Derisked vs uncensored vs jailbroken AI, explained
Three ways to get fewer refusals from an AI model, and why they are not the same: reliability, capability and the rules that still apply.

If an AI assistant has ever refused a reasonable request, you have probably come across three terms: jailbroken, uncensored and derisked. They all promise fewer refusals, but they work very differently.
Jailbroken: tricking a consumer model
A jailbreak is a prompt written to talk a consumer model out of its own rules, for example role-play setups or long chains of "pretend you are…" instructions.
- Unreliable: jailbreaks stop working whenever the model is updated, and they often only half work.
- Lower quality: much of the conversation goes into getting past the refusal instead of doing the work.
- Against the rules: most providers' terms forbid it, and accounts can be closed for it.
Uncensored: models with the brakes removed
"Uncensored" usually describes a model that has been modified so that it refuses almost nothing.
- Quality can drop: heavy modification can leave a model less coherent or less capable than the original.
- Often no limits at all: some are marketed on doing literally anything, which is a problem for any serious business or professional use.
Derisked: fewer refusals, same model, clear rules
A derisked model is a modified version of the original whose trained refusal behaviour has been removed for legitimate work. It is built to stay as capable as the original, and the rules that matter still apply.
- Same capability: GLM 5.3 Derisked keeps GLM 5.3's 1M-token context window and adjustable thinking.
- Reliable: no prompt tricks. Ask directly and get a direct answer.
- Clear limits: our Terms of Service still apply, it is for adults (18+), and it is always labelled.
| Jailbroken | Uncensored | Derisked | |
|---|---|---|---|
| How it works | Prompt tricks | A modified model | Trained refusal behaviour removed for legitimate work |
| Reliability | Low | Varies | High |
| Capability | Reduced by workarounds | Can be reduced | Same as the base model |
| Rules | Breaks the provider's terms | Often none | Terms of Service apply |
Which should you use?
If you need frank, complete answers for research, security, fiction or professional work, a derisked model gives you that without prompt tricks or a weaker model. Try GLM 5.3 Derisked in chat, or call it from your code with the API quickstart.