Why Aviation AI Must Cite Approved Sources

A procedure draft based on a blog post is not enough if the controlled MOE says something different.

The wording may be tidy. The logic may sound reasonable. The document may even look like something a quality department could issue. But if it is not tied to the current approved manual, the right regulation, the relevant form and the evidence actually sampled, it is only a writing exercise.

Aviation teams do not work from general knowledge. A Part-145 organisation works from its MOE, work orders, CRS records, competence records and approved procedures. A CAMO works from its CAME, AMP, reliability process, airworthiness review procedure and contracted arrangements. An operator works from its OM, MEL procedures, occurrence reporting process, SMS material and controlled forms. Those documents are not background reading. They are the operating basis.

AI is useful when it helps people work faster inside that basis. It becomes risky when it writes around it.

The source is part of the answer

In aviation, the first question is rarely "does this read well?" It is "what is it based on?"

A finding without a requirement is weak. A corrective action plan that ignores the company procedure may fix the wrong thing. A MEL deferral summary that does not show the applicable MEL item, operational limitations and maintenance entry can mislead the reviewer. A training gap note that skips the approved training programme may look helpful and still miss the required standard.

The reviewer needs to see the chain:

• the applicable rule or contract requirement;
• the approved manual paragraph, such as MOE, CAME or OM;
• the controlled form, checklist or process used;
• the sampled record, such as a CRS, work pack, audit note or occurrence report;
• the revision status of the material;
• the parts that are still assumptions.

When that chain is visible, a draft can be challenged. When it is hidden, the reviewer is left trusting fluent text.

Not all sources weigh the same

Ask whether a Part-145 procedure is enough, whether an AMC applies or whether an audit finding is supported, and the answer is only half the job. You also need to know what kind of material it rests on. Aviation answers often need to separate:

• the binding regulation;
• the AMC, which is one acceptable way to comply;
• the GM, which explains or guides;
• the organisation's approved exposition or procedure;
• the user's own interpretation;
• the recommended next step.

An answer that blends these may look helpful and still create risk. Guidance presented as a hard rule, or a company habit presented as a requirement, is how clean-looking mistakes start. For how the layers of EASA material relate to each other, see how to read EASA regulations.

This is also why a citation is not the same as a link. In casual research, a link is enough. In aviation work, the citation has to say what type of source it is and which amendment it reflects. General tools can list source names or URLs and still fail to classify them, leaving the user to work out whether the answer rests on hard law, AMC, GM, explanatory content or general reasoning. The user should never have to guess. How to check a cited answer against the rule text, step by step, is in how to check an AI answer against the regulation.

Public information is not company authority

Public regulations and guidance help explain the framework and give useful language. They do not describe how a specific organisation complies.

A generic answer may say that maintenance release requires authorised personnel and records. Fine. The real question inside a Part-145 organisation is more specific: what does the MOE say about CRS issue, certifying staff privileges, independent inspections, shift handover and incomplete work? Which form is used? Where is the record kept? Who checks it during internal audit?

The same applies to CAMO and operator work. A general summary of continuing airworthiness will not know the current CAME paragraph, the local AMP escalation route or the contracted maintenance interface. A general safety answer will not know the operator's occurrence classification process, safety review board terms of reference or risk acceptance limits.

Public text can support a draft. It should never be allowed to overrule the controlled system.

A worked audit-finding example

Take a Part-145 internal audit. The auditor samples three closed work orders and notices that one CRS reference is incomplete. The rough note reads: "CRS detail missing on WO-EXAMPLE-0142, checked against MOE 2.16, talked to certifying staff."

A generic tool might turn that into a polished finding about poor release discipline. That may be too strong, too broad or unsupported by the sample. A better aviation assistant asks for the MOE paragraph, the sampled record, the audit checklist criterion and any previous similar findings. It separates the confirmed evidence from the possible interpretation.

A review-ready draft might say: "Sample WO-EXAMPLE-0142 did not include the required CRS reference described in MOE 2.16, subject to confirmation of the current MOE revision and record-retention procedure." And it flags: "Classification not assigned. Reviewer to confirm extent, repeat finding history and impact."

