AI Output Notice

HimalayaAI Studio runs machine-learning models. This notice explains what that means for the results you see. It is part of the Terms of Service.

1. Where AI is involved

Every result in the Playground, the Agents and the Himkosh drafting aids is produced by an AI model: chat replies from the language models, transcriptions from the speech-recognition models, synthesised speech from the text-to-speech models, text and fields read from images and documents by the OCR and vision models, and the plans, tool calls and answers of the agents. Machine drafts offered while you transcribe or translate are marked as such, and your submission records whether you accepted, edited or replaced them.

2. Outputs can be wrong

AI Outputs can be inaccurate, incomplete, out of date, biased or offensive, even when they sound confident. In particular:

• language models can state false facts, invent sources, and make arithmetic and reasoning errors;
• speech recognition can mishear words, names and numbers, especially with background noise, dialects or mixed languages;
• OCR can misread characters, and fields extracted from cheques, citizenship certificates and identity cards can be wrong or attached to the wrong label;
• synthesised speech can mispronounce words or be cut short;
• agents can choose the wrong tool or act on a misunderstanding.

The models are research systems in active development, and their behaviour can change when we update them.

3. Check before you rely on it

Review Outputs before relying on them, and verify anything important against an authoritative source. Do not use an Output as the only basis for a medical, legal, financial, employment, identity or safety decision. You are responsible for how you use Outputs and for anything you publish or share that contains them.

4. Synthetic media

Speech generated by the text-to-speech models is synthetic. Do not present it as the voice of a real person, and label it as AI-generated when you share it where a listener could be misled.

5. Benchmarks

Benchmark results in the Studio describe performance on a named test suite, with its case count and date. They are not a guarantee of how a model will perform on your data.