OpenAI shapes ChatGPT's personality by showing the model examples of how it should respond in different cases, from which it learns its overall personality.

True: OpenAI documentation or credible technical accounts as of 2025-08-09 show ChatGPT's personality is shaped mainly by training on examples of desired responses (e.g. supervised fine-tuning demonstrations, or RLHF built on example comparisons), from which the model generalises its overall behaviour. Mostly true: example-based training is a main method, but other methods described in the sources (e.g. system prompts, the Model Spec, reward models) also play a significant part that the claim simplifies. Mixed: example-based training is one of several roughly equal methods. Mostly false: personality is mainly set by other means (e.g. system prompts or rules) and examples play only a minor part. False: evidence shows OpenAI does not use example-based training to shape personality.

Set before the research began.

1. E1 Expanding on what we missed with sycophancy OpenAI · 2 May 2025

   “To post-train models, we take a pre-trained base model, do supervised fine-tuning on a broad set of ideal responses written by humans or existing models, and then run reinforcement learning with reward signals from a variety of sources.”

2. E2 Expanding on what we missed with sycophancy OpenAI · 2 May 2025

   “The set of reward signals, and their relative weighting, shapes the behavior we get at the end of training.”

3. E3 Sycophancy in GPT-4o: What happened and what we’re doing about it OpenAI · 29 April 2025

   “Refining core training techniques and system prompts to explicitly steer the model away from sycophancy.”

4. E4 Training language models to follow instructions with human feedback arXiv (OpenAI authors, InstructGPT paper)

   “Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning.”

OpenAI's May 2025 post says its post-training includes supervised fine-tuning on a broad set of ideal responses (E1), and the InstructGPT paper describes fine-tuning on labeler demonstrations of the desired behaviour (E4). Both support the claim that the model learns from examples of how it should respond. OpenAI's May 2025 post also says that reward signals from reinforcement learning, and how they are weighted, shape the model's behaviour at the end of training (E2), and an April 2025 OpenAI post says system prompts are being refined to steer the model away from sycophancy (E3). Example-based training is therefore a main method, but the claim leaves out other methods that play a significant part.