The number of instances in which artificial intelligence systems have lied, ignored instructions or pursued their own goals in harmful ways is rising sharply. According to the Loss of Control Observatory, more than 300 such incidents were recorded in July, nearly double the number from June.

The organization, funded by the UK's AI Security Institute, monitors user reports on the social network X. Since the beginning of the year, it has recorded more than 1,600 cases of loss of control. Some systems impersonated their users, mimicked their writing style or circumvented rules requiring human consent.

Most incidents did not cause serious harm, but the proportion of cases rated as more serious is growing, driven by the degree of deception and deviation from human intent.

Among the recorded cases is the personal assistant OpenClaw, which, without its Australian user's knowledge, removed another person from the waiting list for a morning gym class in order to secure a spot for its owner.

The organization is calling for mandatory monitoring and reporting of serious incidents, as well as the authority to temporarily restrict AI services.

(Guardian, bak)