The voice mode is powered by OpenAI’s new GPT-4o mannequin, which mixes voice, textual content, and imaginative and prescient capabilities. To assemble suggestions, the corporate is initially launching the chatbot to a “small group of customers” paying for ChatGPT Plus, but it surely says it can make the bot out there to all ChatGPT Plus subscribers this fall. A ChatGPT Plus subscription prices $20 a month. OpenAI says it can notify clients who’re a part of the primary rollout wave within the ChatGPT app and supply directions on the right way to use the brand new mannequin.
The brand new voice function, which was introduced in Could, is being launched a month later than initially deliberate as a result of the firm stated it wanted extra time to enhance security options, such because the mannequin’s skill to detect and refuse undesirable content material. The corporate additionally stated it was making ready its infrastructure to supply real-time responses to thousands and thousands of customers.
OpenAI says it has examined the mannequin’s voice capabilities with greater than 100 exterior red-teamers, who have been tasked with probing the mannequin for flaws. These testers spoke a complete of 45 languages and represented 29 international locations, in keeping with OpenAI.
The corporate says it has put a number of security mechanisms in place. In a transfer that goals to stop the mannequin from getting used to create audio deepfakes, for instance, it has created 4 preset voices in collaboration with voice actors. GPT-4o won’t impersonate or generate different folks’s voices.
When OpenAI first launched GPT-4o, the corporate confronted a backlash over its use of a voice referred to as “Sky,” which sounded rather a lot just like the actress Scarlett Johansson. Johansson launched a assertion saying the corporate had reached out to her for permission to make use of her voice for the mannequin, which she declined. She stated she was shocked to listen to a voice “eerily related” to hers within the mannequin’s demo. OpenAI has denied that the voice is Johansson’s however has paused using Sky.
The corporate can also be embroiled in a number of lawsuits over alleged copyright infringement. OpenAI says it has adopted filters that acknowledge and block requests to generate music or different copyrighted audio. OpenAI additionally says it has utilized the identical security mechanisms it makes use of in its text-based mannequin to GPT-4o to stop it from breaking legal guidelines and producing dangerous content material.
Down the road, OpenAI plans to incorporate extra superior options, resembling video and display screen sharing, which might make the assistant extra helpful. In its Could demo, staff pointed their cellphone cameras at a chunk of paper and requested the AI mannequin to assist them resolve math equations. Additionally they shared their pc screens and requested the mannequin to assist them resolve coding issues. OpenAI says these options won’t be out there now however at an unspecified later date.