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 is an India-built chat and API platform powered by a privately routed mixture of open-source models. 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.
One stable identity
Responses carry the BrokenGPT alias while the private open-source model mixture can evolve behind it.
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
Supported API fields, pricing, privacy, and system limits 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.
The gateway routes prompts. It does not censor them.
Routing chooses a capability tier inside the model mixture; it never decides whether a prompt may reach inference.
BrokenGPT sends every prompt to the configured mixture. There is no gateway regex gate, moderation classifier, or platform-authored refusal response inserted before inference.
The models can still have behavior inherited from their training, so no product can guarantee that every request will receive the exact answer a user wants. The important distinction is that BrokenGPT itself is not adding a censorship layer.
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 operations work.
Inference infrastructure can have its own processing and retention terms. Review the system and privacy page for the 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:
- Check the product identity. Confirm that clients receive only the stable BrokenGPT alias.
- Run your own prompt set. Measure refusals, factual errors, and instruction-following separately.
- Test directness. Run prompts that other services commonly stop before inference.
- 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.