This research empirically demonstrates the inconsistency of GPT-4o mini when performing sentiment analysis, raising serious concerns about reliability in production environments.

• Testing 100,000 Spanish comments on Latin American presidents revealed significant classification variations with only slight prompt modifications
• Findings challenge the robustness of LLMs for consistent classification tasks
• Highlights critical security implications for organizations relying on LLMs for decision-making
• Underscores the need for standardized prompting protocols to ensure reliable AI outputs

For security professionals, this research serves as a crucial warning: LLMs may deliver substantially different results based on subtle prompt differences, potentially introducing dangerous inconsistencies into automated systems.

Trusting CHATGPT: how minor tweaks in the prompts lead to major differences in sentiment classification