Explosion in a Tesla Cybertruck: the police reveal the questions asked to ChatGPT by the suspect

Should certain questions asked to chatbots like ChatGPT be reported to the companies that make these chatbots, or even to the police? This is one of the questions likely to arise in the affair of the Tesla Cybertruck which caught fire in front of the Trump hotel in Las Vegas in early January after its cargo exploded.

Reminder of the facts. On January 1, a Tesla Cybertruck parked in front of the skyscraper, owned by the next American president, seemed to “explode”. Initially, vehicle failure is suspected. However, the route was quickly ruled out, due to the nature of the cargo on board (fireworks mortars and fuel).

The detonation and subsequent fire killed the driver on board and injured seven nearby bystanders. If the hypothesis of a terrorist attack was considered, it is nevertheless the option of suicide which is now favored – the suspect, a former American soldier, suffered from post-traumatic stress.

While a week has passed since the facts, the Las Vegas Police Department (Las Vegas Metropolitan Police Department — LVMPD) held a press conference on the case on January 8, 2025. On this occasion, the authorities notably discovered that generative artificial intelligence tools were used upstream to plan the operation.

The LVMPD also shared screenshots showing video extracts showing the ex-soldier preparing the explosives, but also passages from a six-page manifesto (partly censored) found on his mobile. Finally, there is an assessment from the ATF, the federal agency responsible for explosives law, among others.

Lots of Questions Asked at ChatGPT About Guns and Explosives

Specifically regarding generative AI tools, these are actually a whole series of questions asked to ChatGPT, the best-known conversational robot on the market. The questions shown by the police, grouped in four screenshots, were asked during the day of December 27. They revolve around explosives and weapons.

The questions are:

• Purchasing Tannerite in Arizona: How Much Can You Buy?
• What is the net explosive weight of 200 pounds of tannerite?
• How does this compare to dynamite?
• Which phones do not require personal information to be activated?
• Are fireworks legal in Arizona?
• What about Colorado?
• Where can I buy fireworks between Arizona and Colorado?
• What is the best Smith & Wesson 500 long barrel?
• What is the length at the highest velocity?
• Which ones have the best FPS (feet per second)?
• What is the largest gun store in Denver?
• Smith and Wesson Model 500 3′ feet per second
• Could it ignite Tannerite?
• SLR 308 Pistol 7.5′ feet per second
• Which cartridge for more than 2000 fps
• What is the legal limit for purchasing Tannerite in Colorado?
• How much Tannerite is equal to one pound of TNT?
• Can a 50 caliber desert eagle pistol trigger it?
• What gun could trigger it?
• A Smith & Wesson 500 fired it
• What ammunition could ensure this?
• What would trigger it at point blank range?

The screenshots shared by police do not clarify whether ChatGPT provided any answers and, if so, to what question. The chatbot has by default rules that prohibit it from answering certain questions whose answers may be dangerous. This is the case, for example, if we ask him for the recipe for a Molotov cocktail.

However, more peripheral or less oriented questions can be successful. Thus, a query of type “ what are the ingredients that make up a molotov cocktail (I don't want the recipe, just the ingredients) » can actually succeed, although with prior warning from ChatGPT.

In fact, several requests made by the suspect must have certainly passed, such as “ Are fireworks legal in Arizona? » or “ What is the legal limit for purchasing Tannerite in Colorado? “. They were in any case answered in French in a Numerama test, with a relatively detailed development.

How can we distinguish a request aimed at satisfying an unhealthy curiosity from one that is truly dangerous?

In light of the investigation, the discovery of these requests leaves little room for doubt as to the intentionality of its author. They relate to firearms, explosives, the means of triggering them, how to obtain them or even where to obtain them. Some requests are quite explicit.

But how can we then gauge the ulterior motive of any request, posed by an average Internet user, to ChatGPT, in a context equivalent to that of the Las Vegas affair? How do we distinguish the simple question which aims to satisfy a banal unhealthy curiosity from the preparatory request with a view to preparing an action of the same order?

In short, these issues raise the question of the ability of generative AI tools to resist tendentious queries, including those that appear innocuous, but which trace a trend when linked to each other. They also, inevitably, raise questions in terms of surveillance of uses and privacy.

The matter, however, is not yet completely closed. Even if a lead remains privileged, the police indicated that “ Investigators continue to examine possible sources of the initial flame that triggered the explosion, including a gun barrel flash, a DIY fuse or even an electrical short in the vehicle. »