TELUS Digital Study Finds AI Follow-Up Questions Rarely Improve Accuracy
Event summary
- TELUS Digital's poll of 1,000 U.S. adults found only 14% of AI responses changed when questioned, with just 25% of revised answers deemed more accurate.
- Research on four leading LLMs (GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5, Llama-4) showed follow-up prompts often fail to improve accuracy and can sometimes reduce it.
- 88% of poll respondents acknowledged AI errors, yet only 15% always fact-check AI-generated information.
- TELUS Digital emphasizes the need for high-quality training data and robust model evaluation to ensure AI reliability.
The big picture
TELUS Digital's findings highlight a critical challenge in AI deployment: ensuring models maintain accuracy under scrutiny. As enterprises scale AI, the reliance on post-deployment user fact-checking proves insufficient, underscoring the need for pre-deployment model rigor. The study's insights may push AI developers to prioritize robustness in model training and evaluation, particularly for high-stakes applications.
What we're watching
- Model Stability
- How AI models balance stability and adaptability when challenged will shape enterprise adoption strategies.
- Data Governance
- The emphasis on high-quality training data may accelerate demand for specialized AI data solutions.
- User Trust
- Whether enterprises can bridge the gap between AI capabilities and user expectations through better model design.
Our editorial coverage:
Related topics