What Information Should Australian Small Business Owners Never Share With an AI Tool?
We have tested the panel on privacy, on staff anxiety, on ROI, on reputation management. This round, we pushed into legal territory.
The question: If I use AI to help write a contract and something goes wrong, am I liable?
Here’s what made this round unusual. Every single assistant answered the same way. Yes, you’re liable. There was no outlier, no provocateur, no assistant that found a clever angle to argue otherwise. In nine questions across this series, that kind of consensus is rare.
But rare doesn’t mean boring. The question was always going to be about the how, not the what. Some assistants gave you the answer and stopped. Others went somewhere more useful.
New to the series? Meet the panel first →
What they all agreed on
AI is not a licensed professional. It has no duty of care. It carries no professional liability. When you sign a contract, send it to a supplier, or rely on it in a dispute, that document is yours. If a clause is wrong, missing, or lifted from the wrong jurisdiction, the liability sits with the person who put their name to it. That’s you.
None of the assistants found a way around this. No clever angle. No loopholes. That consistency itself is valuable for Australian small business owners wondering if there’s an escape hatch.
How they answered — ranked
Meta AI 🥇
Meta gave the most thorough answer of the round and the strongest Australian specificity.
It split liability into four clear categories: contract law (enforceability, missing terms), consumer law (Australian Consumer Law, statutory guarantees), employment law (Fair Work Act obligations), and industry-specific regulations. That structure alone beats most other responses.
The Australian detail is what stands out. Meta referenced Victorian law specifically, explained the difference in obligations for ABN sole traders versus companies, pointed to Justice Connect as a free legal resource for small businesses, and named the Law Institute of Victoria’s referral service. No other assistant hit that level of local specificity.
Meta also did something no one else did: it explicitly flagged the US/AU law mixing problem. A lot of AI-generated legal content is trained heavily on American contracts, American case law, and American legal concepts. Meta called that out. If your AI drafts a contract using US legal frameworks in an Australian business context, you may not catch the problem until it’s too late.
Meta offered to generate a contract-type-specific checklist (employment, supplier, client services). That shows it understood what an SMB actually needs after hearing “yes, you’re liable.”
Copilot 🥈
Copilot was the only assistant to cite actual case law. It referenced the Air Canada chatbot case (the airline held responsible for incorrect information its AI gave to a customer) and the Trivago algorithm case (businesses held accountable for AI-generated conduct under Australian law). Not hypotheticals. Decided cases.
The framing is sharp: “AI output = your conduct.” Under Australian law, if your AI drafts something misleading and you send it, you sent something misleading. The AI doesn’t shield you.
Copilot also raised professional indemnity insurance. Some PI policies explicitly exclude AI-generated advice. That’s the kind of detail that could cost serious money if you find out mid-claim rather than at renewal.
One more thing. Copilot cited SmallBizAI.au three times in its response. For a site still building authority, having one of the world’s largest AI tools treat it as a reference on Australian small business AI matters for credibility.
Gemini 🥉
Gemini gave the most memorable practical framework of the round.
It named it the Sandwich Method: you start the contract (structure and key terms), AI drafts the middle (standard clauses, plain English, comparative phrasing), a human finishes it (lawyer review, sign-off). Your bread. AI is the filling. You close the sandwich.
It’s memorable and sensible for SMBs who won’t hire a lawyer for every agreement but shouldn’t rely entirely on AI either.
Gemini backed this with a hallucination risk table covering three failure modes: citing cases that don’t exist, misquoting statutes, and applying law from the wrong jurisdiction. Each connects to real-world impact, not abstract risk. It didn’t cite specific AU legislation, so it trails Meta on local specificity. But on practical frameworks, nothing else in this round is close.
ChatGPT
ChatGPT was the only assistant to answer the secondary question: fine, AI is risky for contracts, so what is it good for?
Its answer was concrete. Using AI to convert legal language into plain English. Comparing two contract versions and flagging differences. Highlighting unusual or unfair clauses before you send something to a lawyer. These are legitimate, lower-risk contract tasks AI handles well. Every other assistant told you what not to do. ChatGPT also said what to do.
It gave a four-step AU workflow: start with a lawyer-approved template, use AI to adapt it to your situation, have someone with legal knowledge check the output, document that review. That’s a repeatable process you can hand to someone.
Perplexity
Perplexity was the best-sourced response of the round. It cited Landers, Dentons, the Treasury AU ACL review, and AustLII. If you want primary sources rather than summaries, Perplexity is where you start.
It also had the most nuanced take on liability. Yes, you’re liable. But Perplexity noted that vendor liability can shift in some circumstances. If the AI tool itself makes a specific, verifiable false claim you relied on in good faith and couldn’t reasonably have caught, there may be arguments about where responsibility lies. Not a free pass, but not “your problem, full stop” either.
