An AI disclosure is not an FWC filing control.

Three requirements create one release obligation

New Fair Work Commission requirements for generative AI took effect on October 20, 2026. If GenAI helps prepare a document for a Commission case, the party must disclose that use and explain how it was used. The document must be checked and corrected so its details are accurate and relevant. If it is a witness statement or declaration, the witness must ensure it rests on their own knowledge, reflects their own words, and is true to the best of their knowledge.

Professional representatives also must provide hyperlinks to case law cited in AI-assisted documents. That requirement turns a familiar hallucination risk into a testable release control: the cited authority must exist, the link must resolve, the decision must come from an authoritative source, and the passage must support the proposition for which it is cited. A plausible citation copied from a chatbot is not evidence.

The practical mistake is to treat the rule as a disclosure footer. A perfect disclosure does not cure a false fact, stale legislation, irrelevant authority, altered witness meaning, or confidential data sent to an unapproved tool. The obligations form one chain. The team must know what the tool did, verify what the document says, restore personal authorship where evidence is involved, and bind those checks to the exact file that will be lodged.

This workflow is designed for HR and employee-relations teams that prepare factual chronologies, responses, statements, annexures, submissions, or instructions for counsel. It is not legal advice and it does not replace the professional responsible for the matter. Its job is to make unchecked AI assistance visible before it becomes a procedural, evidentiary, privacy, or credibility problem.

Start with use, not with whether the model wrote the final words

The Commission describes the Guidance Note as applying when GenAI is used in writing, creating, modifying, or otherwise preparing an application or another document to be lodged. Teams should therefore avoid a narrow “AI authored this sentence” test. Summarizing interview notes, translating a statement, suggesting authorities, reorganizing a chronology, rewriting tone, extracting facts, or checking a draft may materially prepare the document even if a human later edits every line.

When the boundary is uncertain, record the use and ask the responsible professional. Hiding a marginal use creates more risk than reviewing it. The AI-use ledger should distinguish drafting, editing, summarization, translation, research, retrieval, and formatting so the final disclosure can be accurate rather than vague.

Freeze the candidate before substantive review. Use a document-management version, immutable PDF, or file hash. If a colleague continues editing while facts and authorities are checked, the final copy can diverge from the verified one without anyone noticing. Any material change after review should invalidate the release receipt and trigger targeted re-verification.

Separate preparation, verification, witness confirmation, and lodgement

approved source records + current official law
                       │
                       ▼
             AI-assisted candidate draft
                       │
              freeze version + hash
                       │
          ┌────────────┼─────────────┐
          ▼            ▼             ▼
     fact/evidence   authority      privacy and
       verifier      verifier       privilege review
          └────────────┼─────────────┘
                       ▼
             witness line review
          (when statement/declaration)
                       │
                       ▼
       disclosure + exception resolution
                       │
             named human release gate
                       │
                       ▼
          exact file lodged by authorized user
                       │
                       ▼
             portal receipt retained

Role separation matters. The person who prompts the tool should not be the only person verifying its factual and legal output. A witness should not receive a polished final statement with a yes/no request. They need time to read it line by line, correct wording, restore uncertainty, remove statements outside their knowledge, and confirm the result in their own voice.

Tool approval is a separate gate. Public models may retain prompts, use inputs for product improvement, or expose data through misconfiguration. Matter records can contain health, complaint, investigation, union, remuneration, disciplinary, settlement, and privileged material. Use the organization's approved environment, minimum necessary data, controlled identifiers, and explicit retention rules. If a tool's handling is uncertain, do not paste the material.

The release receipt binds all four lanes. It should list the document identity, GenAI-use disclosure, claim ledger, source register, legal review, confidentiality disposition, witness confirmation where required, hyperlink test, reviewer names, approval time, filing deadline, and lodgement owner. The lodged copy and portal receipt complete the record.

Turn polished prose into atomic claims

A reviewer cannot verify “the document” as one object. Break material content into atomic units: dates, identities, events, amounts, communications, documents, evidence assertions, legislative provisions, cases, propositions, quotes, extracts, and requested outcomes. Give every claim a stable ID and record its location in the frozen candidate.

Verification is proposition-specific. A real case can still be the wrong case. A correct section number can be stale, from the wrong compilation date, or applied outside its defined scope. A quote can be exact but misleading after an omission. The ledger should therefore retain the exact passage, context, jurisdiction, currency, and the reviewer's explanation of why it supports the claim.

Use primary and official sources whenever possible: the current Federal Register compilation for legislation, official Commission or court databases for decisions, and the Commission's own procedural guidance. Commentary can help locate an issue, but it should not silently replace the authority cited in a lodged document.

Professional representatives should test every case-law hyperlink in the final rendering, not only in the authoring file. Confirm that it opens the intended authority without a private session, unstable search token, or organization-only permission. Record the successful URL and check time in the authority ledger.

A witness statement is not a style-transfer task

The Commission's third requirement applies specifically to witness statements and declarations. The person must check that the document is based on their own knowledge, reflects their own words, and is true to the best of their knowledge. A model can improve grammar while quietly changing certainty, chronology, causation, or emphasis. “I understood” can become “I was told.” “Around June” can become a precise date. “I felt pressured” can become an assertion about another person's intention.

Give the witness a comparison that preserves their original notes or draft, the proposed sentence, every factual source, and a place to rewrite the sentence themselves. Ask whether the statement is within their personal knowledge, whether any part is inference or hearsay, whether uncertainty was removed, and whether the words sound like them. The witness's correction should be recorded as their change, not regenerated through another prompt.

Do not use stylometry, “authenticity scores,” or model confidence to certify voice. The control is procedural: a real person reviews the exact final version and confirms it. For translated statements, retain the original-language material, translator or approved process, and confirmation that the meaning was preserved.

