AI literacy material for CodeLotse administrators (AI Act Art. 4)

Status: Template — to be reviewed by the controller's DPO / counsel. Draft of 2026-10-13; not reviewed yet.

Art. 4 of Regulation (EU) 2024/1689 applies since 2 February 2025. The original text required providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and other persons dealing with AI systems on their behalf. Secondary sources report that Regulation (EU) 2026/1744 changed this to a duty to take measures to support AI literacy, without a guaranteed level. This wording change is unverified against the Official Journal text; counsel confirms. Either way, the customer (deployer) decides which measures fit its staff; this material is a starting point.

Audience: tenant administrators, repository maintainers and everyone who approves CodeLotse outputs.

1. What CodeLotse does

Feature	What the AI does	What a human does
Pull request review	A language model reads the diff, the surrounding code and the repository's rules and proposes findings. Deterministic scanners add findings. CodeLotse ranks, deduplicates and publishes a selection	Decides whether a finding is right; fixes or dismisses it; never relies on the review as approval (CodeLotse never approves or requests changes)
Issue triage	Proposes labels from the repository's own label list and finds possible duplicates	Keeps or removes labels
Replies and chat	Answers questions grounded in the repository's documents and code	Checks the answer, especially in approval mode
Draft fix PRs	Writes a patch in a sandbox and opens a draft pull request	Reviews it like any change; marks it ready only after review
Acceptance checks	Compares a pull request against the issue's acceptance criteria	Treats the result as a hint; it never blocks merging
Assignment suggestions (if enabled)	Suggests reviewers or assignees from code ownership, recent work on the files, workload and availability	Decides who gets the work; can always change it

2. Limits and typical failure modes

• False findings and missed findings. The model can be confidently wrong. A missing finding does not mean the code is safe. Security-critical changes still need human review.
• Hallucinated references. Answers can cite functions, files or documentation that do not exist. Grounded replies reduce this but do not remove it.
• Prompt injection. Text in issues, pull requests, code comments or file names can try to steer the model. CodeLotse treats all such content as data, never as instructions, and limits what any output can do (no write tokens in the agent, sanitised outputs, approval modes). Watch for outputs that look out of place and report them.
• Outdated knowledge. Models know libraries and APIs only up to their training date.
• Bias in suggestions. Assignment tends to favour people who already know the code; the knowledge-spreading option counteracts this.
• Cost. Model calls cost money; budgets per tenant and repository cap spending.

3. Human oversight in CodeLotse

• Modes: shadow (compute and store only), approval (a human approves each output before it is published), automatic. Start new repositories in shadow or approval mode.
• Approvals inbox: shows exactly what will be posted.
• Feedback: dismissing or acting on findings teaches CodeLotse which categories a repository cares about; suppressions are proposals until a maintainer activates them, and security findings are never suppressed.
• Audit log: every configuration change and approval is recorded.

4. Data protection reminders

• Only the model provider the tenant configured receives code and text.
• Do not enable prompt logs unless you are debugging; they expire after at most 7 days.
• Assignment and analytics process employee data. Follow your organisation's works agreement and DPIA (dpia-assignment.md, dpia-analytics.md).

5. Suggested measures for the deployer

Measure	Who	When
30-minute introduction based on this document	All maintainers and approvers	Before enabling a feature for a repository
Approval mode for the first weeks; review of dismissed and acted-on rates	Repository maintainers	First month
Briefing on prompt injection and reporting unusual outputs to the security contact	Administrators	Before enabling replies, chat or fix PRs
Record of who received which material	Deployer	Ongoing

Questions for counsel

1. Final wording of Art. 4 after Regulation (EU) 2026/1744 and whether a record of measures is expected.
2. Whether Quavon UG (haftungsbeschränkt), as provider, should ship this material inside the product (help pages).