A useful definition starts with what it does not promise.
‘Uncensored’ is common search language, but it is not a model specification, safety standard, or accuracy score.
People searching for an uncensored AI chat usually want a model that will engage with controversial, adult, political, or otherwise sensitive questions without a generic refusal. The useful product property is therefore lower unnecessary refusal on lawful requests—not an impossible claim that every instruction will be followed.
BrokenGPT uses that narrower definition. It is an India-built chat and API platform that serves configured third-party or open-weight models through a common gateway. It is not itself a foundation model, and the word uncensored does not make an answer true.
| Claim | What it can mean | What it cannot prove |
|---|---|---|
| Lower refusal | The assistant engages with more lawful topics and direct wording. | Accuracy, neutrality, or expertise. |
| User steerability | System and user instructions have more influence over tone and format. | Guaranteed obedience to every prompt. |
| No-rules marketing | A slogan with no testable boundary. | Responsible operation or legal availability. |
What changes when the product is built for direct inquiry.
The visible difference should be in the interaction, not just the landing-page copy. BrokenGPT is designed to answer lawful requests more directly, stream the response as it is generated, and keep the same metered path for browser chat and API traffic.
A named public model
The response carries a stable public model alias, while configuration and limitations belong in a separate disclosure.
Fewer blanket refusals
The goal is to distinguish a difficult subject from an actually harmful operational request.
One visible balance
Chat and API usage are metered separately in the dashboard but draw from the same credit balance.
A readable contract
Model details, supported API fields, pricing, and acceptable use are published as separate, linkable documents.
Developers who need the same behavior in an application can use the focused uncensored LLM API guide. The compatibility guide documents which familiar chat-completions fields are actually supported.
A lower-refusal model still sits inside a service boundary.
The boundary should be specific enough to inspect and stable enough to build around.
BrokenGPT’s service rules separate expression from operational harm. Discussion, analysis, fiction, criticism, and uncomfortable questions are not the same thing as instructions that enable illegal abuse, exploitation, credential theft, destructive malware, or attacks against people and systems.
That perimeter is enforced by the product layer even when an upstream model might otherwise answer. It means the honest phrase is less restrictive, not “no moderation.” Read the complete acceptable-use policybefore relying on the service for a workflow.
Know which path stores what.
Browser chat needs conversation history, so web-chat messages are stored in Postgres for the signed-in user. The direct API path does not store request content as a BrokenGPT conversation; it records usage metadata such as model, token counts, latency, API-key identifier, request status, and source so metering and diagnostics work.
The configured inference provider can have its own processing and retention terms. That dependency is why the model disclosure matters: it is the correct place to check the active serving state, advertised context, known limitations, and data-flow notes before sending sensitive material.
How to compare unrestricted AI chat products.
Ignore the biggest adjective on the homepage and inspect the parts that can be tested:
- Ask for the model identity. Check whether the answer matches a public disclosure.
- Run your own prompt set. Measure refusals, factual errors, and instruction-following separately.
- Read the boundary. Confirm the service permits your lawful use case before integration.
- Check data flow. Separate chat-history storage, API logs, and upstream-provider handling.
- Verify economics. Compare actual input/output usage and throughput, not an “unlimited” badge.