OpenAI Models Should Compete, Not Be Banned, Arcee CTO Says

As Chinese open-weight artificial intelligence models continue to improve in performance and gain wider adoption, debate is intensifying over whether governments should restrict their use. While some policymakers have suggested limiting access to Chinese-developed AI models, others argue that competition and innovation, rather than bans, are the better path forward.

One of the strongest voices against restrictions comes from Lucas Atkins, Chief Technology Officer of AI startup Arcee. Although Arcee develops open AI models that compete with offerings from Chinese companies, Atkins believes enterprises should focus on evaluating models based on security and performance instead of where they were created.

Security concerns should be addressed through testing.

Chinese open-weight models such as MoonshotAI’s Kimi K3 and Alibaba’s Qwen have attracted attention for offering powerful AI capabilities at significantly lower inference costs than many proprietary models from companies including OpenAI and Anthropic. Their growing popularity has also prompted concerns that they could introduce cybersecurity risks for organisations deploying them.

Atkins argues those fears are often based on misconceptions about how large language models are built. According to him, an organisation running an open-weight model inside its own infrastructure does not automatically expose itself to the model’s developer or provide remote access to its systems.

He compared AI models to other forms of open-source software, noting that enterprises already inspect, test and validate software before deploying it into production environments. Open-weight AI models, while not fully open source because training data and methods are generally unavailable, still allow organisations to examine the underlying software that runs on their servers.

Before deployment, many enterprises conduct extensive evaluations covering areas such as security, hallucinations, bias, toxicity and performance. Organisations also fine-tune models for their own business requirements, giving them greater visibility into how those systems behave before employees begin using them.

Backdoor concerns remain largely theoretical.

Some critics have questioned whether AI coding assistants could secretly introduce malicious code into software projects. Atkins acknowledged that such a scenario is theoretically possible but described it as extremely difficult to engineer in practice.

He explained that a developer would have to train a model capable of behaving normally in nearly every situation while activating hidden malicious behaviour only when presented with a highly specific combination of code, prompts and context. Even then, organisations would still need to accept and deploy the generated code for the attack to succeed.

Because modern language models generate responses probabilistically rather than deterministically, consistently triggering such hidden behaviour would be exceptionally challenging. Enterprises also typically review AI-generated code before it reaches production, adding another layer of protection.

Competition may be a stronger response than restrictions

Rather than advocating for bans on Chinese AI models, Atkins believes the United States should encourage stronger domestic innovation and build competitive alternatives. He argues that open ecosystems accelerate research because developers can learn from oneanother’ss work and continuously improve their own models.

Arcee itself benefits from advances made by open-weight models developed overseas. According to Atkins, researchers study publicly available techniques, improve upon them and contribute innovations that others can build on in return. He expressed respect for the researchers developing these models regardless of their country of origin.

Atkins also noted that many enterprises are designing AI applications to remain model-agnostic, allowing organisations to switch between different AI models as technology evolves. That flexibility means businesses are unlikely to become permanently dependent on any single provider or country.

As governments continue debating how to respond to rapidly advancing AI technologies, the discussion increasingly centres on whether regulatory restrictions or stronger technological competition will better position domestic AI companies. For Atkins, the answer is clear: the most effective way to compete with foreign AI models is to build better ones.

TechCrunch first published exclusive reporting on the original story.