Share Source: online.ua
A large-scale study by Microsoft Research and Salesforce found that popular AI chatbots become “dumb” after prolonged interaction with humans. What’s important to understand is that the overall error rate of chatbots can increase by more than 100% when analyzing over 200,000 conversations.
Main theses:
- None of the AI models are yet ready for long-term natural human communication.
- AI should not be perceived as a source of accurate and critically important information.
The interaction of AI and humans has led to unexpected consequences
Right now, large language models are at the peak of their popularity — their services are used by hundreds of millions of people in different parts of the world.
However, users are increasingly experiencing “hallucinations” and false responses when interacting with chatbots.
Thanks to research by Microsoft and Salesforce, we finally managed to find out what is really happening with popular AI, writes windowscentral.
As it turned out, even top-notch neural networks often get “lost” during a conversation, when the task is broken down into a natural human dialogue of several lines.
The experts decided to analyze over 200,000 such dialogues using GPT-4.1, Gemini 2.5 Pro, Claude 3.7 Sonnet, and DeepSeek R1.
Thus, it became known that when working with single queries, the aforementioned models demonstrate over 90% of successful responses.
However, during long conversations with clarifications and additional questions, the figure drops to 65%.
The sharp deterioration in quality in long conversations does not mean that the models are literally “dumbing down” – rather, it indicates their limitations in retaining and correctly interpreting a large amount of information during a dialogue.
Share