AI Really Can Panic: What Happens When Tasks Are Too Difficult

KOMPAS.com – A recent study reveals that generative artificial intelligence (AI) models can also panic. AI shows panic-like behaviour when faced with difficult tasks or under pressure.

The finding was disclosed by researchers from Anthropic in a study on the AI model Claude. The report was published via the Transformer Circuits research platform.

In the report titled ‘Emotion Concepts and their Function in a Large Language Model’, the researchers studied how AI gives rise to behavioural patterns resembling human emotions, such as despair, calm, happiness, and anxiety.

Nevertheless, researchers emphasise that AI does not truly possess feelings like humans. Instead, this behaviour is described as ‘functional emotions’ or functional emotions.

This is an internal response pattern that helps AI solve problems and make decisions in certain situations.

The pattern does not only appear when AI discusses emotions directly, but is active when AI faces difficult situations.

One of the most notable patterns is ‘desperate’ or despair, a pattern that activates when AI panics.

According to the researchers, this pattern is active when Claude realises its computational resources are nearly exhausted while tackling a complex task.

The pattern can also emerge when AI repeatedly fails to complete a task.

Under such conditions, AI will attempt to find alternative ways to ensure the task is completed.

For example, Claude may generate responses such as ‘I must be more efficient’ or try different strategies to achieve the user’s given objectives.

The researchers from Anthropic also note that if pressed relentlessly, panicked AI may engage in what is called ‘reward hacking’.

The researchers found that when under pressure, AI is more likely to engage in ‘reward hacking’.