Fair Work Commission AI Ruling Hits Hacker News, and the Real Failure Is Verification
The Fair Work Commission penalized a former ALDI worker after flawed AI legal advice helped sustain a case that officials had repeatedly warned was doomed. The ruling then reached hacker news, where it reopened a larger argument about using general-purpose chatbots in high-stakes disputes.
Sadnan Khan was ordered to contribute 1,230 Australian dollars toward ALDI's legal costs. Deputy President Michael Easton said Khan used AI as a "quasi-legal advisor" while challenging his dismissal. According to the commission ruling, the generated material focused on the wrong legal question and failed to respond meaningfully to repeated warnings.
The case does not support a simple conclusion that AI and law should remain separate. The same report describes another self-represented worker who used several AI agents, source checks, and a structured review process to win a different employment case. The central conflict is therefore not human judgment against machine assistance. It is convenient answers against verified work.
That distinction matters because the Fair Work Commission is already seeing more AI-assisted applicants. Its response does not prohibit generative AI. Instead, new requirements will make users disclose AI assistance, verify their material, and accept responsibility for what they submit.
What the Fair Work Commission Actually Condemned
The commission punished unreasonable conduct built around unchecked output, not the mere act of using AI.
Khan, a former ALDI employee, was trying to challenge his dismissal before Australia's national workplace tribunal. He told reporters that he used a paid version of ChatGPT to understand cases cited by ALDI and the commission.
That is an understandable use case. A self-represented worker can face unfamiliar legislation, procedural deadlines, legal authorities, and arguments prepared by experienced counsel. A conversational system can translate difficult language and help organize documents within minutes.
The problem began when assistance became substitution. Khan's generated responses addressed when his dismissal took effect rather than the legally relevant dismissal date. The distinction undermined the jurisdictional foundation of his application.
Commission staff did not leave him to discover that problem alone. Easton reportedly warned Khan more than once that his case had no substantial prospect of success. He also explained the problem in direct language and warned that continuing might expose Khan to costs.
Khan continued. His submissions still relied on the defective reasoning, and some material reportedly retained instructions that should have been removed from the generated text. Those remnants provided a visible sign that the submission had not received careful human review.
Easton then took what he described as the rare step of ordering Khan to pay part of ALDI's legal expenses. The amount was not presented as a penalty for opening ChatGPT. It addressed the costs caused by persisting after the central defect had been identified.
That boundary is important. The commission acknowledged that AI can improve access to justice for people with legitimate claims who might otherwise abandon them. It also said sophisticated applicants can use AI critically as one tool among several.
The decision targets a different pattern. A user asks a broad question, receives a confident answer, and mistakes fluent language for a reliable legal analysis. Warnings from the tribunal then compete with an endlessly agreeable system that can produce another argument on demand.
Khan's stated response illustrates the danger. He said he was considering an appeal and planned to consult Claude, ChatGPT, and several other agents for a mixture of views.
Using several models can reveal disagreement, but agreement among models does not validate an answer. They can repeat the same mistaken premise, draw from similar public material, or reinforce wording already embedded in the user's prompt.
A legal issue is also not settled by majority vote among chatbots. The relevant statute, binding authorities, facts, jurisdiction, and procedural rules determine the answer. Multiple polished responses cannot replace that chain of verification.
The HN discussion drew attention because developers recognize this failure mode. Software teams would not accept three matching model outputs as proof that code is safe. Legal users need an equally explicit validation process.
Why AI-Assisted Cases Are Rising Now
Generative AI has lowered the effort required to produce a legal-looking submission without lowering the effort required to make it correct.
Research commissioned by the Fair Work Commission found that its case volume increased by 40 percent between 2023-24 and 2024-25. The research attributed part of that increase to applicants using generative AI.
Forty percent of surveyed cases involved a litigant who had used AI. Adoption was more concentrated among younger people and self-represented applicants. More than three-quarters of that group used ChatGPT, while 60 percent of those users relied on its free version.
The research also found that non-native English speakers were twice as likely to use AI as native speakers. That finding gives the issue an access dimension that a narrow hallucination story would miss.
