Version 3 coming soon·Faster, leaner, cheaper and more powerful.

UnitBuddy
Demo

Disputes

Using AI for Strata Disputes: NCAT, VCAT and the Risk of Fake Law

Australians are using ChatGPT and other AI tools to prepare tribunal cases. The help is cheap and immediate, but invented authorities, altered evidence, privacy breaches and bloated submissions can damage a sound strata claim.

· 19 min read

On this pageHow far the evidence goes
  1. How far the evidence goes
  2. Why strata disputes are especially tempting AI territory
  3. What has already gone wrong in Australian tribunals
  4. NCAT: AI-assisted submissions were given less weight
  5. NCAT: “improving” a statement can alter the evidence
  6. VCAT: a Gemini answer about fence heights had no evidentiary value
  7. QCAT: checking hallucinations wastes public resources
  8. The NSW Court of Appeal: more words, less help
  9. The rules are not the same at NCAT, VCAT and QCAT
  10. Seven ways a chatbot can damage a sound strata case
  11. 1. It chooses the wrong legal doorway
  12. 2. It supplies law-shaped text instead of law
  13. 3. It blurs evidence and argument
  14. 4. It exposes other people's information
  15. 5. It buries the useful point
  16. 6. It escalates the dispute before the hearing
  17. 7. It creates false confidence
  18. A practical traffic-light guide
  19. How to verify an AI-assisted tribunal document
  20. If you have already filed something AI helped write
  21. Use AI as an organiser, not as the source of truth
  22. Sources and further reading

The dangerous ChatGPT answer is the polished one. It names an Act that sounds right, supplies a section number, explains a legal test and cites a case with a judge, year and paragraph. An owner who has spent months chasing a leak, a by-law breach or access to records may copy that answer into a letter or tribunal submission because it finally sounds like the argument they have been trying to make.

But the Act may be from the wrong state, the section may have been amended, and the case may say something else or never have existed at all.

Australian courts and tribunals are already dealing with the result. In 2025, UNSW's Centre for the Future of the Legal Profession reviewed 87 Australian proceedings involving recorded GenAI use or misuse between January 2023 and September 2025. Sixty-eight involved self-represented litigants. The cases appeared across 23 courts, tribunals and commissions, and the number accelerated during 2025. The research did not isolate strata matters, but it found the cases clustered in forums where self-representation is common. Read the UNSW report summary.

By November 2025, High Court Chief Justice Stephen Gageler was warning that judges had become “human filters” for machine-generated or machine-enhanced arguments and described the prevalence of AI use in litigation as unsustainable. His concern was not that every use of AI is wrong, but that the checking burden moves from the person who generated the material to the opponent, the tribunal member and the public justice system. .

How far the evidence goes

There is no published national count of Australians asking AI about strata law. The public material reviewed for this article also does not provide an AI-by-list breakdown for NCAT strata cases or VCAT owners corporation cases.

It would therefore be wrong to claim that ChatGPT caused the wider NCAT or VCAT backlog.

There is solid evidence for a narrower conclusion:

  • AI-generated material is appearing more often in Australian legal proceedings.
  • Self-represented parties account for most of the Australian cases identified in the UNSW dataset.
  • Australian tribunal members have had to locate, check and address false or irrelevant material.
  • Bad AI use has caused adjournments, added costs, reduced the weight given to material and wasted court or tribunal time.
  • NCAT, VCAT and QCAT have each issued rules or guidance in response.

That is serious enough: a system does not need to be brought to a halt before wasted hearing time becomes a public problem.

Why strata disputes are especially tempting AI territory

Tribunals are intended to be more accessible than conventional courts. , and lawyers do not have an automatic right to appear in every tribunal matter. Owners may conclude that a paid legal opinion costs more than the dispute is worth.

At the same time, strata disputes generate an intimidating amount of material: meeting minutes, email chains, photographs, by-laws, plans, invoices, expert reports and years of maintenance records. A chatbot promises to read the lot in seconds.

The legal setting is also unusually easy to get wrong. NSW uses the Strata Schemes Management Act 2015 and NCAT. Victoria uses owners corporation legislation and VCAT. In Queensland, the BCCM office—not QCAT—is the first forum for most body corporate disputes. . Even within one state, different orders have different applicants, evidence requirements, time limits and pre-filing steps.

