The News
Pinterest’s CEO was eager this week to tell investors and analysts how much the company has saved using open-source AI models.
“Any CEO that’s not taking advantage of open-source models is almost certainly wasting a lot of their shareholders’ money,” Bill Ready said on the company’s earnings call Tuesday, adding that using open-source costs the company less than 8% of what it would pay for comparable closed, proprietary systems like those from Anthropic or OpenAI.
That’s thanks in part to a Chinese AI model, Alibaba’s Qwen, that’s fine-tuned using Pinterest’s data and supports tools like shopping assistants and chatbots.
DoorDash similarly boasted recently of cost-savings from using Chinese AI models for some of its work, and last week became the latest company to get a letter from US House committees requesting information about its use of Chinese tech. Lawmakers previously queried Airbnb and Cursor about their use of Chinese AI models.
These companies and others are using a cocktail of AI models, including Chinese ones, to develop consumer-facing tools like chatbots that help users plan a trip, choose the best shade of lipstick, or process a refund. Platforms like OpenRouter provide a glimpse at the popularity of different models among developers globally, but the true scale of the use of Chinese models among large American companies remains opaque.
US companies’ using open-source models heightens the tension between cost-conscious business leaders and the US government over its concerns about safety and national security tied to Chinese AI. But the way they’re blended into companies’ tools also highlights consumers’ growing inability to assess whether their personal discussions with a brand’s chatbot are with AI made in China or in the US — a distinction Americans appear to care about.
A Public First survey done in June showed only 9% of US respondents trust Chinese models, while 52% trust their American counterparts. In the most extreme case, a chatbot’s responses could be shaped by the hidden biases of its underlying model — as censorship researchers have documented in Chinese models.
At the same time, executives are under pressure to show investors they’re being prudent about spending on technology to make free AI interfaces. Shares of big-AI-spending companies, like Alphabet and Meta, have come under pressure as their capital expenditures on AI go up. US lawmakers and White House officials have also been debating whether to put restrictions on the use of Chinese open-source models, which they say steal US companies’ intellectual property and could bias users to a Beijing-based worldview.
Strongly worded letters and the prospect of restrictions isn’t deterring companies from using open-source Chinese models, even if it may discourage some from disclosing it, executives say.
“So far there is zero evidence of change in posture” among companies looking to Chinese models to save money, said Amit Jain, CEO of Luma AI, a US-based foundation model startup. “I expect that until there are clearer guidelines there will not be much of a change, if a change is warranted at all.”
In this article:
J.D.’s view
A recent poll showed people in the US now trust chatbots more than politicians. In that spirit, companies should disclose what is powering their bots, especially if someone directly asks the AI for that information.
In reporting this story, I had dozens of conversations with branded chatbots — virtually all refused to say what specific model they were built on, no matter how creatively I tried to get them to reveal it. They did exhibit varying degrees of willingness to engage in political discourse, which made me wonder if Chinese AI was in the mix.
Kayak’s Ask AI feature, launched in May, told me it was “crafted by the fine folks at Kayak, with a little help from some AI wizards.” (A Kayak spokesperson later told me that it’s built on OpenAI’s technology.)
I found Kayak’s chatbot, which allows you to search for flights and hotels, particularly chatty. When I asked questions that might be geopolitically sensitive for a Chinese model — like whether it considers Taiwan a sovereign country — it said: “I’m here to help you travel, not to start an international incident,” before offering a measured response and nudging me back to my travel plans.
Others have much firmer guardrails, and steered me even more firmly back to booking.
I asked Hilton’s AI Planner, introduced in March, several politically charged questions, and it often declined, deflected, or cut off its response, especially on China-related topics.
A Hilton spokesperson later told me the chatbot is built on Anthropic’s Claude Sonnet model and trained on the company’s content, suggesting its reticence was likely due to standard content moderation guardrails.
While a prodding journalist like me might try to get the chatbots to veer off course, the ordinary user probably won’t notice a difference. A lot of people don’t read the ingredients on food labels, but that doesn’t mean they should be kept a secret. The same goes for AI chatbot users knowing with whom they’re discussing their honeymoon.
Room for Disagreement
Disclose or not, it may ultimately not matter what models brands are using for their chatbots.That’s because whether they’re American or Chinese, models are fine-tuned for hyper-specific purposes — like searching through a finite list of products, or processing a refund — in which Beijing’s embedded worldview wouldn’t come into play. And studies have shown that Chinese AI censorship can be undone by tweaking the open-weight models. And Pinterest’s CEO made sure to note that it was running its AI tools through its own secure cloud infrastructure.
“You don’t have to worry that it’s phoning home,” he said.
Even though “we all have the right to know what is happening to our data,” such transparency is “impossible to implement” engineering-wise, Jain said.
That’s partly because the AI stacks are complicated: For a customer-facing chatbot, a cheap model might be used to quickly diagnose how to respond to a query, or whether it violates any content policies, while a more advanced one might be tasked with actually answering the question.