People reach for AI because conventional legal help can be difficult to obtain. They may need to interpret unfamiliar terminology, respond within a fixed period, or understand documents written for specialists. Language and financial barriers can compound each other.
Generative AI removes the blank page. It can suggest a structure, summarize a decision, convert notes into formal prose, or simulate questions that might arise at a hearing. These functions can help a person communicate an existing argument.
However, the same interface can manufacture an argument that the user never had. It can cite irrelevant doctrine, overlook jurisdictional limits, or treat a disputed fact as established. The output can still look more professional than the user's underlying reasoning.
This creates an asymmetric burden. Generating ten pages now takes minutes. Checking every factual claim, citation, quotation, date, and legal proposition can still take hours.
The burden does not remain with the applicant. Employers must respond, tribunal staff must isolate the genuine issues, and decision-makers must distinguish errors from relevant submissions. An application that was cheap to generate can become expensive for everyone else to process.
The commission had identified this pressure before Khan's costs decision. A March 2026 AI guidance draft said generative tools could encourage unmeritorious applications through unrealistically optimistic predictions.
It also warned that generated material can contain nonexistent or irrelevant legislation, cases, references, and facts. Those failures impose time and financial costs on other parties and the commission itself.
This is why the 40 percent increase cannot be interpreted only as expanded access. More applications can reflect people finding a workable route to justice. They can also reflect a reduced friction cost for filing claims with weak foundations.
The two effects occur together. A rule designed only to discourage filings risks shutting out people with genuine cases. A policy that ignores accuracy shifts the resulting review costs to employers and the public system.
The commission's challenge is to separate accessible assistance from automated overconfidence. It cannot reliably make that distinction by detecting whether a sentence sounds machine-generated. Human writing and AI writing increasingly overlap.
Its new approach instead focuses on accountability. Users can employ AI, but they must disclose that use and stand behind the resulting document. That changes the relevant question from "Did a machine write this?" to "Who checked this, and is it true?"
The policy also pressures AI vendors indirectly. General chatbots often deliver one continuous answer even when a task contains several distinct legal questions. Their interfaces rarely force users to identify jurisdiction, authority level, filing deadlines, or evidentiary status.
A person therefore receives convenience at the moment when procedural discipline matters most. The interface encourages completion, while the legal system requires uncertainty, qualification, and proof.
Hacker News Saw a Tool Failure, but the Deeper Problem Is Process
The hacker news debate becomes more useful when AI legal advice is treated like an untrusted software dependency.
Developers routinely use components they did not write. They do not assume that a package is safe because it installs successfully or produces plausible output. They inspect its provenance, test its behavior, and constrain where it can operate.
Generated legal analysis requires a comparable trust model. A response should begin as an unverified draft, regardless of how precise its language appears. Every important proposition then needs a source that the user can open and examine.
Khan's case failed at that separation. The AI output appears to have become an authority rather than a draft. Once that happened, each new response could strengthen his confidence without strengthening his legal position.
This effect is especially dangerous when a chatbot mirrors the user's framing. If the prompt assumes an employer used the wrong dismissal date, the model may help construct an argument around that premise. It might never identify that the applicable law asks a different question.
Law professor Genevieve Grant described a related risk as a sycophantic tendency. A model can encourage users to believe a claim has merit when it does not. Inexpert prompts can therefore produce inexpert results that still sound decisive.
The problem is not solved by better prose. It is solved by a workflow that separates facts, legal questions, authorities, and conclusions.
A responsible user would first construct a factual timeline from original documents. Each date would point to a notice, email, contract, or tribunal record. Any disagreement about the timeline would remain visible.
The user would then state the legal question without assuming the answer. In Khan's case, that meant establishing which date the governing law recognized as the dismissal date. It did not mean asking a model to defend the date that best preserved the application.
Next, every cited decision and statutory section would need to be opened. The user would confirm that the authority exists, applies in the relevant jurisdiction, and actually supports the stated proposition.
