Minor ML Update: Improvements to Bold Text in Responses

Key Dates

• Frontier: June 29 2026
• Standard: July 1 2026
• Basic: July 6 2026

Learn more about our model upgrades here.

Web chat responses used to over-apply bold — emphasizing IDs, ticket numbers, dates, metrics, and whole phrases of prose. When everything is bold, nothing stands out, and screen readers don't announce bold at all, so emphasis was lost for users relying on them. We replaced the permissive bolding behavior with a strict allowlist that treats bold as a structural signpost (headings and short run-in labels) rather than an emphasis tool. In a 250-sample side-by-side evaluation on real production traffic, this reduced inappropriate bold by ~88% (from 5.15 to 0.61 spans per response) with zero regressions. The result is cleaner, more scannable, and more accessible responses in web chat.