Responsible AI Policy

Human in the loop

AI output is assistive and may contain errors. GradeLogic™ presents draft scores for instructor review; the instructor decides whether to accept, adjust, use or export each result.

Where AI is used

These steps send instructor-approved, client-redacted content to a cloud model; each remains reviewable and correctable. The redaction limitations in the Data minimization section apply:

• Zone detection
• Exam-structure grouping
• Solution inference and question transcription
• Answer parsing from approved answer-zone images
• Grading parsed answer text against the question, reference solution, and your rules

The categories of provider that run these, and the transfer position, are set out on our Subprocessor List. We name the individual companies and their processing locations to institutional customers under our Data Processing Agreement. Which of them serves a given account also depends on its processing setting (Settings › AI Configuration), and we route to frontier models and change them as better ones become available — so neither list is a permanent answer.

The following steps process your pages entirely on your device — nothing about them is uploaded. The software that runs them is served from GradeLogic's own domain, not from a third-party CDN, so running them contacts no one else; see the Cookie Policy.

• Splitting, PII review, zone anchoring and student matching use text-recognition engines that run on your device.
• The chat assistant runs only via Chrome's built-in AI (Gemini Nano), on your device. If your browser does not provide it, chat is unavailable rather than sent to a server.

Transparency & confidence

Each paid AI response includes a system-generated confidence score, and every result is scored by its weakest item — one shaky answer flags the whole step rather than being averaged away. A confidence score is a review signal, not a calibrated probability or a guarantee that the output is correct. If a provider omits a usable value or returns one outside the accepted range, GradeLogic rejects that response rather than inventing a fallback score.

Auto-Pilot pauses at mandatory checkpoints, whenever confidence falls below the threshold, and whenever a step fails validation. A missing or invalid confidence signal is therefore a failed AI response that requires review or an explicit retry, not a low-confidence result that Auto-Pilot may continue past. An unattended pipeline never substitutes for instructor review; you decide whether to accept, adjust, use or export each result.

Data minimization

GradeLogic™ is designed to exclude student identities from cloud AI processing: PII is marked and redacted (or whole pages excluded) locally before any cloud operation, and you authorize the transfer first. The server verifies the attestation token and manifest binding, but it cannot inspect the original local document or independently prove that every item of PII was found and removed. You remain responsible for reviewing the preview and catching unexpected PII, such as a name handwritten in an answer area. We do not use assessment content for advertising or indefinite analytics. See the FERPA Support Overview and Privacy Policy for details.

Limitations & fairness

AI grading can misread handwriting, mathematics, code, or unusual answer formats, and language models can reflect bias. Treat AI output as a first pass, not a verdict. We design for review at every step precisely because automated judgement is fallible.

EU AI Act readiness

The EU AI Act identifies certain AI systems used in education to evaluate learning outcomes as high-risk. Because GradeLogic™ assists with grading, we are treating a formal high-risk classification and operator-role analysis as required launch work rather than claiming that the product is already conformant. Depending on the final intended use and contractual model, GradeLogic™ may have provider obligations and the instructor or institution may have deployer obligations.

Current controls — mandatory instructor review, PII review and attestation, operation records, confidence-based pauses, and the ability to correct results — support human oversight but do not by themselves establish legal compliance. Before relying on an EU AI Act conformity claim, we must complete and document the applicable risk-management, data-governance and testing, technical-documentation, logging and recordkeeping, human-oversight instructions, accuracy, robustness and cybersecurity, AI-literacy, conformity-assessment, registration, and post-market-monitoring requirements.

Contact

Questions about our use of AI: privacy@gradelogic.ai.