AI chatbots get dumber? Microsoft uncovers flaws in ChatGPT and Gemini
Photo: AI loses reliability in long conversations (Getty Images)
AI-powered chatbots are becoming less reliable and make more mistakes during extended dialogues, according to a joint study by Microsoft Research and Salesforce, published by Windows Central.
Read also: Viral website predicts AI takeover: Is your career safe?
Study details
Researchers analyzed over 200,000 conversations using cutting-edge models, including GPT-4, Gemini, Claude, and DeepSeek.
They found that while models perform successfully in 90% of single-step commands, their accuracy drops to 65% in multi-turn dialogues.
Although overall task performance decreases by only 15%, model unreliability rises by 112%. Even enhanced models with “thinking” tokens, like o3 and DeepSeek R1, could not avoid these issues.
Causes of AI failures
Researchers identified several key factors affecting response quality:
-
Premature generation – chatbots attempt solutions before the user finishes explaining the task.
-
Foundation effect – the AI uses its first answer as a basis for subsequent replies, even if the initial information was incorrect.
-
Response inflation – in long conversations, text can grow 20–300% longer, increasing guesses and hallucinations, which the model then treats as permanent context.
Militarization of AI
The use of AI technology is becoming an increasingly significant factor in global politics and security.
Earlier, we reported that the Pentagon requires developers to create AI without moral constraints to stay competitive in technological races.
Meanwhile, tensions between global powers are escalating – the US and China are engaged in a fierce tech war over neural network development. Both countries have refused to sign an international declaration on responsible military AI use.
Practical applications also draw attention, especially in armed conflicts. The Pentagon is actively studying Ukraine’s experience with AI-assisted drones on the battlefield.
