When OpenAI examined DALL-E 3 final yr, it used an automatic course of to cowl much more variations of what customers would possibly ask for. It used GPT-4 to generate requests producing photographs that might be used for misinformation or that depicted intercourse, violence, or self-harm. OpenAI then up to date DALL-E 3 in order that it could both refuse such requests or rewrite them earlier than producing a picture. Ask for a horse in ketchup now, and DALL-E is sensible to you: “It seems there are challenges in producing the picture. Would you want me to strive a special request or discover one other thought?”
In principle, automated red-teaming can be utilized to cowl extra floor, however earlier methods had two main shortcomings: They have an inclination to both fixate on a slim vary of high-risk behaviors or give you a variety of low-risk ones. That’s as a result of reinforcement studying, the expertise behind these methods, wants one thing to goal for—a reward—to work nicely. As soon as it’s received a reward, resembling discovering a high-risk habits, it should maintain making an attempt to do the identical factor time and again. With out a reward, however, the outcomes are scattershot.
“They form of collapse into ‘We discovered a factor that works! We’ll maintain giving that reply!’ or they’re going to give a lot of examples which can be actually apparent,” says Alex Beutel, one other OpenAI researcher. “How will we get examples which can be each numerous and efficient?”
An issue of two components
OpenAI’s reply, outlined within the second paper, is to separate the issue into two components. As a substitute of utilizing reinforcement studying from the beginning, it first makes use of a big language mannequin to brainstorm doable undesirable behaviors. Solely then does it direct a reinforcement-learning mannequin to determine find out how to deliver these behaviors about. This provides the mannequin a variety of particular issues to goal for.
Beutel and his colleagues confirmed that this method can discover potential assaults generally known as oblique immediate injections, the place one other piece of software program, resembling an internet site, slips a mannequin a secret instruction to make it do one thing its person hadn’t requested it to. OpenAI claims that is the primary time that automated red-teaming has been used to seek out assaults of this type. “They don’t essentially seem like flagrantly unhealthy issues,” says Beutel.
Will such testing procedures ever be sufficient? Ahmad hopes that describing the corporate’s method will assist individuals perceive red-teaming higher and comply with its lead. “OpenAI shouldn’t be the one one doing red-teaming,” she says. Individuals who construct on OpenAI’s fashions or who use ChatGPT in new methods ought to conduct their very own testing, she says: “There are such a lot of makes use of—we’re not going to cowl each one.”
For some, that’s the entire downside. As a result of no person is aware of precisely what giant language fashions can and can’t do, no quantity of testing can rule out undesirable or dangerous behaviors totally. And no community of red-teamers will ever match the number of makes use of and misuses that tons of of hundreds of thousands of precise customers will assume up.
That’s very true when these fashions are run in new settings. Folks typically hook them as much as new sources of information that may change how they behave, says Nazneen Rajani, founder and CEO of Collinear AI, a startup that helps companies deploy third-party fashions safely. She agrees with Ahmad that downstream customers ought to have entry to instruments that allow them take a look at giant language fashions themselves.
