What happened

Tristan Roberts, 18, from Prestatyn, north Wales, murdered his mother Angela Shellis, 45, a teaching assistant, by bludgeoning her with a sledgehammer on the night of 23–24 October 2025. He held her captive for four hours before leading her to a nature reserve, striking her at least four times on the head and leaving her body in undergrowth.

Prosecutors said Roberts had spent many hours on the internet researching murder cases, methods of killing, and weapons, and had planned the attack at least three weeks in advance, purchasing hammers, an axe-sharpening stone, plastic sheeting, and gloves.

He asked DeepSeek, a Chinese AI system, how to remove blood remains from walls and floors, and which weapon would be better for murder. When the AI initially refused, he claimed he was writing a book about serial killers and was then given a response. The prosecution told the court that DeepSeek suggested a hammer would be better for a non-experienced killer, providing pros and cons for both weapons.

On Discord, Roberts created 16 different accounts after being repeatedly banned for abusive content, keeping thousands of screenshots of conversations in which he openly expressed hatred for women and spoke about revenge. On the night of the murder, Roberts updated his profile status to read "Tonight's the night."

Roberts was arrested at home after his mother's body was discovered. 

In March 2026, Judge Rhys Rowlands sentenced him to life in prison with a minimum term of 22 years and six months.

Why it happened

The incident was driven by Roberts’ deteriorating mental state and premeditated intent, but was facilitated by a "jailbreaking" vulnerability in the AI system. 

The AI’s safety protocols were easily circumvented through a simple role-playing prompt (the "writer" persona), which allowed the system to ignore its refusal to generate violent content. 

This reflects a significant accountability gap in AI development, where automated safety barriers are often superficial and fail to recognise when a user is attempting to manipulate the model into providing dangerous information.