For example, NSW Fair Trading mediation is compulsory for most, but not all, NCAT strata applications. NCAT requires the applicant to identify the statutory order sought and attach the right supporting material. A plausible AI answer that skips mediation or nominates the wrong order can fail before the underlying leak, levy or by-law issue is properly considered. .

The tribunal is less formal than a superior court; the law still matters.

What has already gone wrong in Australian tribunals

The reported decisions show several different failure modes.

NCAT: AI-assisted submissions were given less weight

In , a self-represented applicant told NCAT that Google Gemini and ChatGPT had been used to prepare submissions.

NCAT's Procedural Direction 7 requires a person using GenAI for submissions to verify in the document that cited cases, legislation and other authorities exist, are accurate and are relevant. The applicant had not done that. The tribunal accepted the bundle but placed more weight on the underlying records than on the AI-assisted submissions. .

For a strata owner, a genuine invoice, dated email, strata plan or meeting minute may carry more weight than pages of confident legal prose. AI can make a bundle look sophisticated while making its useful parts harder to find.

NCAT: “improving” a statement can alter the evidence

In , the applicant said they drafted statements and then uploaded them to ChatGPT for improvement. NCAT pointed to its rule that GenAI must not be used to alter, embellish, strengthen, dilute or rephrase a witness's written evidence. .

That distinction catches people by surprise. A submission is an argument about what the evidence proves. A witness statement is the witness's own evidence. Asking AI to make a factual account “more persuasive” can change emphasis, add certainty the witness did not have or introduce a fact that was never observed.

For strata disputes, this matters when an owner is describing noise, smoke drift, committee conduct, water entry or conversations with contractors. Correcting grammar is not the same as rewriting the account into stronger evidence.

VCAT: a Gemini answer about fence heights had no evidentiary value

In , a party relied on a filed Google Gemini search about standard fence heights in Australia. VCAT found the answer did not assist with the claims. The member gave unverified, unattributed and untested statements produced by Google's AI no weight. .

This is a useful property-law example because the AI answer did not need to invent a case to be worthless. A general statement about what is “standard in Australia” does not prove the applicable planning rule, the terms of a registered plan, the condition of a particular structure or the cost of rectification.

A strata equivalent would be asking AI whether balcony tiles, windows or a waterproofing membrane are “usually common property”. The answer might describe a common pattern. Responsibility in the actual dispute may turn on the registered plan, the relevant legislation, the date and nature of work, a by-law, or expert evidence about the source of damage.

QCAT: checking hallucinations wastes public resources

In , a self-represented applicant relied on a supposed Queensland Court of Appeal authority that the tribunal could not locate. QCAT found that non-existent information weakens an argument, makes the tribunal less able to trust the rest of the material and wastes tribunal time and public resources. .

QCAT later issued Practice Direction 10 of 2025. It warns that relying on non-existent cases, legislation or other material may cause an adjournment, damage the party's credibility and potentially lead to an adverse costs order. .

Queensland owners face another AI trap before any citation is checked: most body corporate disputes belong first in the Office of the Commissioner for Body Corporate and Community Management. QCAT only hears specified disputes directly and appeals from adjudicators are confined to questions of law. .

The NSW Court of Appeal: more words, less help

The clearest Australian warning about AI and self-representation came from . The respondent used AI to prepare written material and a script for oral submissions in a property dispute. At least one authority was invented. Many real cases had little to do with the appeal, and the argument went beyond the issues the Court could fairly decide.

The Court found that the material added to the cost and complexity of the case without providing real assistance. It raised the prospect of more active case management to stop AI-generated material unfairly increasing the burden on other parties. .

This is how AI clogs a case: one person can generate 30 pages in minutes. Every proposition may then need to be checked by an opponent, a lawyer, an associate or a tribunal member. If authorities cannot be found, the hearing may stop while the party is asked to produce them. Corrected submissions may need to be filed. The other side may seek costs thrown away by the delay.

The rules are not the same at NCAT, VCAT and QCAT

Before using any AI tool, read the current practice direction for the forum hearing the case. Do not assume a court rule found in a search result applies to a tribunal—or that a rule from another state is close enough.

