1. Status of models and features
The website may describe research directions, experimental results, preview capabilities and features already made available. The specific status must be understood in light of each project’s official page; it must not be inferred merely from a name, video or interactive illustration that something is available for use.
If APIs, weight downloads, paid use or enterprise deployment are added in the future, access conditions, licenses, service boundaries and the corresponding documentation must be published separately. These disclosures do not create any standalone commitment to deliver a service.
2. Brand visuals and demonstration imagery
Images, videos and animations on the website may be used to present visual style, research directions or interaction concepts, and some of this material is produced with the assistance of artificial intelligence. Its purpose is determined by its labels, source notes and context taken together.
Unless a page expressly states the model version, input conditions and experimental procedure, brand imagery should not be regarded as direct measured evidence of any model capability. Work that has been edited, dubbed or post-processed should likewise not be treated as unprocessed raw output.
3. Limitations of generative systems
Generative models may produce factual errors, biased expression, inconsistent character behavior or unsupported explanations. Video and audio generation may also show identity changes, timing errors, abnormal movements or audio–video desynchronization. Natural, fluent expression does not guarantee accurate content.
Output that affects the rights and interests of others, is to be shared publicly or informs significant decisions should be reviewed appropriately before use. The system may not fully understand your intent; important constraints should be stated explicitly and checked to confirm that they have actually been followed.
4. Speed, quality and controllability metrics
Generation speed and latency depend on conditions such as hardware, precision, resolution, duration, sampling settings, batch size and network. Time to first frame, time to complete output and sustained generation throughput are different metrics; comparisons must use consistent definitions.
Conclusions about quality and controllability should be supported by adequate samples and evaluation methods. Selected success cases can show what is possible, but cannot on their own indicate average performance, failure rates or stability in complex environments.
5. Interpreting evaluations and leaderboards
Evaluation results depend on the task, samples, model version, prompts and scoring method. Automated judges may show preferences and instability, and human evaluation is also affected by the composition of the evaluators and how tasks are defined.
When using conclusions from SociaBench or other evaluations, refer to the methodology page and reproduction notes. A change in ranking on one task is not enough to infer which is better across all product scenarios. Comparative citations should state the applicable date and avoid mixing different versions of an evaluation.
6. Inputs, outputs and third-party rights
In any product that allows material to be submitted, you need to confirm that you have the right to process that input material, especially material involving real people’s likenesses or voices, confidential documents or personal information. Each product should explain how user content is processed and stored.
Use of generated output may also be subject to contracts, intellectual property, privacy and content rules. A model provider cannot, merely through a technical capability, replace the user’s own judgment about the actual context of publication or the necessary permissions.
7. Feedback and revisions
Useful feedback on an issue can include the project name, version, time, input conditions, expected result and actual behavior. Before submitting, please remove credentials, unrelated personal information and any material you do not have the right to disclose.
Once a formal feedback channel has been established, issues should be handled by the relevant product or maintenance team. The website’s disclosures should be revised on an ongoing basis as model status, evaluation methods or conditions of use change, with a clear date of last update.
Read this document together with the other documents in the Legal Center that apply to the same service.
All legal documents