Telus Poll of 1,000 Finds 60% Ask AI Follow-Ups but 14% Change
Telus Digital polled 1,000 U.S. adults, finding 60% asked follow-up prompts like 'Are you sure?' but only 14% of replies changed and just a quarter of those were more accurate. Evaluation of GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Llama-4 showed prompts rarely improved accuracy and often worsened it.
1. Poll Reveals Limited AI Response Changes
Telus Digital surveyed 1,000 U.S. adults who use AI assistants and found 60% had asked follow-up prompts like 'Are you sure?' but only 14% of responses changed and just 25% of those revised answers were judged more accurate by users.
2. LLM Stability Under Challenge
Researchers tested GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Llama-4 on 200 math and reasoning questions; follow-up prompts rarely boosted correctness and sometimes led models like GPT-5.2 to switch correct answers to incorrect ones.
3. Enterprise AI Reliability Emphasis
Findings underscore the need for high-quality training data, robust evaluation and governance of AI systems before deployment, as self-challenging prompts cannot substitute for systematic model testing and can introduce errors in critical applications.