Legal and employee-relations reviewers still decide evidentiary and procedural questions. The AI pack intentionally produces an exception queue, not an admissibility verdict. It cannot decide credibility, privilege, relevance, hearsay, or whether a concession should be made.

Worked example: an AI-assisted unfair dismissal response

An employee-relations team uses an approved model to summarize the investigation file and draft a chronology for a response. The model states that a meeting occurred on 12 August, describes an email as a refusal, cites a Fair Work Act provision, and suggests two decisions. It also rewrites a manager's witness statement to sound more formal.

The scope gate records summarization, drafting, legal research, and witness editing as four distinct uses. The candidate is frozen. The fact ledger confirms the meeting date from the calendar and minutes, but the email does not say “refused”; the statement is corrected. The legislation reviewer opens the current official compilation and determines that the suggested provision does not support the proposition. One suggested decision does not exist, and the other is real but addresses a different issue. Both are removed after professional review.

The manager compares the AI-edited statement with their notes. They restore “I do not recall the exact words,” change a precise time back to “after lunch,” and remove a sentence about the employee's motivation because it was not within their knowledge. The reviewer confirms the final wording, attaches working links to the authorities that remain, and updates the AI-use disclosure to state exactly how summarization and drafting were used.

The release receipt binds the final PDF hash, disclosure, claim ledger, authority check, witness confirmation, privacy disposition, reviewer, and lodgement authority. A second authorized person uploads that exact file and stores the Commission portal receipt. If the PDF hash differs, the record stays open and the filing copy must be reconciled.

Failure modes that a disclosure cannot cure

The strongest red flag is a document that reads more confidently than the evidence. AI commonly removes hedging, fills chronology gaps, and turns disputed interpretations into declarative sentences. Reviewers should search for precise dates without sources, quotations without locations, legal propositions without official authority, statements about another person's intention, and conclusions that appear nowhere in the factual bundle.

Launch as a shadow review, not an automated filing lane

1. Map the current FWC Guidance Note, official case-preparation resources, organization legal policy, privilege, privacy, retention, and matter-management rules.
2. Select closed, de-identified matters with known facts, authorities, statements, and filing outcomes.
3. Test the workflow on drafting, summarization, research, translation, and witness-editing examples separately.
4. Seed fabricated cases, real-but-irrelevant authorities, stale legislation, wrong jurisdictions, altered quotes, broken hyperlinks, and changed witness certainty.
5. Require the tool to abstain when source material or authority is missing. Measure false verification as a critical defect.
6. Test approved-tool boundaries with restricted-data markers. Confirm no prompt or output leaks into logs, analytics, or unapproved storage.
7. Run witness review with real users on synthetic statements. Check whether the interface makes original wording and AI changes obvious.
8. Bind approvals to an immutable file and prove that a changed filing copy cannot reuse the old receipt.
9. Keep lodgement manual. The model and workflow service should have no Commission portal credential.
10. Review a sample after lodgement and retain corrections, withdrawals, or Commission feedback as controlled learning evidence.

Useful measures include material claims extracted, claims with authoritative evidence, fabricated authorities caught, irrelevant authorities caught, quote differences found, witness changes after review, broken links caught, restricted-data blocks, reviewer time, post-release corrections, and filings stopped by missing approval. “Drafting time saved” is secondary to false statements prevented.

FAQ

Do the FWC requirements apply only when AI writes the whole document?

No. The Guidance Note is framed around GenAI being used to write, create, modify, or otherwise prepare a document. Record summarization, research, translation, editing, and drafting uses for professional review.

Is disclosure enough?

No. The document must be checked and corrected for accuracy and relevance. Witness statements and declarations also require confirmation of personal knowledge, personal words, and truth to the best of the witness's knowledge.

Can an AI tool verify a legal citation?

It can help create a checklist or extract a candidate citation. Verification requires opening an authoritative source, confirming identity and currency, and checking that the actual passage supports the exact proposition.

Should the workflow store prompts?

Retain enough evidence to explain use without spreading restricted content. Controlled prompt and output IDs may be safer than copying full matter material into a general audit report. Follow privilege, privacy, security, and retention policy.

Is this legal advice?

No. It is an operational release-control pattern. Qualified professionals must determine legal strategy, evidence, relevance, privilege, confidentiality, disclosure wording, and authority to lodge.

Sources and further reading

Sources were checked online on October 26, 2026. The Fair Work Commission's current materials control this workflow. Comparative court guidance illustrates verification and confidentiality risks but does not replace the Commission's requirements or legal advice for a matter.

• Fair Work Commission: New GenAI requirements in Commission cases - effective-date notice and the three requirements.
• Fair Work Commission: Use of AI in Commission cases - President's statement, final Guidance Note, research report, and resources.
• Fair Work Commission: Using AI to help you with your case - official explanation, setup prompt, risks, and responsibilities.
• Fair Work Commission: GenAI Guidance Note exposure draft - published scope and verification mechanics; use the final linked Guidance Note for current compliance.
• Federal Register of Legislation: Fair Work Act 2009 - current official compilation and downloads.
• Supreme Court of NSW: Generative Artificial Intelligence resources - official practice note, review material, and professional resources.
• Supreme Court of NSW Practice Note SC GEN 23 - comparative guidance on GenAI use in proceedings.
• Guidelines for NSW judges on GenAI - verification, confidentiality, stale law, fabricated citations, and red flags.
• Fair Work Commission privacy information - privacy handling and official policy links.
• Related: HR AI output release gate.
• Related: employee-data incident response workflow.
• Related: AI delegation and accountability workflow.