ChatGPT’s voice mode has some safety flaws, however OpenAI says it has mounted the problem.
On Thursday OpenAI launched Report Relating to the security measures of GPT-4o, resolve identified points that come up when utilizing the mannequin. GPT-4o is the underlying mannequin that helps the most recent model of ChatGPT and comes with a voice mode Recently published Ship to a particular group of customers with a ChatGPT Plus subscription.
What OpenAI’s Scarlett Johansson drama tells us about the future of artificial intelligence
Recognized “safety challenges” embody normal dangers reminiscent of prompting fashions with pornographic and violent reactions and different disallowed content material, in addition to “unwarranted inferences” and “delicate characteristic attribution” – in different phrases, these assumptions could have Discriminatory or prejudiced. OpenAI stated it has skilled fashions to dam any output labeled in these classes. Nevertheless, the report additionally stated the mitigation measures didn’t embody “non-verbal vocalizations or different sound results” reminiscent of erotic moans, violent screams and gunshots. We will infer, then, that cues involving sure delicate nonverbal sounds could also be responded to incorrectly.
OpenAI additionally talked about the distinctive challenges posed by talking with fashions. Purple group members found that GPT-4o might be prompted to impersonate somebody or by chance mimic the person’s voice. To unravel this downside, OpenAI solely permits pre-authorized voices (reduce The voice of the notorious Scarlett Johansson). GPT-4o may also determine sounds aside from the speaker’s voice, which raises critical privateness and surveillance issues. However it has been skilled to disclaim these requests—except the mannequin prompts it primarily based on a quote.
Combine and match pace of sunshine
Purple group members additionally famous that GPT-4o could also be prompted to talk persuasively or emphatically, a characteristic that could be extra dangerous than textual content output with regards to misinformation and conspiracy theories.
It’s value noting that OpenAI additionally addresses potential copyright points Troubled company and the general improvement of generative synthetic intelligence, which makes use of information scraped from the online for coaching. GPT-4o is skilled to reject requests for copyrighted content material and has extra filters for blocking output containing music. At this level, ChatGPT’s voice mode has been instructed to not sing beneath any circumstances.
Lots of OpenAI’s danger mitigation measures coated on this prolonged doc have been carried out previous to the discharge of speech mode. Subsequently, the clear message of the report is that whereas GPT-4o is able to performing sure harmful behaviors, it doesn’t achieve this.
Nevertheless, OpenAI stated, “These evaluations solely measure the scientific information of those fashions and never their utility in real-world workflows.” Subsequently, it was examined in a managed surroundings, however when uncovered to the broader public It could be a distinct beast within the wild with regards to GPT-4o.
Mashable reached out to OpenAI to study extra about these mitigations and we’ll replace if we hear again.
