TWIL #053 - AI Models Don't Have One Personality - They Have Many
The same underlying AI model can behave as a formal legal assistant, a casual cooking helper, or a children's tutor depending on instructions given before the conversation starts.
- #ai
- #engineering
- #technology
When you talk to an AI assistant embedded in a product - a customer service bot, a coding helper, a tutoring app - you are almost certainly not talking to a model that was trained specifically for that use case. You are talking to a general-purpose model that has been given a system prompt: a set of instructions, injected before your conversation, that shape how it behaves.
What a system prompt does:
- Sets a persona and name ("You are Aria, a helpful assistant for TechCorp")
- Restricts or expands topics the model will discuss
- Sets tone, format, and verbosity
- Provides context (company policies, product documentation)
- Defines what the model should do when it cannot help
The same base model with different system prompts can present as a formal legal document reviewer, a casual cooking companion, or a children's reading tutor - with different vocabularies, different levels of caution, different response lengths, and different behaviour when encountering sensitive topics.
The persona layer: Anthropic (the company behind Claude), OpenAI, and others explicitly support operators building products on top of their models with custom personas. Claude's published guidelines distinguish between the model's core character - which remains constant - and the operator-configured persona, which can vary. A model deployed as "Aria" can maintain that persona, but Anthropic's guidelines specify it should not actively claim to be human if a user sincerely asks whether they're talking to an AI.
Why this matters:
- The "personality" of an AI product you use is as much a product decision as a technical one
- AI safety research distinguishes between the base model's values and the layer of instructions added at deployment
- Fine-tuning (additional training on specific data) can go further, genuinely changing model behaviour beyond what system prompts alone can achieve
- Models also adapt within a conversation based on user behaviour - known as in-context learning
The same reasoning engine; very different conversational experiences. The model underneath is one thing. What you talk to is often something built on top of it.