A Guardian commentary argues that it should. Its case is that free Chinese open-weight models carry security and intent risks that Western alternatives at least try to address. The cost savings are real, so the trade-off is a matter of debate.
Key takeaways
- A commentary published by the Guardian argues that British businesses and government should not adopt AI models developed in China.
- Chinese developers such as DeepSeek and Xiaomi regularly release “open-weight” models that organisations can download and run without paying an ongoing licence fee.
- The central argument is that Western AI comes with at least some effort towards security and benign intent, and that Chinese AI offers no such assurance.
- Open-weight models can be run on an organisation’s own hardware, which changes the risk compared with sending data to a foreign-hosted service.
- The debate sets lower costs and wider access against concerns about trust, hidden behaviour and national security.
What is happening with Chinese AI models in the West?
The Guardian reports that Chinese AI developers, including DeepSeek and Xiaomi, put out new versions of their models about once a month. These are “open-weight” models. The core trained software, meaning the numerical parameters that decide how the model behaves, is published so that anyone can download it, run it and adapt it.
For businesses and public bodies in Britain and elsewhere in the West, this means a capable AI system can be had almost free. The alternative is usually a paid licence or a per-use subscription from a commercial provider. According to the Guardian piece, the low price is exactly what makes these models tempting, and also what makes them a cause for concern.
The commentary argues that there is much to fear from AI built in China and handed to Western organisations at little or no cost. It says Western technology comes with at least some attempt to provide security and benign intent, and that Chinese technology does not.
Why is this in the news now?
The article is a comment piece published in the Guardian’s business section. It is an argument, not a news report. It comes as Chinese labs keep releasing open-weight models at a fast pace, so organisations looking to cut the cost of AI are repeatedly faced with the choice.
The source summary does not name any particular incident, policy announcement or government decision behind the piece. It is not known from the available material whether the article responds to a specific British procurement decision, a security finding or a new model release. Readers should treat it as part of a continuing policy debate, not as a reaction to one event.
What does a newcomer need to know about open-weight AI?
Large language models and similar AI systems are built by training software on very large amounts of data. Training produces a set of “weights”, the numbers that encode what the model has learned. Whoever holds the weights can run the model.
Developers share their models in broadly three ways:
- Closed models are reached only through the developer’s own service. The user sends a request and receives an answer, but never sees the weights.
- Open-weight models publish the weights, so users can run the model on their own computers or cloud servers, often with few restrictions.
- Fully open-source models go further and also publish training code, and sometimes the training data, so outsiders can study how the model was made.
Open-weight release lowers costs and lets organisations keep sensitive data on their own systems, because nothing has to be sent to the developer. It also lets organisations fine-tune a model for specialised tasks.
The weakness is that the weights are hard to inspect. A model’s behaviour comes from billions of learned parameters, not readable instructions. That makes it hard to be sure a model contains no biases, hidden triggers or deliberate weaknesses. Critics of Chinese models focus on this opacity.
Who is affected and how?
Several groups have a stake in the argument.
British businesses that want to use AI in products or internal operations could save money by building on free open-weight models instead of paying licence fees. Smaller firms and start-ups that cannot afford large subscriptions may find the saving especially attractive.
Government departments and public bodies handle sensitive information and provide essential services. Concerns about trust and security are sharper for them, because a compromised or manipulated system could affect citizens directly.
Developers and researchers often use open-weight models to experiment, compare performance and build new tools. Limits on which models they can use would change how they work.
Western AI companies that sell licensed models compete with free alternatives. The argument for avoiding Chinese models would, in effect, favour their products, which is worth keeping in mind when weighing the debate.
Chinese AI developers depend on worldwide adoption to build influence, gather user feedback and establish their technology as a standard. Wide avoidance in Britain or elsewhere would limit that reach.
Where do informed people disagree?
The Guardian commentary takes a clear position, but the wider debate has several sides.
Those who share the commentary’s concerns argue that a model’s origin matters. Their points include that a model trained under a different legal and political system may reflect that system’s priorities, and may give skewed or censored answers on some topics. They also argue that the opacity of weights makes it hard to rule out deliberate tampering, and that heavy dependence on foreign technology could become a strategic weakness. On this view, a low price may be a way of building dependence.
Those who take a more permissive view make other points. Running an open-weight model on an organisation’s own infrastructure means no data needs to flow back to the developer, which removes one of the main surveillance concerns. Open weights can be tested, benchmarked and examined by independent researchers in ways closed models cannot. Some also argue that the security assurances offered by Western developers are imperfect and hard to check from outside, so the line between “trusted” and “untrusted” sources is less sharp than it looks.
