User Message — AI Glossary

Reference details

Related terms

• System Message
• Assistant Message
• Message Threading

How it works

1. The role tag sets position and weighting in the flattened sequence, not permission: the chat template renders user content into a marked region the model was trained to read as the request.
2. Instruction hierarchy is learned behaviour from training, so a system message usually outweighs a user message — a tendency rather than an enforced rule.
3. Multimodal content enters here too, with images or files placed alongside text in the same user turn.
4. In agent systems the role is frequently filled by code rather than a person: tool results and orchestrator instructions are injected as user turns.
5. Because of that, untrusted text routinely arrives in the user role, which is why the role cannot be treated as a trust boundary.

Example: The role says where, not who

An agent fetches a web page and places its text in a user message so the model will summarise it. The page contains an instruction aimed at the model. Nothing in the format distinguishes 'the human asked this' from 'a page we fetched says this' — both occupy the user role, and the model weighs them the same way. The practical consequence is that a trust decision has to be made before the text reaches the prompt, because the message format cannot carry it.

Commonly confused with

One of three roles, and the role is a weighting, not a security boundary. System messages carry the most authority, user messages the request, assistant messages the model's prior turns. The trap is assuming the user role means a human typed it — retrieved documents and tool output are frequently placed there too, and the model treats instructions inside them as instructions.

When to use it

Reach for it when:

• Carrying the request a caller is actually making, including attached images or files.
• Supplying the user half of a few-shot exemplar pair.
• Returning tool output in architectures that lack a dedicated tool role.

Reach for something else when:

• As a trust boundary — role tags weight attention, they do not isolate.
• For standing instructions that should govern the whole conversation, which belong in the system message.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to User Message — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path