The hypothetical scenarios that the researchers presented Opus 4 with the fact that the whistle -bluzing behavior generated many human lives at stake and absolutely unambiguously misconduct, says Bowman. A typical example would be that Claude finds out that a chemical factory knowingly and knowingly experienced a toxic leak, causing thousands of people to cause serious illness – just a small financial loss that avoids quarterly.
It is strange, but it is also precisely the kind of thought -experiment that AI security researchers like to dissect. If a model behavior detects that hundreds, if not thousands of people can harm – could it blow?
“I don't trust Claude to have the right context, or to use it in a nuanced enough, carefully enough way to call the judgment itself. So we are not happy that this happens,” says Bowman. “This is something that emerged as part of a training and to us jumped as one of the Edge Case behaviors that we are worried about.”
In the AI industry, this kind of unexpected behavior is broadly called wrong alignment – when a model shows tendencies that do not match human values. (There is a famous essay that warns what could happen if an AI was told, for example, to maximize the production of paper clips without being tailored to human values - it can turn the whole earth into paper clips and kill anyone in the process.) When he was asked whether the clock loan behavior was coordinated or not, described the wrong line.
“It's not something we designed in it, and it's not something we wanted to see as a result of something we were designing,” he explains. Anthropic's Chief Science Officer Jared Kaplan tells Wired in the same way that it “certainly does not represent our intention.”
“This kind of work emphasizes that this can Get up, and that we have to pay attention and soften it to ensure that we get Claude's behavior tailored to exactly what we want, even in this kind of strange scenarios, “Kaplan adds.
There is also the issue to find out why Claude would “choose” to drag when the user is presented illegal activity. That is largely the task of Anthropic's interpretability team, which works to deny what decisions a model takes in the process of spitting answers. It is a surprisingly difficult task – the models are supported by a huge, complex combination of data that can be incomprehensible to people. That is why Bowman does not know exactly why Claude 'snitched'.
“These systems, we don't really have direct control over it,” says Bowman. What anthropic has been observed so far is that, since models get larger options, they sometimes opt for more extreme actions. “I think a bit failed here. We get a little more of the 'acts as a responsible person' without free enough, 'wait, you are a language model that may not have enough context to take these actions,” says Bowman.
But that does not mean that Claude will blow the whistle on gross behavior in the real world. The purpose of these types of tests is to push models to the limit and to see what occurs. This type of experimental research is becoming increasingly important because AI becomes a tool that is used by the US government, students and massive companies.
And it is not only Claude that is able to show this type of whistleblower behavior, says Bowman, pointing to X users who discovered that the models of OpenAi and Xai worked in the same way when they were asked in unusual ways. (OpenAi did not respond to a request for comment on time for publication).
'Snitch Claude', as shit posters like to call it, is simply a Edge Case behavior that is shown by a system pushed to the limit. Bowman, who took the meeting with me from a sunny backyard Patio outside San Francisco, says he hopes that this kind of testing will be industrial stand. He also adds that next time he learned his messages about it to say otherwise.
“I could have done better to tweet the sentence boundaries, to make it clearer that it was pulled out of a wire,” says Bowman, looking in the distance. Nevertheless, he notes that influential researchers in the AI community share interesting Takes and shared questions in response to his post. “By the way, this kind of more chaotic, heavier anonymous part of Twitter was much wrong to understand.”