There is also an economic argument. Restricting access to cheap, capable models could raise costs for British organisations and slow their use of AI compared with competitors in countries that impose no such limits. Supporters of restrictions reply that the hidden costs of a security failure could be far larger than any saving.
Which view is right depends partly on the technical facts of specific models. Those facts are not set out in the source material and are best judged case by case.
What are the practical implications?
For organisations considering AI adoption, the debate points to several practical questions, wherever a model comes from:
- Where does the model run? A model hosted on the organisation’s own systems carries different risks from one reached through an external service.
- What data does it touch? Using a model for low-stakes drafting differs greatly from using it to process personal, financial or security-sensitive information.
- How has it been tested? Independent evaluation for bias, accuracy and unexpected behaviour can lower uncertainty, though it cannot remove it.
- What happens if it must be replaced? Building core systems around one model, from any supplier, creates dependence that may be expensive to undo.
- Are there procurement rules? Public bodies especially may face guidance or restrictions on technology from particular countries. The current state of any such British rules for AI models is not set out in the source material.
For policymakers, the question is whether to leave such choices to individual organisations, issue guidance, or set formal restrictions. Each approach balances security, cost and openness differently.
What should readers watch next?
Several developments will shape how the debate plays out. The first is the pace and capability of new Chinese open-weight releases, which the Guardian says arrive roughly monthly. If they keep closing the gap with paid Western models, the pull of a low price will grow.
The second is any British government response. That could take the form of procurement guidance, security advice for businesses, or broader rules on AI supply chains. No specific policy is mentioned in the source.
The third is independent research into how open-weight models behave, including whether hidden behaviours can be reliably detected. Better tools for inspecting models could ease some concerns or confirm them.
Finally, it is worth watching how Western developers respond. They may release more open-weight models of their own, cut prices, or stress security assurances as a selling point against free alternatives.
Frequently asked questions
What is an open-weight AI model?
An open-weight AI model is one whose trained parameters, called weights, are published for anyone to download. Users can run the model on their own computers or servers and adapt it to their needs without paying an ongoing licence fee. This differs from closed models, which can only be reached through the developer’s own service. Open-weight release does not always mean the training data or code are also published.
Why are Chinese AI models so cheap?
Many Chinese developers, including DeepSeek and Xiaomi, release their models as open weights, which means organisations can use them without paying ongoing licence fees. The Guardian reports that new versions come out roughly monthly. Developers’ exact commercial motives are not set out in the source material, but free release generally helps a technology spread widely and gain influence among users.
Is it safe to use DeepSeek in a business?
There is no single answer, as safety depends on how and where the model is used. Running an open-weight model on an organisation’s own systems avoids sending data to the developer, but it does not remove concerns about bias or hidden behaviour built into the weights. Organisations handling sensitive data should test any model carefully and check relevant security guidance before deploying it.
Can AI model weights contain hidden backdoors?
Researchers have raised the concern that AI models could be trained to behave differently when given particular triggers, and that such behaviour would be hard to spot. Weights are vast collections of numbers, not readable code, so inspecting them directly is difficult. That opacity is a central reason critics urge caution about models from developers they do not trust, whatever the country of origin.
Why does a Guardian article say Britain should reject Chinese AI?
The commentary argues that Western AI technology comes with at least some effort towards security and benign intent, and that AI from China offers no such assurance. It warns that the low cost of Chinese open-weight models makes them tempting for businesses and governments, but says the risks outweigh the savings. It is an opinion piece, not a statement of government policy.
Are Western AI models more secure than Chinese ones?
That is disputed. The Guardian commentary argues that Western developers at least try to provide security and benign intent. Others point out that security claims from any developer are hard for outsiders to verify, and that open-weight models from any source can be independently tested. In practice, security depends on the specific model, how it is deployed and what data it handles.
Sources and further reading
- The Guardian: business section commentary arguing that Britain should avoid cheap Chinese AI models.
- Published developer documentation for open-weight AI models, which explains licensing terms and how weights are released.
- UK government cyber security guidance for organisations adopting AI and other third-party technology.
- Academic and independent research on AI model evaluation, transparency and the detection of hidden model behaviour.
Surfaced from the rss:guardian_tech signal “Chinese AI adoption debate”. AI-assisted draft, editorially reviewed.