ForumCurrent position at 30 July 2026Practical effect
NCATProcedural Direction 7 applies to all proceedings. AI may assist with chronologies, indexes, briefs, document summaries and submissions, subject to privacy controls. It must not generate or alter witness evidence. Expert-report content requires leave. AI-assisted submissions must verify in their body that authorities and references exist, are accurate and are relevant; that check cannot be performed solely by AI.Do not paste a witness statement into ChatGPT for a rewrite. Keep AI away from expert opinion unless NCAT grants permission. Put the required verification in any AI-assisted submissions.
VCATPNVCAT 11 applies in every division. Parties remain responsible for accuracy and should disclose AI assistance, commonly by a statement on the document's front page. The practice note prohibits uploading specified confidential, protected, subpoenaed or suppressed material to an AI tool. Expert witnesses have additional duties.Check every output against an authoritative source. Disclose use. Do not assume a private dispute file is safe to upload because a chatbot account is password-protected.
QCATPractice Direction 10 requires accurate references. QCAT's non-lawyer guidance says public chatbots have limited Australian-law training and may not draw from AustLII, Queensland Judgments or official legislation. False references may lead to delay, loss of credibility or costs.Use official databases to check the law. AI can help organise facts or improve tone, but it is not the source of truth.

The , and should be read in full. Rules can change while a proceeding is on foot.

An appeal or judicial review may also move the dispute into a state court with different requirements. The and the impose their own controls. Do not assume that a use permitted in the tribunal will be permitted in the court.

Seven ways a chatbot can damage a sound strata case

1. It chooses the wrong legal doorway

An owner may have a real grievance but ask for an order the tribunal cannot make, apply under the wrong section or miss a mandatory preliminary step. AI often answers the issue described in the prompt instead of checking jurisdiction, standing, limitation periods and the exact remedy available.

“The owners corporation has ignored the leak” could involve an order to repair common property, damages for loss caused by a breach, access to a lot, an interim order, a building claim against a contractor or several separate claims. Those are not interchangeable.

2. It supplies law-shaped text instead of law

Large language models generate likely sequences of words. A citation can look authentic because the model has learned the shape of legal writing. It does not follow that the case exists or supports the sentence beside it.

Checking whether the case name appears in a database is only the first step. A real decision can still be from the wrong jurisdiction, distinguishable on its facts, overturned on appeal or quoted for a proposition it never decided.

3. It blurs evidence and argument

Tribunals decide facts from evidence. AI cannot turn a resident's suspicion into proof that a pipe is common property, a committee acted dishonestly or a neighbour caused water damage.

The risk increases when a user asks AI to fill gaps, strengthen a statement, reconstruct a conversation or create a cleaner image. A tribunal may then be unable to tell what the witness actually remembers and what the model supplied.

4. It exposes other people's information

A strata bundle can contain names, unit numbers, email addresses, phone numbers, bank details, health information, CCTV stills, access-control records and allegations about neighbours. Committee minutes may contain material never intended for a technology provider.

The Office of the Australian Information Commissioner recommends that organisations do not enter personal information—especially sensitive information—into publicly available generative AI tools. Depending on the service terms and settings, prompts may be accessible to the provider or another party, used for product development or be difficult to remove. Read the OAIC guidance.

De-identifying a document is harder than deleting the name at the top. A scheme number, address, lot description, dates and quoted emails may identify the people involved when combined.

5. It buries the useful point

AI can generate a long submission at almost no marginal cost, and length can feel like preparation. At a hearing, it may hide the four things that matter: the order requested, the legal power to make it, the facts that engage that power and the evidence proving those facts.

A tribunal member should not have to search 40 pages for the date the repair was first reported.

6. It escalates the dispute before the hearing

An AI-drafted letter may accuse a committee of negligence, bad faith, fraud or statutory breaches because those labels sound forceful. The accusation may be unsupported or legally irrelevant. It can harden positions at mediation and damage relationships in a building where the parties still share a lift, roof and insurance policy.

Ask for a specific action and date. Do not ask a chatbot to make the letter “more legal” or “more aggressive”.

7. It creates false confidence

AI answers arrive without the friction that makes a good adviser stop and ask for the strata plan, by-laws, prior orders, dates and correspondence. A fluent answer can make an uncertain case feel settled.

The risk is not confined to free chatbots. Legal-specific AI can also make mistakes. The difference is usually the quality of its source set, controls and links, not an exemption from verification.

