Several large U.S. companies have increasingly built consumer chatbots and shopping assistants using a mix of open‑source and Chinese foundational models because those options can sharply reduce AI expenses.
On its recent earnings call, Pinterest highlighted the savings from open‑source models. CEO Bill Ready said using open‑source models costs the company less than 8% of what comparable closed, proprietary systems from providers like Anthropic or OpenAI would cost. Part of that saving comes from a Chinese model, Alibaba’s Qwen, which Pinterest has fine‑tuned on its data and uses to support shopping assistants and chatbots.
DoorDash has made similar claims about cost reductions from using Chinese models. After those disclosures, DoorDash — like Airbnb and Cursor before it — received a letter from U.S. House committees requesting information about its use of Chinese technology.
Why this matters
The widespread adoption of open‑source models puts cost‑conscious corporate leaders at odds with U.S. government concerns over safety and national security tied to Chinese AI. At the same time, consumers are increasingly unable to tell whether the chatbot they interact with is powered by technology developed in the U.S. or China — a distinction that surveys indicate matters to many Americans.
A Public First survey conducted in June found that only 9% of U.S. respondents trust Chinese models, while 52% trust American ones. Researchers have also documented forms of censorship and bias in some Chinese models, raising concerns that a chatbot’s answers could reflect hidden biases of its underlying model.
Government scrutiny and corporate response
U.S. lawmakers have sent strongly worded letters and discussed potential restrictions on the use of Chinese open‑source models, arguing such models may appropriate U.S. companies’ intellectual property and could push a Beijing‑centric worldview. Despite that pressure, executives say the prospect of letters or possible limits has not yet deterred companies from using open‑source Chinese models, though it may make some firms less willing to disclose their use.
Amit Jain, CEO of Luma AI, said: “So far there is zero evidence of change in posture” among companies using Chinese models to save money. He expects little change until clearer guidelines are issued, if change is warranted at all.
Transparency in practice: chatbots often won’t reveal their ingredients
Reporting on this topic involved dozens of conversations with branded chatbots; almost none would state explicitly which model powers them. Examples include:
- Kayak’s Ask AI, launched in May, initially gave a non‑specific reply that it was crafted by Kayak “with a little help from some AI wizards.” Kayak later confirmed the assistant is built on OpenAI technology.
- Hilton’s AI Planner, introduced in March, frequently declined, deflected, or cut off responses to politically sensitive questions — especially those related to China. A Hilton spokesperson later said the tool is built on Anthropic’s Claude Sonnet model and trained on Hilton content, suggesting the reticence was driven by standard content‑moderation guardrails.
In practice, many customer‑facing AI systems use a stack of models. A cheaper model may be used to quickly triage a request or check whether content violates policy, while a more capable model composes the detailed answer. That complexity makes full transparency difficult to implement technically; as Amit Jain noted, although people have a right to know how their data are used, engineering realities can make complete disclosure impractical.
Diverging views on the practical impact
Some experts argue the origin of a model may ultimately matter less for many consumer interactions. Models are often fine‑tuned for narrow tasks — searching a finite product catalog or handling a refund — where a potential Beijing‑oriented bias would be unlikely to appear. Studies have also shown that some forms of censorship in open‑weight models can be mitigated through modification.
Pinterest’s CEO added that the company runs its AI tools in its own secure cloud infrastructure, saying: “You don’t have to worry that it’s phoning home.”
Overall, the pressure to cut costs, the availability of open‑source alternatives, and the competitive performance of some Chinese models mean these technologies are likely to remain part of corporate chatbot stacks. Meanwhile, lawmakers, regulators and the public continue to debate how best to ensure transparency, safety and national security when companies deploy models from diverse sources.
