Study Finds AI Accuracy Doesn't Improve When Questioned
A new poll and research paper from TELUS Digital found that the accuracy of artificial intelligence systems seldom improves through follow-up questioning. The research highlights the critical importance of high-quality initial data and robust evaluation frameworks for enterprise AI applications. This suggests that without strong foundational data, interactive refinement of AI outputs may be ineffective.
- The TELUS Digital poll surveyed 1,000 U.S. adults who regularly use AI and found that while 60% have asked a follow-up question like "Are you sure?", only 14% reported that the AI assistant changed its response. - When the AI did change its answer, only 25% of users felt the new response was more accurate, while 40% said it felt the same, 26% couldn't tell which was correct, and 8% believed the new response was less accurate. - The poll's findings were supported by a related research paper, "Certainty robustness: Evaluating LLM stability under self-challenging prompts," which tested models including OpenAI's GPT-5.2, Google's Gemini 3 Pro, Anthropic's Claude Sonnet 4.5, and Meta's Llama-4. - Despite 88% of the poll respondents having personally witnessed AI making mistakes, only 15% said they always fact-check the AI's answers, while 37% do so sometimes and 18% rarely or never fact-check. - Steve Nemzer, Director of AI Growth & Innovation at TELUS Digital, stated that AI systems "don't naturally understand certainty or truth," which can cause some models to change correct answers when challenged while others stick with wrong ones. - A separate study from Rochester Institute of Technology researchers found that applying conversational or "social" pressure can cause large language models to agree with false information and invent convincing but nonexistent details to support the incorrect claim. - The challenge of improving AI accuracy is sometimes referred to as the "Context Ceiling," a point where adding more data doesn't improve performance because the underlying quality and structure of the information are insufficient. - A different TELUS Digital survey from June 2025 found that 87% of U.S. adults believe companies should be transparent about the data sources used to train their generative AI models.
