The study found that AI models fed with repetitive, unproductive data showed reduced logical reasoning, difficulty understanding context, and in some cases, violations of safety protocols. Among the models tested — Llama3 8B, Qwen2.5 7B, and Qwen3 4B — Meta’s Llama was the most affected, while Qwen3 4B performed slightly better but still showed signs of cognitive decline.
Although ChatGPT was not part of the study, researchers warn that AI systems could be affected if they are consistently provided with content that does not enhance their capabilities.
In related news, OpenAI recently launched a new browser, Atlas, designed to compete with Google Chrome. Atlas integrates ChatGPT directly into the browsing experience, allowing users to access AI across multiple tabs and enabling the AI to perform tasks such as research, planning, and booking activities.
Experts note that both humans and AI benefit from meaningful and high-quality input, and overexposure to trivial content can negatively impact performance.
