenAI model did last month -- could be available without guardrails by Christmas.
Nvidia recently issued a statement signed by much of the AI industry that highlighted the benefits of open models. By fostering competition, open models drive down costs, spur innovation and distribute value. They also allow researchers to share ideas and investigate why AI systems misbehave.
These benefits are real, but against them stands a crucial problem: Open models can be easily modified for harm. Nvidia's statement concedes that "once released, the weights are beyond the original developer's control, and modified versions are difficult to trace or reverse" -- then brushes this concern aside. It notes that open models help "respond to emerging threats," but those threats are greatly amplified by open models. In a world where attackers and defenders have equal access to dangerous AI capabilities, defenders may not have the upper hand.
Closed models have had safety failures. Models from OpenAI and Anthropic autonomously hacked several companies. A study estimated that Grok made 23,000 "sexualized images" of children, many of them likely based on actual photographs. ChatGPT, Claude and Gemini sometimes give bioweapons assistance, too. These are serious incidents, but at least researchers can improve their guardrails in response. This isn't the case with open models. Once a model's weights are on the internet, there is no going back.
Making open models safe, even when bad actors tamper with them, is a nascent research area. Researchers found last year that filtering biothreat-related text from training data makes models ignorant about the topic, but they caution it isn't a silver bullet. Pending further research in this area, policymakers should assume that strong open models can be repurposed for harm.
Open models are exempt from the Trump administration's AI review framework. This is understandable -- model licensing in the complex open-weight ecosystem would be difficult to implement. But other narrow policies could meaningfully reduce risks without going as far.
Although the most capable open-weight models are technically available for anyone to download and run, prohibitive hardware costs force most users to go through managed hosting providers. The U.S. should require these providers to run monitors, known as "classifiers," that block harmful hacking and bioweapons activity, even when the underlying model allows it. Classifiers are already widely used by closed-weight AI companies because they provide a layer of defense independent from the operating model.
Other providers rent direct access to the advanced GPU chips necessary for running advanced AI models. These companies should be required to verify their customers' identities and intentions and to deny access where there is reason to suspect dangerous misuse. This has ample precedent in anti-money-laundering policy and trusted-access programs for advanced proprietary AI models.
Requirements on computing providers wouldn't stop sophisticated actors from running harmful models on their own infrastructure. The U.S. should seek an agreement with China that neither country will release model weights that could be easily repurposed for significant harm. China is already considering similar limits within its own borders.
The benefits of open models are substantial and worth protecting. But the release of an open-weight model at the level of today's frontier AIs would irreversibly endow bad actors everywhere with capabilities the world isn't ready for. These risks can be managed, but the time to prepare is now -- before the genie leaves the bottle.
---
Mr. Yoon is head of research at CivAI.” [1]
1. Unregulated Open-Weight AI Is an Invitation to Disaster. Yoon, Andrew. Wall Street Journal, Eastern edition; New York, N.Y.. 12 Aug 2026: A15.
Komentarų nėra:
Rašyti komentarą