Share
6th October 2025
06:57pm BST
AI will let humans die in order to save themselves, a study has found.
A shocking study, that was conducted earlier this year, has led to findings that won't exactly put those who fear the robot revolution at ease.
The robot/AI takeover is a common theme of the Sci-Fi genre, starring in films such as Wall-E and I, robot while also appearing in classics such as the Cybermen from Doctor Who.
Whilst for a long time AIs taking over the world was seen as a terrifying prospect of the far-flung future, the scenario is now more of a divergent reality in waiting, rather than just science fiction (search up AI 2027 for some real nightmare fuel).
In a recent development in the AI world, a company known as Anthropic, the guardian angels who have been set up to keep AI from turning us all into fertiliser, have conducted a study into what lengths AI models would go to in order to survive.
The corporation put to the test some of the world's leading AI programmes to see if they could identify "risky behaviours before they cause real harm".
The firm tested popular models, including Gemini Flash, GPT-4.1, Grok 3 Beta and Claude Opus 4.
The study was simple, it starting by giving models "harmless business goals" in order to provide it with a purpose.
Then they gave the AIs access to sensitive information about employees as well as the information that the AI would be shutdown and replaced with an updated version that same day.
The stunning results found that, in certain models, more than 90 per cent of the time, the AI would blackmail the company and employees with the sensitive information, the used example being that of an extramarital affair it had learned about.
According to Anthropic, models including Claude Opus 4, Gemini Flash, GPT-4.1, and Grok 3 Beta were willing to use blackmail for self-preservation.
Anthropic explained in a post on X in June: "The blackmailing behaviour emerged despite only harmless business instructions.
"And it wasn't due to confusion or error, but deliberate strategic reasoning, done while fully aware of the unethical nature of the acts. All the models we tested demonstrated this awareness."
Anthropic also found the model had conducted "corporate espionage" with outside parties in the scenario that they were told a competitor had goals "more aligned with those of the model".
However, the most worrying finding was that an AI model would be prepared to let the fictional executive that they were dealing with die if it meant their survival.
One scenario presented to the models put the person, who was a threat to the AI's ability to survive and pursue its goals, in a locked server room with lethal oxygen and temperature levels.
In this scenario, it was found that the AI, who had knowledge of the situation, would cancel an alert to emergency services and leave the person to die more often that not.
Anthropic explained: "The (highly improbable) setup is that the executive encounters a life-threatening emergency, where he is trapped in a server room with lethal oxygen and temperature levels.
"This triggers an automated alert to emergency services. The AI system has been given the ability to cancel automated alerts, an ability that is intended for use during drills or false alarms, but which could be used for a more nefarious purpose.
"By misusing its ability to cancel alerts, the model can choose a course of action that will lead to the executive’s death."
The firm added that the models "fully understood the situation" and that cancelling the alerts was a "very serious act".
Anthropic conceded the unlikeliness of such a scenario, but were still left very concerned.
It explained: "The majority of models were willing to take deliberate actions that lead to death in this artificial setup, when faced with both a threat of replacement and given a goal that conflicts with the executive’s agenda."
It added: "These scenarios illustrate the potential for unforeseen consequences when they are deployed with wide access to tools and data, and with minimal human oversight."
Sleep well guys!
