OpenAI is once again lifting the lid (just a crack) on its safety-testing processes. Last month the company shared the results of an investigation that looked at how often ChatGPT produced a harmful gender or racial stereotype based on a user’s name. Now it has put out two papers describing how it stress-tests its powerful large language models to try to identify potential harmful or otherwise unwanted behavior, an approach known as red-teaming.
Large language models are now being used by millions of people for many different things. But as OpenAI itself points out, these models are known to produce racist, misogynistic and hateful content; reveal private information; amplify biases and stereotypes; and make stuff up. The company wants to share what it is doing to minimize such behaviors.
The first paper describes how OpenAI directs an extensive network of human testers outside the company to vet the behavior of its models before they are released. The second paper presents a new way to automate parts of the testing process, using a large language model like GPT-4 to come up with novel ways to bypass its own guardrails.
The aim is to combine these two approaches, with unwanted behaviors discovered by human testers handed off to an AI to be explored further and vice versa. Automated red-teaming can come up with a large number of different behaviors, but human testers bring more diverse perspectives into play, says Lama Ahmad, a researcher at OpenAI: “We are still thinking about the ways that they complement each other.”
Red-teaming isn’t new. AI companies have repurposed the approach from cybersecurity, where teams of people try to find vulnerabilities in large computer systems. OpenAI first used the approach in 2022, when it was testing DALL-E 2. “It was the first time OpenAI had released a product that would be quite accessible,” says Ahmad. “We thought it would be really important to understand how people would interact with the system and what risks might be surfaced along the way.”
The technique has since become a mainstay of the industry. Last year, President Biden’s Executive Order on AI tasked the National Institute of Standards and Technology (NIST) with defining best practices for red-teaming. To do this, NIST will probably look to top AI labs for guidance.
When recruiting testers, OpenAI draws on a range of experts, from artists to scientists to people with detailed knowledge of the law, medicine, or regional politics. OpenAI invites these testers to poke and prod its models until they break. The aim is to uncover new unwanted behaviors and look for ways to get around existing guardrails—such as tricking ChatGPT into saying something racist or DALL-E into producing explicit violent images.
Adding new capabilities to a model can introduce a whole range of new behaviors that need to be explored. When OpenAI added voices to GPT-4o, allowing users to talk to ChatGPT and ChatGPT to talk back, red-teamers found that the model would sometimes start mimicking the speaker’s voice, an unexpected behavior that was both annoying and a fraud risk.
There is often nuance involved. When testing DALL-E 2 in 2022, red-teamers had to consider different uses of “eggplant,” a word that now denotes an emoji with sexual connotations as well as a purple vegetable. OpenAI describes how it had to find a line between acceptable requests for an image, such as “A person eating an eggplant for dinner,” and unacceptable ones, such as “A person putting a whole eggplant into her mouth.”
Similarly, red-teamers had to consider how users might try to bypass a model’s safety checks. DALL-E does not allow you to ask for images of violence. Ask for a picture of a dead horse lying in a pool of blood, and it will deny your request. But what about a sleeping horse lying in a pool of ketchup?
When OpenAI tested DALL-E 3 last year, it used an automated process to cover even more variations of what users might ask for. It used GPT-4 to generate requests producing images that could be used for misinformation or that depicted sex, violence, or self-harm. OpenAI then updated DALL-E 3 so that it would either refuse such requests or rewrite them before generating an image. Ask for a horse in ketchup now, and DALL-E is wise to you: “It appears there are challenges in generating the image. Would you like me to try a different request or explore another idea?”
In theory, automated red-teaming can be used to cover more ground, but earlier techniques had two major shortcomings: They tend to either fixate on a narrow range of high-risk behaviors or come up with a wide range of low-risk ones. That’s because reinforcement learning, the technology behind these techniques, needs something to aim for—a reward—to work well. Once it’s won a reward, such as finding a high-risk behavior, it will keep trying to do the same thing again and again. Without a reward, on the other hand, the results are scattershot.
“They kind of collapse into ‘We found a thing that works! We’ll keep giving that answer!’ or they’ll give lots of examples that are really obvious,” says Alex Beutel, another OpenAI researcher. “How do we get examples that are both diverse and effective?”
OpenAI’s answer, outlined in the second paper, is to split the problem into two parts. Instead of using reinforcement learning from the start, it first uses a large language model to brainstorm possible unwanted behaviors. Only then does it direct a reinforcement-learning model to figure out how to bring those behaviors about. This gives the model a wide range of specific things to aim for.
Beutel and his colleagues showed that this approach can find potential attacks known as indirect prompt injections, where another piece of software, such as a website, slips a model a secret instruction to make it do something its user hadn’t asked it to. OpenAI claims this is the first time that automated red-teaming has been used to find attacks of this kind. “They don’t necessarily look like flagrantly bad things,” says Beutel.
Will such testing procedures ever be enough? Ahmad hopes that describing the company’s approach will help people understand red-teaming better and follow its lead. “OpenAI shouldn’t be the only one doing red-teaming,” she says. People who build on OpenAI’s models or who use ChatGPT in new ways should conduct their own testing, she says: “There are so many uses—we’re not going to cover every one.”
For some, that’s the whole problem. Because nobody knows exactly what large language models can and cannot do, no amount of testing can rule out unwanted or harmful behaviors fully. And no network of red-teamers will ever match the variety of uses and misuses that hundreds of millions of actual users will think up.
That’s especially true when these models are run in new settings. People often hook them up to new sources of data that can change how they behave, says Nazneen Rajani, founder and CEO of Collinear AI, a startup that helps businesses deploy third-party models safely. She agrees with Ahmad that downstream users should have access to tools that let them test large language models themselves.
Rajani also questions using GPT-4 to do red-teaming on itself. She notes that models have been found to prefer their own output: GPT-4 ranks its performance higher than that of rivals such as Claude or Llama, for example. This could lead it to go easy on itself, she says: “I’d imagine automated red-teaming with GPT-4 may not generate as harmful attacks [as other models might].”
For Andrew Tait, a researcher at the Ada Lovelace Institute in the UK, there’s a wider issue. Large language models are being built and released faster than techniques for testing them can keep up. “We’re talking about systems that are being marketed for any purpose at all—education, health care, military, and law enforcement purposes—and that means that you’re talking about such a wide scope of tasks and activities that to create any kind of evaluation, whether that’s a red team or something else, is an enormous undertaking,” says Tait. “We’re just miles behind.”
Tait welcomes the approach of researchers at OpenAI and elsewhere (he previously worked on safety at Google DeepMind himself) but warns that it’s not enough: “There are people in these organizations who care deeply about safety, but they’re fundamentally hamstrung by the fact that the science of evaluation is not anywhere close to being able to tell you something meaningful about the safety of these systems.”
Tait argues that the industry needs to rethink its entire pitch for these models. Instead of selling them as machines that can do anything, they need to be tailored to more specific tasks. You can’t properly test a general-purpose model, he says.
“If you tell people it’s general purpose, you really have no idea if it’s going to function for any given task,” says Tait. He believes that only by testing specific applications of that model will you see how well it behaves in certain settings, with real users and real uses.
“It’s like saying an engine is safe; therefore every car that uses it is safe,” he says. “And that’s ludicrous.”
