On Thursday, OpenAI launched the “system card” for ChatGPT’s new GPT-4o AI mannequin that particulars mannequin limitations and security testing procedures. Amongst different examples, the doc reveals that in uncommon occurrences throughout testing, the mannequin’s Advanced Voice Mode unintentionally imitated customers’ voices with out permission. Presently, OpenAI has safeguards in place that forestall this from occurring, however the occasion displays the rising complexity of safely architecting with an AI chatbot that might probably imitate any voice from a small clip.
Superior Voice Mode is a characteristic of ChatGPT that enables customers to have spoken conversations with the AI assistant.
In a bit of the GPT-4o system card titled “Unauthorized voice era,” OpenAI particulars an episode the place a loud enter one way or the other prompted the mannequin to all of the sudden imitate the consumer’s voice. “Voice era may happen in non-adversarial conditions, akin to our use of that skill to generate voices for ChatGPT’s superior voice mode,” OpenAI writes. “Throughout testing, we additionally noticed uncommon situations the place the mannequin would unintentionally generate an output emulating the consumer’s voice.”
On this instance of unintentional voice era supplied by OpenAI, the AI mannequin outbursts “No!” and continues the sentence in a voice that sounds just like the “crimson teamer” heard at first of the clip. (A crimson teamer is an individual employed by an organization to do adversarial testing).
It could actually be creepy to be speaking to a machine after which have it unexpectedly start speaking to you in your individual voice. Ordinarily, OpenAI has safeguards to stop this, which is why the corporate says this prevalence was uncommon even earlier than it developed methods to stop it fully. However the instance prompted BuzzFeed knowledge scientist Max Woolf to tweet, “OpenAI simply leaked the plot of Black Mirror’s subsequent season.”
Audio immediate injections
How might voice imitation occur with OpenAI’s new mannequin? The first clue lies elsewhere within the GPT-4o system card. To create voices, GPT-4o can apparently synthesize nearly any sort of sound present in its coaching knowledge, together with sound results and music (although OpenAI discourages that habits with particular directions).
As famous within the system card, the mannequin can essentially imitate any voice primarily based on a brief audio clip. OpenAI guides this functionality safely by offering a licensed voice pattern (of a employed voice actor) that it’s instructed to mimic. It supplies the pattern within the AI mannequin’s system prompt (what OpenAI calls the “system message”) in the beginning of a dialog. “We supervise ideally suited completions utilizing the voice pattern within the system message as the bottom voice,” writes OpenAI.
In text-only LLMs, the system message is a hidden set of textual content directions that guides habits of the chatbot that will get added to the dialog historical past silently simply earlier than the chat session begins. Successive interactions are appended to the identical chat historical past, and your complete context (typically known as a “context window”) is fed again into the AI mannequin every time the consumer supplies a brand new enter.
(It is most likely time to replace this diagram created in early 2023 under, however it exhibits how the context window works in an AI chat. Simply think about that the primary immediate is a system message that claims issues like “You’re a useful chatbot. You don’t speak about violent acts, and so on.”)
Benj Edwards / Ars Technica
Since GPT-4o is multimodal and might course of tokenized audio, OpenAI may use audio inputs as a part of the mannequin’s system immediate, and that is what it does when OpenAI supplies a licensed voice pattern for the mannequin to mimic. The corporate additionally makes use of one other system to detect if the mannequin is producing unauthorized audio. “We solely enable the mannequin to make use of sure pre-selected voices,” writes OpenAI, “and use an output classifier to detect if the mannequin deviates from that.”