That is useful. It gives the quality manager something to check. It does not pretend the tool has classified the finding or approved the report.

The same discipline carries through the rest of the system. That one weak point may drive a corrective action. The action may change an MOE paragraph. The change may need training, which affects competence records, which affect who can certify. A corrective action draft should not just say "brief staff". It should name the affected procedure, the training record, the owner, the due date, the effectiveness check and where the record lives. Working from approved documents is what keeps the draft inside the organisation's actual system.

Revision status and version awareness

Old sources create clean mistakes.

A CAME paragraph may have changed after an authority finding. An OM procedure may have been updated after a safety action. A Part-145 form may have moved to a new revision. On the regulatory side, rules, AMC and GM are amended too, so a saved summary or an old checklist can be right today and misleading next year. If the assistant uses the wrong revision, the output is wrong even when the writing is excellent.

Professional work needs simple status signals. The draft should show whether the material came from a current controlled document, a user upload with unknown status, a public reference or a pasted note. If the version is missing, it should say so: "Manual revision not confirmed." "Procedure status unknown." "Based on user-provided text only."

A general chat tool can summarise an amendment if you give it the text. It is not your change-management system. How versioned regulation tools handle this is covered in EASA version-aware regulation tools.

Missing sources should slow the answer down

Sometimes the most professional response is not a complete answer. It is a request for the missing basis.

• Asked for a finding, the assistant should ask for the requirement and the sampled evidence.
• Asked for a MEL deferral note, it should ask for the MEL item, defect text, operational limitation and maintenance status.
• Asked for an occurrence report summary, it should ask whether the report is draft, submitted, investigated or closed.
• Asked for a procedure change, it should ask for the current controlled paragraph and the approval route.

The tool can still help while it waits. It can prepare a skeleton, list missing records, draft reviewer questions or organise the facts it has. It should not dress uncertainty as authority.

Aviation professionals already work this way. They ask for the manual, the form, the evidence and the version before accepting the conclusion. AI should follow the same habit.

Approved sources do not approve anything

Using the right material does not make the assistant the accountable person.

The tool may retrieve references, compare paragraphs, draft wording, flag missing records and prepare the review pack. The final acts stay with people: approving a procedure, classifying a finding, accepting a corrective action, accepting safety risk, issuing an authority response, releasing a record or closing an action.

That line should be visible in the output. "Draft for review" is not a weak label. "Classification to be confirmed" is not a failure. "Check current approved manual before use" is exactly the kind of warning aviation work needs. Good source control makes the human review easier. It does not remove it. Where that review sits for each type of work is covered in the human review gate.

How Avioverse handles sources

In Avioverse, source handling is part of how the assistant works, not a note at the bottom of an answer.

Metis, the assistant in the workbench, looks rules up rather than recalling them. It searches the stored text of its library, which includes EASA regulations, Airworthiness Directives and Safety Information Bulletins, and answers from what it found. Each conversation shows which regulation library is in scope, EASA by default, with FAA available. Grounded answers carry numbered markers and a single Sources list. Selecting a marker shows the passage actually retrieved, and the popover says plainly when the text is a paraphrase rather than a quotation. When a regulation is quoted verbatim, the block carries the amendment reference and a source-type label such as Regulation, AMC or GM.

Library content is versioned. Each item keeps its status and effective dates, and older versions are kept rather than overwritten, so a citation continues to point at the text it quoted.

When you want the sources without a generated answer, a separate search mode returns attributed passages with the version, date and status information available for each, and lets you compare them before you ask for anything to be written.

Each answer also shows its check result: Checked against sources, Review suggested or Not checked. The chat carries a standing reminder: "You're chatting with an AI assistant. Answers can contain errors — verify against official sources before acting on them." The organisation's own MOE, CAME or OM remains the controlling document for how it complies. Metis can work with an extract you provide, but it does not replace the approved exposition or the accountable person. What Metis is explains how its answers are checked.

If an AI output cannot show the manual paragraph, the record sampled, the revision status and the open assumptions, it is not ready for aviation work. Use AI for the desk work: collect the notes, compare the controlled references, flag the missing evidence and prepare wording for review. Do not let it turn public knowledge into company authority.