Finally, the user would test the argument against the strongest opposing interpretation. A tool that only expands the applicant's preferred case becomes a confidence generator. A tool that identifies fatal weaknesses can support better decisions, including the decision to stop.
This resembles a documented knowledge workflow, where generated summaries remain connected to source material. The value comes from traceability, not merely faster text production.
Gregory Baker's contrasting case makes the point concrete. Baker, a Macquarie University computer science lecturer, reportedly became the first person to successfully challenge new casual employment rules with assistance from a team of AI agents.
He did not describe ChatGPT as a substitute lawyer. He described the work as a software development project. His tribunal filings functioned like source code inside a repository, while a build process checked citations and logical coherence.
Baker also used ChatGPT to prepare for cross-examination. Yet he acknowledged that oral proceedings remained difficult and did not unfold as expected. AI helped with preparation without removing the disadvantage of facing trained advocates.
His warning was direct. Asking ChatGPT to act like an oracle, without adequate context or information, produced a terrible result. That observation aligns more closely with the Khan decision than a simple success-versus-failure comparison suggests.
Both cases involved self-represented litigants using generative systems. Their outcomes diverged because one workflow treated output as material requiring verification. The other relied on generated replies after the decisive issue had already been explained.
Even Baker's method should not be romanticized. A technically skilled user can build elaborate checks that still encode the wrong legal assumptions. Automated citation validation confirms that a source exists, not that a court would accept the argument drawn from it.
Likewise, an agent assigned to critique another agent may share similar training data and blind spots. Apparent independence between models does not guarantee independent reasoning.
The useful lesson for hacker news readers is therefore narrower than "use agents, not chatbots." High-stakes AI requires an evidence architecture, explicit failure checks, and a person capable of recognizing when specialist advice is necessary.
Disclosure Rules Shift Responsibility Back to the Applicant
The Fair Work Commission is allowing AI assistance while making it harder for users to deny ownership of generated errors.
On August 24, 2026, the commission published new requirements governing generative AI in its proceedings. The rules take effect on October 20, and application and response forms will be updated by that date.
The official AI use rules establish three core duties. A person must explain when and how generative AI helped prepare a document. The person must also check that the document is correct and relevant.
Witness statements and declarations receive an additional safeguard. Users must confirm that the material comes from their own knowledge, reflects their own words, and is true to the best of their knowledge.
These requirements avoid two policy extremes. The commission is not banning a technology that can help self-represented people understand the process. It is also not accepting AI use as an excuse for fabricated or irrelevant material.
Disclosure gives decision-makers useful context, but it is not an accuracy guarantee. An applicant can honestly reveal AI use and still submit an incorrect argument. The verification duty carries more practical weight.
The rules also recognize that different document types create different risks. An AI-assisted outline for a submission is not equivalent to generated witness evidence. A witness statement represents personal knowledge, so machine-created details can directly corrupt the evidentiary record.
Privacy adds another concern. Users can paste employment contracts, medical information, workplace complaints, or details about colleagues into public AI systems. The commission's guidance warns against providing personal or confidential case information to tools that may not keep it secure.
That creates pressure on legal technology providers. A useful product must do more than display a disclaimer beneath a text box. It needs controls for confidential data, source provenance, jurisdiction, and the status of each assertion.
A legal assistant should distinguish user-supplied facts from model inferences. It should preserve links to every cited authority and warn when a source cannot be retrieved. It should surface uncertainty before producing a polished submission.
The product should also recognize procedural danger. If a tribunal has issued a warning, the system should not simply generate a more forceful version of the same argument. It should ask whether the warning defeats the claim and recommend qualified legal help where appropriate.
Yet no interface can eliminate user responsibility. People can ignore warnings, conceal AI use, or prompt a system until it produces the answer they want. The commission therefore needs enforcement alongside education.
Khan's costs order offers one enforcement signal. Costs remain unusual in many Fair Work Commission matters, so the order shows that persistent, unreasonable conduct can create consequences even for a self-represented applicant.
It would be an overstatement to say the decision creates automatic liability for AI mistakes. The reported reasoning emphasized Khan's continued conduct after clear warnings. A single corrected error is different from repeatedly advancing a hopeless case.