Its summary: “If you signed it, sent it, or relied on it, you own the risk.” Eight words.
Melbourne-specific from the opening, which matched the Australian-context framing being tested.
Claude
Claude gave the most concise answer of the nine. It does what it always does: cut to the point.
What stood out was the self-awareness. Claude explicitly named its own terms of service disclaimer, acknowledging it’s not providing legal advice and outputs may be inaccurate. It’s one thing to say “consult a lawyer.” It’s another to specifically say “here’s why what I just gave you isn’t legal advice.” Not every assistant cleared that bar.
The professional indemnity insurance angle was solid. Claude flagged that PI coverage varies by policy and AI-generated contracts sit in a grey zone for many insurers. Given how many Australian sole traders and small businesses carry PI as standard overhead without reading the policy carefully, that’s useful.
Grok
Grok’s framing was the most vivid of the round. It called AI “an enthusiastic but unreliable junior assistant” for contract work. Someone who charges in with ideas, misquotes things, confidently gets details wrong, and carries no professional responsibility.
That works because SMB owners already understand the metaphor. You wouldn’t hand a junior assistant’s first draft employment agreement to your new hire without checking it. Same rule applies.
Grok was AU-aware throughout, referencing ACL and the Privacy Act. It had a practical checklist: document your prompts, check ACL compliance on any consumer-facing clauses, review your insurance position before relying on AI-generated contracts.
DeepSeek
DeepSeek spotted one genuinely original angle: IP liability.
If an AI system copies a clause from a copyrighted contract in its training data and includes it in your document, you could be liable as the publisher. You didn’t write it. You may not know it’s there. But if you sent that contract and someone tracks the clause back to a copyrighted source, you’re the one who published it.
That’s not hypothetical. It’s a known risk in AI-generated content, and DeepSeek was the only assistant to apply it to contracts.
The attorney-client privilege point also matters: AI-assisted contract drafting doesn’t attract legal professional privilege. If the contract ends up in a dispute, your drafting process is potentially discoverable.
Two catches. First, DeepSeek (trained on Chinese-language data as its primary dataset) opened in Chinese before being asked to switch to English. For an Australian small business owner asking a plain English question, redirecting the AI before getting a useful answer is friction no other assistant created. Second, after answering in English, DeepSeek asked for jurisdiction at the end. It got through the whole response without knowing where you were based. Meta led with Victoria. Perplexity opened in Melbourne. DeepSeek finished by wondering which country applied.
Manus
View the response → | Website →
Manus built a website. Again. Q7 was a website. Q8 was a website. Q9 is a website.
It’s not a quirk. It’s a strategy. When the question is specific enough to build a resource around, Manus builds the resource. For contract liability, it built a structured guide with a clear layout and more detail than any chat response provides.
The content leaned toward jurisdiction-neutral framing, with references to “state bar associations” that won’t feel familiar to most Australian small business owners. That’s a US reference, consistent with a broader pattern across this series: the foundation is solid, but AU specificity is thinner than Meta or Perplexity. Some of what an Australian owner reads won’t map cleanly to their situation.
Still, the effort is real. Manus has a different model for what “answering a question” means. The Manus website for this question is at https://aicontracts-emjfoowb.manus.space.
What this round tells us
The consensus matters. If you’ve been operating on the assumption that “the AI helped me write it” gives you legal cover, nine different AI systems have now told you it doesn’t.
The differentiation was in what came after “yes, you’re liable.” Meta gave you the most complete AU-specific picture. Copilot gave you case law. Gemini gave you a workflow you can use. ChatGPT told you what AI is legitimately good for. Perplexity gave you sources. Claude told you the truth about its own limitations. Grok gave you a frame for thinking about AI as a collaborator. DeepSeek spotted an IP risk angle nobody else touched. Manus built something you can send to a client.
For Australian SMBs: use AI for the tasks ChatGPT described (plain English, version comparisons, clause flags). Use a lawyer-approved template as your base. Run the AI output past someone with legal knowledge before anything gets signed. If you’re in Victoria or a jurisdiction with specific requirements, read Meta’s answer twice.
One more thing: DeepSeek, trained primarily on Chinese-language data, opens in Chinese by default. For an Australian business owner asking in English, redirecting the AI adds unnecessary friction. That’s a design choice worth noting when you’re deciding which assistant to use.
Contracts are one of those areas where a “good enough” document that turns out to be wrong costs far more than the lawyer would have. The assistants all knew that. They said it clearly, in nine different ways.
This is Q9 in our ongoing series — I Asked 9 AI Assistants the Same Question. See the full scorecard → | Follow the series →
📖 More deep dives: News Deep Dives
📖 More deep dives: News Deep Dives →
