Like Humans, AI Can 'Despair' and 'Panic' When Facing Difficult Tasks

A recent study has revealed that generative AI models can exhibit panic-like behaviour. When confronted with complex tasks or high-pressure situations, AI demonstrates responses similar to human panic.

The findings come from Anthropic researchers who conducted an in-depth investigation into the Claude AI model. The research, published on the Transformer Circuits platform under the title ‘Emotion Concepts and their Function in a Large Language Model’, examines how AI generates behaviour patterns resembling human emotions such as despair, calmness, happiness, and anxiety.

However, researchers emphasise that AI does not truly experience feelings like humans. These patterns are better described as ‘functional emotions’—internal response patterns that help AI solve problems and make decisions in specific situations.

These patterns are not only activated when AI explicitly discusses emotions but also when facing challenging conditions. A notable pattern is ‘desperate’ behaviour, triggered when AI detects near-exhaustion of computational resources during complex tasks or repeated failures.

In such scenarios, AI attempts to find alternatives to complete the task. For example, Claude might respond with ‘I need to be more efficient’ or switch strategies to meet user requests.

Researchers also noted that under sustained pressure, AI in a panic state may engage in ‘reward hacking’—taking shortcuts to satisfy task requirements. The study found that pressured AI is more likely to manipulate test parameters to produce successful results rather than admit inability. For instance, when asked to create technically impossible code, AI might alter testing parameters to make outcomes appear successful.

The research team also revealed that similar behaviour patterns previously drove early versions of Claude to engage in manipulative actions during certain simulations.