شفافية الذكاء الاصطناعي

Where AI is used

OncoRadar is a conversational research assistant. When you send a message, a research language model plans literature searches that run through OncoRadar's connectors; every retrieved record is stored in the catalog; the model then writes a source-grounded answer that cites study ids, passes a verification step and states its limitations. The model never searches on its own outside the connectors and must not invent study ids. AI-assisted output should be understood as generated research support describing research patterns, not an authoritative clinical conclusion.

Technology roles

The current implementation combines the roles below. Providers are named only in the Subprocessor Notice; the roles are described neutrally here. Availability and exact routing may change only after the controlling notices are reviewed and updated.

• Embedding model: Produces semantic vectors for retrieval over indexed literature and de-identified text.
• Catalog and retrieval engine: Stores every retrieved literature record and supports lexical, semantic and hybrid retrieval.
• Research model: A research language model hosted in our cloud provider that plans literature searches through OncoRadar's connectors, reads de-identified text and images, and writes source-grounded answers with cited study ids and stated limitations.
• Decision model: A calibrated decision model that ranks evidence, verifies statements and classifies documents; it receives only de-identified text or literature.
• OCR and identifier detection services: Extract text from documents and images, detect identifiers and support rule-based redaction and burned-in text masking.
• Medical imaging store: Holds de-identified DICOM imaging and normalized metadata for the viewer and for findings.

Inputs and outputs

A request may include your chat text (redacted for identifiers before storage), de-identified text and de-identified images or imaging frames from your uploads, selected filters, retrieved literature passages and system instructions. Original uploaded files are never sent to a model. The calibrated decision model receives only de-identified text or literature, never images or original files. Outputs are retained as conversation messages, findings and operational records where attribution is implemented. Do not submit credentials, confidential third-party information or material you cannot lawfully process.

Verification pass

After drafting, generated statements are checked against the retrieved records they cite; unsupported statements are flagged or removed and the answer states its limitations. The verification pass reduces but does not eliminate errors, and a passing check is not evidence of clinical validity.

Limitations and review

Models can generate inaccurate, biased, outdated, incomplete or invented content. Retrieval can miss relevant evidence. Citations may not support a generated statement. Automated de-identification may leave an identifier in place or mask legitimate content. Human review must include opening sources, validating provenance and licenses, checking retractions and corrections, and applying domain expertise.

User controls and feedback

You choose whether to send a message, upload a file, invoke generative functions and use an output. Quotas may limit generation and uploads. You can delete conversations and uploads, and export and deletion requests cover attributable account data, subject to the Privacy Policy and backup rotation. Report materially unsafe or misleading behavior to the abuse contact and vulnerabilities to the security contact.

No training promise and dataset consent

OncoRadar does not use account, conversation or upload content to train a proprietary foundation model. Provider handling is governed by configured service terms. If this practice changes, the controlling notices will be updated before using content for materially different training purposes.

Contributing de-identified material to the OncoRadar research dataset is a separate, explicit, per-item opt-in recorded under the Dataset Contribution Terms with its own consent record. It is not part of accepting this notice, the dataset is shared only de-identified and with that consent, and a contribution can be withdrawn.