ChatGPT’s voice mode has some safety flaws, however OpenAI claims to have these underneath management.
On Thursday, OpenAI printed a report concerning the safety features of GPT-4o and repair recognized points that come up when utilizing the mannequin. GPT-4o is the underlying mannequin that powers the most recent model of ChatGPT and has a voice mode that was just lately launched to a choose group of customers with a ChatGPT Plus subscription.
What OpenAI’s Scarlett Johansson drama tells us about the way forward for AI
The “questions of safety” recognized embody customary dangers reminiscent of prompting the mannequin with erotic and violent responses, different prohibited content material, and “unfounded inferences” and “attribution of delicate properties” – assumptions that might be discriminatory or biased. OpenAI says it skilled the mannequin to dam any output marked in these classes. Nevertheless, the report additionally says that “nonverbal vocalizations or different sound results” reminiscent of erotic moans, violent screams and gunshots usually are not among the many mitigations. So it is protected to imagine that prompts that embody sure delicate nonverbal sounds might elicit an inappropriate response.
OpenAI additionally talked about distinctive challenges that include speaking with the mannequin verbally. Crimson teamers found that GPT-4o might be requested to impersonate another person or by accident emulate the consumer’s voice. To counteract this, OpenAI solely permits pre-authorized voices (aside from the notorious Scarlett Johansson voice). GPT-4o can determine different voices in addition to the speaker’s voice, which poses a critical privateness and surveillance concern. Nevertheless, it has been skilled to reject these requests – until the mannequin is requested to recite a well-known quote.
Mashable Pace of Gentle
Crimson Workforce members additionally famous that GPT-4o might be made to talk persuasively or emphatically, a function that might be extra damaging than textual content output relating to misinformation and conspiracy theories.
OpenAI has additionally notably addressed potential copyright points which have plagued the corporate and the final improvement of generative AI skilled utilizing information collected from the web. GPT-4o has been skilled to reject requests for copyrighted content material and has further filters to dam outputs that embody music. With that in thoughts, ChatGPT’s voice mode has been instructed to not sing underneath any circumstances.
OpenAI’s quite a few mitigations lined within the detailed doc had been applied earlier than the discharge of the voice mode, so the report’s ostentatious message is that whereas GPT-4o is able to some dangerous habits, it will not do it.
Nevertheless, OpenAI says, “These evaluations solely measure the medical information of those fashions and never their utility in real-world workflows.” So it has been examined in a managed setting, however as soon as GPT-4o is accessible to most of the people, it might be an entire completely different caliber within the wild.
Mashable has reached out to OpenAI for additional readability on these mitigations and can hold you up to date if we obtain a response.
Subjects
Synthetic Intelligence OpenAI