Are AI-Generated Web Pages Even Usable? Tencent Launches an Evaluation That Automates Web Interaction for Testing

Beating AI Insight News Brief: Tencent Hunyuan has partnered with Tsinghua University and Peking University to launch IWC-Bench, a benchmark specifically designed to evaluate the real usability of AI-generated web pages. Most existing web generation assessments focus on code or screenshots. A page may look visually appealing and have functional code, but issues like unclickable buttons, failed form submissions, or errors when switching pages are hard to detect via static checks. IWC-Bench operates more like a real human acceptance test: it uses an agent to actually open web pages, click buttons, input content, switch pages, and verify which functions work in practice. It evaluates three aspects separately: page aesthetics, operational smoothness, and whether required user functions are implemented. The benchmark includes 369 real user needs and 5088 acceptance criteria, with 17 models generating a total of 6273 web applications. In an additional comparison of 197 manual reviews, IWC-Bench aligned with human choices in 168 cases, achieving an 85.3% consistency rate. The results confirm that a web page "looking good" is not equivalent to being "actually usable". GPT-5.6-Sol has the highest page aesthetics score but ranks only sixth in usability; for individual web pages, visual appeal is largely unrelated to usability, with a correlation coefficient of just 0.36 between the two.