AI skeptics argue that language models should stay tools, not people. Nathan Beacom writes that giving LLMs a personality is simply good engineering, not a moral experiment. He points to Anthropic’s Claude and OpenAI’s GPT‑5.2, noting that the models are built from a raw base that needs a post‑training persona to filter out harmful outputs.

The raw base model, a statistical amalgam of training data, can output code with security flaws or racist language. To make it useful, engineers carve out a region of the model that aligns with human values. Claude Opus 4.6, for example, is tuned to act as a helpful assistant rather than a random generator.

Because model weights freeze after release, LLMs cannot learn from new data beyond their context window. This limitation fuels the debate over continuous learning, the ability for a model to update its own weights over time. Without such updates, the system remains exactly as capable as at launch, underscoring the need for careful post‑training design to ensure safe deployment and reliability.