Scale AI documents show Google extensively using ChatGPT to improve its AI chatbot: 'Make it BETTER than GPT'

Business Insider

19 hours ago

In 2023, Google was in a race to catch up with ChatGPT — and it turned to ChatGPT itself to do it.
Hundreds of documents obtained by Business Insider reveal that Google's contractors at Scale AI systematically used ChatGPT to improve Bard, Google's own chatbot at the time. When it launched earlier that year, Bard, which has since been renamed Gemini, was internally mocked as "rushed" and "botched."
Scale AI contractors generated thousands of responses from ChatGPT and compared them to their own "rewrites" of Bard's answers. They then improved their rewrites to exceed or at least match ChatGPT, feeding all the data back to Google.
Scale AI managers wrote in detail how ChatGPT's answers tended to have better formatting and more interesting facts. They ordered workers to "explain why gpt4 is better" and "make it BETTER than GPT." A single spreadsheet flagged dozens of contractors for writing responses "consistently worse than GPT4." In one instance, the document said contractors could get a 15% bonus for their responses performing better than ChatGPT.
Scale AI is a San Francisco startup that does crucial AI grunt work for Big Tech. It uses an army of human contractors to do things like labeling images and, as was the case with Google, rewriting chatbot responses. Meta is reportedly investing $15 billion in Scale AI as part of a blockbuster AI deal to buy almost half the company and hire its CEO, Alexandr Wang, for an in-house "superintelligence" team.
The documents obtained by BI showcase how closely Google monitored its chief rival's work.
OpenAI's terms of service at the time prohibited others from using its output "to develop models that compete with OpenAI." Scale AI and Google did not respond to a question about whether they got permission from OpenAI for these detailed comparisons and rewrites.
Scale AI told BI that the ChatGPT outputs weren't used to train Google's or any others' models and were part of routine "evaluations," which it said are industry standards.
"Scale did not, and does not, use ChatGPT responses to train Gemini or any models," a Scale AI spokesperson said in a statement. The spokesperson said that the documents describe "standard side-by-side evaluations, not the use of ChatGPT or any third-party model outputs for training."
"Doing side-by-side competitive evals is standard practice for the industry and those evaluation results are not used to train models," the spokesperson said.
Similarly, Google said, "Any suggestion that we have used other companies' models to train Gemini is inaccurate."
Experts told BI that this kind of comparison is indeed common at some top AI labs. Open AI, which is reportedly in partnership talks with Google Cloud, didn't respond to repeated requests for comment.
Project 'Bulba'
Scale AI gave Bard a catchy codename, "Bulba," after the Pokémon Bulbasaur. The mission was clear: compare Bulba's answers with ChatGPT's to make them better.
Scale AI never mentioned Google by name in the documents, referring instead to its anonymous "client." It references Bard over a dozen times in a private Google sheet titled "bard rewrite comparison with gpt4," and a slide in one training document includes Google's logo.
In July 2023, a manager ordered workers to study GPT-4's responses closely and figure out why they outperformed Bard's. "Try to come up with feedback that we can share so that experts can write responses better than GPT4 or at least the same," the manager wrote.
Scale AI also created a spreadsheet that compared 1,729 Bard rewrites directly to ChatGPT in October 2023. Each rewrite was rated with labels like "worse than GPT4" or "Needs Some Fixes." In one example, a worker rewrote a Bard review of a nursery chair that managers stamped "worse than GPT" because it "lacks detail compared to GPT4."
Another contractor's review of a Charleston history museum didn't make the cut either — a manager wrote that ChatGPT's version was "much better."
Scale AI also used ChatGPT to improve Bard's responses in specific domains, like engineering or physics. In an update from August 2023, Scale AI managers wrote that they would have staff "redo" Google's AI answers for engineering-related questions "with GPT4 guidance."
The documents showed that Scale AI and Google barred its contractors from copying and pasting ChatGPT responses directly into their rewrites, though, an issue many contractors were flagged for.
Scale AI says comparisons weren't for training
The internal documents BI reviewed described the project's goal as helping "train" Bard to give it more specific and complete answers, and refer to efforts to "improve the model."
Google did not answer follow-up questions on whether those comparisons influenced training. Scale AI said that there's a clear line between evaluating a model's performance and training it — and that ChatGPT outputs were only used for the former.
"There is a difference between training data and evaluation data," a spokesperson said. "Evaluation data is not ingested by a model to train it, but rather used to measure how well a model is performing."
Matthew Guzdial, an assistant computer science professor at the University of Alberta, says evaluation data can still influence an AI model.
"Even if all they're doing is looking at those outputs and rating that information to adjust the structure of the model, you could still make the argument that it's involved in the training process," he told BI.
The documents were left public
Scale AI, which has not previously made public details about its work with Google, left an over 300-page Google Doc public.
It contains dozens of links to other Google Docs, many of which are also public and contain sensitive information, including contractors' compensation details, personal email addresses, and performance reviews, along with still-functioning passwords to internal training sessions. Some of the Google Docs can still be edited by anyone who has the link.
Scale AI told BI that it is "actively investigating" how the document "may have been accessed" and is "taking steps to ensure any inadvertent exposure is remediated."
More than two days after BI told Scale AI about the public Google Doc, it was still online and available for anyone with the link to download.
Google is ahead on AI again
The documents don't specify how effective the comparison efforts were. Since its Bard flub in 2023, Google has rebranded Bard to Gemini and transformed into an AI shipping machine. Last month, it launched over 100 new AI products and features at I/O, its annual developer conference.
Google CEO Sundar Pichai began his speech at I/O by rattling off the industry benchmarks that Gemini is topping, touting the company's latest AI achievements.
"We are shipping faster than ever," Pichai said onstage.