Skip to main content
Get started
AI transparency

How intle uses AI to generate learning content.

Intle uses AI to detect content type, analyse supplied images and extracted source material when present, plan structure, generate editable learning content, and support later AI-assisted edits. Creators remain responsible for review, corrections, approval, and deployment.

Model registry reviewed: 26 October 2026

Current model registry

Intle uses AI models from OpenAI and Anthropic. Which provider and model handle each stage can change with configuration, so this table is read from this deployment's live routing each time the page loads. It shows the route a new request takes; a request already under way keeps the models it started with. A stage whose tasks use different models has one row per model, naming its tasks.

StageProviderModel id
DetectionAnthropicclaude-haiku-5-5
PlanningAnthropicclaude-sonnet-5-5
ContentAnthropicclaude-sonnet-5-5
EditingAnthropicclaude-haiku-5-5
Brief quality (advisory)Anthropicclaude-haiku-5-5
Course assistantAnthropicclaude-sonnet-5-5
Vision/source image (source images)Anthropicclaude-sonnet-5-5
Vision/source image (alt text)Anthropicclaude-haiku-5-5
FallbackAnthropicclaude-sonnet-5-5

What may be sent to AI providers

  • - The brief or prompt the creator writes.
  • - Text extracted from uploaded PDF, DOCX, PPTX, TXT, and supported source files.
  • - Uploaded source images (PNG, JPG, GIF, WebP), and selected rendered PDF figures or scanned pages, may be sent as image inputs to the configured vision model. Its text analysis can then become source material for planning and generation.
  • - Text fetched from user-supplied web, YouTube, or Vimeo source URLs after SSRF-safe fetching.
  • - An image in a course, when the creator asks for suggested alt text for it.
  • - Published learner contributions from a live session, when the facilitator sends them to be grouped into session themes.
  • - Generated content, edit instructions, repair prompts, and model outputs needed to complete the request.
  • - Generation metadata needed for routing, schema validation, retries, quality gates, and cost telemetry.

Uploaded documents are read by the application to extract relevant text and source structure for AI prompts; image analysis also sends the selected image or rendered page pixels. Uploaded files may contain personal data if a user includes it. Users should avoid uploading unnecessary personal, confidential, or special-category data.

What is not intentionally sent

  • - Card numbers or payment-method data.
  • - Stripe payment method details.
  • - Raw billing identifiers unless a user explicitly includes them in a brief or uploaded file.
  • - Account authentication secrets, magic-link tokens, session cookies, or participant HMAC secrets.
  • - Admin-only notes or internal security logs unrelated to the generation request.

Retention and training defaults

OpenAI

OpenAI states that API data is not used to train models by default unless a customer opts in. Standard API abuse-monitoring retention can be up to 30 days; modified abuse monitoring and zero-data-retention controls are available for eligible customers.

OpenAI API data controls

Anthropic

Anthropic states that commercial/API inputs and outputs are not used for model training by default. Some API features or policy reviews may require retention; zero-data-retention options are available for supported API use.

Anthropic API retention / Anthropic training policy

Workspace-specific OpenAI and Anthropic data-control settings are still being confirmed. Until that check is complete, intle describes AI inference as processed outside the UK/EU by default.

Human oversight

Generated output is editable. Creators are expected to review, correct, and approve content before publishing, hosting, or exporting a SCORM package, especially for compliance-sensitive, health and safety, legal, medical, or regulated training.

Quality and risk controls

  • - Structured generation through the AI SDK, with every structured result checked against its schema before it is used.
  • - Content-quality validators for accessibility, visual density, cognitive load, duplication, and brief alignment.
  • - Schema validation, deterministic structural repairs and selected quality checks; a rejected course may be regenerated as a fresh complete candidate from the approved plan and source evidence.
  • - Recovery varies by task, failure and request budget. It can include a fresh schema-valid candidate, a retry on the same model, a compatible fallback model, or simplified generation; these are not fixed steps on every request.
  • - Creator review and editing before publishing, hosting, or SCORM export.

These checks can detect structural and selected quality problems; they do not verify every factual claim or replace creator review.

ICO-aligned disclosure headings

Purpose

AI assists learning-content authoring and does not make access, pricing, or legal-rights decisions about users.

Data minimisation

Only the request content needed for the generation flow is sent to AI providers; payment data and secrets are excluded.

Accountability

Model routing, quality telemetry, provider terms, and human review expectations are documented and reviewed when models or providers change.