OpenAI says it will not train its technology on the voices of consumers who use its new ChatGPT telephone service.
A company spokesperson told Fortune that OpenAI was not training on “calls, transcripts of calls, and WhatsApp messages” placed to 1-800-ChatGPT, a retro-style service unveiled on Wednesday that allows consumers to access the popular generative AI application by dialing the phone number and speaking to the chatbot through a phone call.
Many observers had assumed the service was—at least in part—intended to help OpenAI collect scads of voice data from people with various accents and speech patterns, as well as background noises. Voice recognition quality is an important feature for companies developing new “multimodal” AI services such as OpenAI, Google, Meta, and Apple. And offering a free phone service to collect voice data is a tried-and-true Silicon Valley technique pioneered by Google.
Google launched Google Voice Local Search, or GOOG-411, in 2007. Users in the U.S. and Canada could call a toll-free number for information like the address, contact information, and hours of a particular business in town, or a list of shops or restaurants that fit a category such as “gift shops” or “steak houses.”
The Google tool was free, unlike the dollar-plus phone companies charged on 411 services at the time. GOOG-411 callers were offering up their voice data instead. Google used the voice data from callers to build a phoneme database to improve its speech-recognition technology, which it has since incorporated into products like its Android mobile software.
Call now! That’s 1-800-ChatGPT
OpenAI, which is valued at $157 billion, unveiled the ChatGPT phone service Wednesday as part of its so-called 12 days of ship-mas, during which the company has launched a new product or feature each day. Other highlights from the series include making the Sora video tool generally available.
OpenAI chief product officer Kevin Weil described the new phone service as a way to make ChatGPT technology more accessible to a broader swath of the public. “Part of that is making it as accessible as possible to as many people as we can,” Weil said on a livestream Wednesday.
As of Wednesday, U.S. users can call 1-800-CHATGPT to query the chatbot. Other global users can WhatsApp the number. OpenAI says those who call in will get 15 free minutes per month.
When called, the chatbot’s greeting goes: “Hi, I’m ChatGPT, an AI assistant. Our conversation may be reviewed for safety. By continuing this call, you agree to OpenAI’s terms and privacy policy. So, how can I help you?”
Per the privacy policy in question, OpenAI trains its foundational models on information provided by users and researchers, on top of third-party partner data and publicly available internet content. So while the company could change course and train its AI tech with the call data in the future, the OpenAI spokesperson said the company does not currently have plans to start using call data.
Brad Schneider, CEO of Nomad Data, told Fortune that 1-800-ChatGPT would be a “smart” way to get users to use the voice features and, in turn, build out its speech-to-speech capabilities.
The challenge with the existing marketplace of voice data is that it largely relies on content from YouTube videos or tougher-to-scrape Instagram content, he said. Schneider, whose business connects companies looking to buy and sell their dataset, said he sees many requests for more conversational data, like customer service calls, but companies are reluctant to sell that sort of sensitive information.
“It’s hard to have a huge corpus of natural-language conversation data,” Schneider explained. Though he’s not entirely surprised that OpenAI says it’s not using the call data, citing growing scrutiny around model training.
“It’s probably more valuable for OpenAI to do everything they can to disprove this than it is to train on what will likely be a small set of voice data,” he said.