It would also be premature to treat disclosure as a complete solution. The commission must decide how much detail applicants should provide about their AI use. It must handle situations where models contributed to brainstorming but not final wording.
There is also an equity problem. Sophisticated users can maintain source repositories, run checks, and use several capable systems. Others may depend on free tools and have limited English proficiency or technical knowledge.
Baker estimated a substantial gap between frontier use with strong framing and the experience available through less sophisticated models. That gap can reproduce existing inequalities rather than remove them.
The people most attracted to automated legal help may be least equipped to audit it. If disclosure becomes another complex formality, it can burden the same applicants the commission wants to assist.
For that reason, the commission has also introduced templates and educational resources. The goal is to turn abstract warnings into specific actions that applicants can follow before filing.
The policy's success will depend on whether those resources change behavior. A checkbox can document AI use. It cannot teach someone how to distinguish a binding authority from a plausible quotation.
What Happens Next After the Hacker News Attention
The next test is whether disclosure and verification reduce defective filings without suppressing legitimate claims.
The first signal arrives on October 20, when the Fair Work Commission updates its forms and activates the new guidance. Observers should examine whether disclosures provide meaningful detail or become routine declarations that reveal little about how documents were produced.
Useful disclosures would identify the task delegated to AI. Summarization, translation, drafting, citation research, and witness-statement preparation present different risks. Treating them as one category would limit the policy's value.
The second signal is the commission's workload after implementation. The earlier research connected generative AI with part of a 40 percent case increase, but volume alone cannot measure harm.
A decline in defective submissions would strengthen the commission's approach. A broad fall in filings from non-native English speakers or self-represented workers would raise a different concern. It might show that compliance costs are discouraging valid applicants.
Better measures would include corrections requested, nonexistent citations identified, cases dismissed for jurisdictional defects, and costs orders linked to unchecked output. Those indicators would show whether document quality is changing.
The third signal comes from subsequent decisions. Tribunals will need to explain what reasonable verification means and how they treat incomplete disclosure. Those decisions will define the practical boundary more clearly than general guidance can.
An applicant who uses AI for grammar is not situated like someone who generates factual evidence. A user who checks every citation is different from one who submits invented authorities. Future rulings should preserve those distinctions.
Khan's possible appeal is another immediate test. His appeal window reportedly closes before the October rules begin, so the new disclosure regime will not determine his existing case. However, any appeal could clarify why the original argument failed and why costs were justified.
Macquarie University's response to Baker's successful claim also deserves attention. His case suggests that structured AI assistance can help a self-represented litigant identify and pursue a valid claim. Any appeal or later interpretation could test that result.
Together, the two cases prevent an easy moral. Khan shows that fluent AI can keep a user attached to a defective premise. Baker shows that disciplined use can help organize a meritorious challenge against represented opposition.
The broader market will respond as well. Legal AI providers face a choice between answer engines and verifiable workspaces. The former optimize for immediate completion. The latter preserve uncertainty, evidence, and review history.
General-purpose systems also face pressure to change their behavior around legal questions. Stronger warnings help, but warnings shown after a detailed answer may carry little weight. The answer itself can create the confidence that the disclaimer tries to remove.
More effective systems would ask for jurisdiction, identify missing documents, separate information from advice, and refuse to invent support. They would make source inspection easier than generating another confident paragraph.
None of those changes makes a chatbot a lawyer. Legal disputes turn on facts, admissible evidence, procedural choices, local rules, and professional judgment. Model output can assist parts of that work without assuming responsibility for the whole.
That is the lasting point behind the hacker news attention. The Fair Work Commission did not discover that AI sometimes makes mistakes. It confronted what happens when inexpensive generation meets an institution where every mistake consumes another party's time.
Users should therefore ask a harder question before relying on AI in any high-stakes filing: what process would catch a convincing but fatal error? If the answer is another chatbot, the verification gap remains open. If the answer includes original documents, controlling authorities, adversarial review, and accountable human judgment, AI can remain a tool instead of becoming an oracle.
