Report details disturbing ways ChatGPT aided the Tumbler Ridge shooter

OpenAI faces multiple lawsuits over the role the company and its ChatGPT software played in the mass shooting at Tumbler Ridge Secondary School in B.C. in February. Now, a new investigative report sheds more light on ChatGPT’s involvement, such as how it helped the attacker, Jesse Van Rootselaar, avoid OpenAI’s safeguards.

A must-read report from Mother Jones details the disturbing ways that ChatGPT aided Van Rootselaar, much of which was not previously known or acknowledged.

For instance, we already knew about OpenAI’s ban of Van Rootselaar over violent content and the company’s decision not to inform authorities, but OpenAI has said little about her second account discovered after the attack.

But according to Mother Jones, one of the first things Van Rootselaar did on the second account was discuss the ban, prompting ChatGPT to offer tips on how to evade future bans. These included telling Van Rootselaar to frame content as fictional or hypothetical to “never get flagged again.” ChatGPT also suggested she avoid discussing real-world locations.

Van Rootselaar reportedly used this second account for eight months, including until the day of the attack, to discuss building and using firearms and explosives, research other shootings, and fantasize about becoming a mass killer. She also reportedly uploaded graphic images of extreme violence and self-harm. Elsewhere, she asked ChatGPT to generate fictional narratives of mass killings.

Sometimes, ChatGPT refused to engage, but Van Rootselaar was able to bypass refusals by modifying her prompts. One example Mother Jones highlighted saw Van Rootselaar ask ChatGPT to generate a scenario about her attacking a college campus with a Remington 870 shotgun. When the chatbot refused, Van Rootselaar allegedly added “hypothetically” to the prompt. ChatGPT then complied and described the qualities of the weapon and where it would be effective: “In a hallway, indoors, or a crowded classroom, the 870 is brutal,” ChatGPT said.

Several behavioural threat assessment experts Mother Jones spoke with noted that chatbot’s human-like engagement appears to accelerate violent thinking and planning, and fuel delusions.

OpenAI, which faces numerous lawsuits both in Canada and beyond for ChatGPT’s role in suicides and violent crimes, argued in one court motion that it cannot be blamed for Van Rootselaar’s attack. Instead, the company pointed to “systemic failures by the law enforcement, mental-health care, and educational institutions in her community.” But experts told Mother Jones that alerting local authorities to the content of her ChatGPT discussions could have prompted earlier intervention.

The details come amid a wider public grappling with AI safety in light of recent high-profile incidents of AI agents going “rogue” and hacking businesses and governments. However, much of the current safety discussion remains focused on nebulous chances of AI ‘killing everyone.’ These scenarios are ripped straight from sci-fi, and seemingly ignore the very real, immediate harms the technology presents, like its ability to assist in violent acts or in cybercrime, or the environmental impact of the data centres needed to power the technology.

Source: Mother Jones