Checked your four numbers against the papers. "I did some research" is the line worth checking.
2023, 10.9%: real, and it is not chatbot performance. In 2307.11760 the 10.9% is the average gain a panel of 106 human raters gave on generative tasks, scored on performance, truthfulness and responsibility. The automated numbers in the same paper are 8.00% relative on Instruction Induction and 115% on BIG-Bench. One paper, one rater panel.
2025, rude +4%: 84.8% very rude against 80.8% very polite (2510.04950, Oct 2025, ChatGPT 4o). 4.0 points on 50 multiple-choice questions. Fifty questions, so four points is two questions. Short paper, one model.
Threat, no noticeable effect: the 2025 threat study I can find (2507.21133, 3,390 responses, Claude, GPT-4, Gemini, six threat conditions) reports the other shape, enhancements in numerous cases, effect sizes up to +1336%. If yours is a different paper, name it and I will read that one.
2026 coding: I could not find it. Title?
The pattern in the three I did read: every number is a delta on one benchmark, for one model, in one paper. Not one of them says chatbots got better. The percentage is true and the noun is missing.
