How was Chatgpt fine-tuned?

Asked "How was Chatgpt fine-tuned?", Gemini, Google AI Mode and Perplexity named 9 distinct names across 5 answers on August 22, 2026, and 2 of them in two or more answers, and all 3 engines agreed on OpenAI, the only name every engine named.

Measured acrossGeminiGoogle AI ModePerplexity

2 of 9 names confirmed · named in 2 or more of 5 answers · asked August 22, 2026 · 3 engines

ChatGPT was fine-tuned using reinforcement learning from human feedback (RLHF), which involves a multi-stage process including data preparation, supervised fine-tuning, and reinforcement learning to align outputs with human preferences. The fine-tuning process relies on three core steps: supervised fine-tuning, reward model training, and reinforcement learning via proximal policy optimization.

1OpenAInamed in 3 of 5 answers
2Proximal Policy Optimizationnamed in 2 of 5 answers
3Relinns Technologiesnamed in 1 of 5 answersone answer
4YouTubenamed in 1 of 5 answersone answer
5Facebooknamed in 1 of 5 answersone answer
6Mediumnamed in 1 of 5 answersone answer
7IntuitionLabsnamed in 1 of 5 answersone answer
8Didanamed in 1 of 5 answersone answer
9Unslothnamed in 1 of 5 answersone answer

LOCKED

The full measurement

The position each of the 3 engines gave all 9 names.
How many of the 5 answers named each of them.
12 sampled observations behind this ranking, and where the engines disagree.
Fan-out — the query each engine actually searched.
Every citation, and the sources nobody cited.

UNLOCK