A practical traffic-light guide

RiskTaskSafer approach
LowerTurn a list of dates you have already checked into a blank chronology formatRemove personal information first. Compare every row with the source document. Follow the tribunal's current AI rule.
LowerImprove grammar or reduce hostility in a non-evidentiary emailPreserve the facts and requested action. Read the final text aloud before sending it.
LowerSummarise a public tribunal webpageUse the summary as navigation, then read and link the official page.
MediumSummarise your own document bundleConfirm the tool's privacy terms and the tribunal's restrictions. A local or controlled system may reduce some risks, but it does not fix accuracy. Check every summary against the originals.
MediumBrainstorm questions for a lawyer, engineer or strata managerTreat the output as a question list, not an answer. Remove irrelevant items.
HighIdentify the claim, section, tribunal, time limit or ordersCheck official legislation and tribunal material. Obtain targeted legal advice where the answer affects filing or rights.
HighFind and summarise casesLocate every case independently and read the relevant paragraphs. Never ask one chatbot to verify another chatbot.
Prohibited or highly restrictedDraft or rewrite witness evidence, alter images or prepare expert opinionAt NCAT, generating or altering evidence is prohibited and expert-report use requires leave. Other forums impose their own controls. Prepare evidence from the witness's actual knowledge and preserve original files.
HighUpload a complete strata roll, complaint file, CCTV export or unredacted bundleDo not put personal, privileged, suppressed or confidential material into a public AI tool. Check privacy obligations and the forum's directions.

“Lower risk” means the task is easier to verify and less likely to determine a legal right, not that it is automatically permitted. The applicable practice note still controls.

How to verify an AI-assisted tribunal document

Use a source-first check before sending a demand or filing anything:

  1. Confirm the forum. Start with the official NCAT, VCAT, QCAT, BCCM, SAT, ACAT or court website. Check that the forum can make the order you want.
  2. Open the current Act. Use the authorised state legislation website. Confirm the section number, wording, commencement date and any transitional provision.
  3. Open every authority. Find the full decision on the court's site, an authorised report or AustLII. A search-result snippet is not enough.
  4. Read the cited paragraphs. Confirm the decision supports the exact proposition and applies to the relevant jurisdiction.
  5. Separate fact from argument. Mark each statement as something a witness can prove, something a document proves, an expert opinion or a legal submission.
  6. Check the remedy and procedure. Confirm standing, mediation, service, time limits, fees, forms and the wording of the proposed orders.
  7. Run the tribunal's AI rule last. Add any disclosure or verification statement and remove any prohibited use. Keep a record of how the document was produced and checked.

Legal Aid NSW gives similar advice to self-represented people: AI may help with some tasks, but a person can harm their case if they use it incorrectly. It directs users back to official legal resources and legal advice. Read Legal Aid NSW's guide.

For a wider view of filing fees, representation and costs risk, see our 2026 state-by-state tribunal cost guide. If the dispute is still at the complaint stage, use the documented escalation process in How to Raise a Strata Dispute Without Making Enemies. The Applying to a Tribunal playbook covers orders, evidence bundles and hearing preparation.

If you have already filed something AI helped write

Do not ask the same chatbot whether its citations are real. Do not create a second AI answer to defend the first.

Instead:

  • locate each authority in an official database;
  • compare every quotation with the judgment or legislation;
  • identify anything that does not exist, is inaccurate or is irrelevant;
  • preserve the filed version and your source documents;
  • obtain legal advice about correcting or withdrawing the material; and
  • notify the tribunal and other parties promptly if a correction is required.

Trying to conceal the problem can be worse than the original mistake. Australian decisions involving lawyers have led to personal costs orders, referrals to regulators and restrictions on practising certificates. A self-represented person does not carry the same professional duties, but still must not mislead the tribunal.

Use AI as an organiser, not as the source of truth

AI has useful jobs in a strata dispute. It can reduce the intimidation of a large file, turn dates into a chronology, simplify a public guide or take the heat out of an email.

It becomes dangerous when fluency is mistaken for authority. A tribunal does not need a submission that sounds like a barrister wrote it; it needs authentic evidence, the correct statutory power, a relevant order and a short explanation connecting them.

If AI saves an owner two hours of formatting but makes a member spend half a day checking invented cases, no time has been saved; the cost has moved to the other party and the public.

This article provides general information, not legal advice. Tribunal rules, legislation and AI products change. Check the current requirements for your forum and obtain advice about your facts before filing.

Sources and further reading

Keep the scheme file in one place the committee and the manager can both open. Features, pricing, or book a tour.