Every AI gets things wrong. Sometimes a small thing. Sometimes a confident, polished, completely fabricated thing. The What AI Cannot Do page covers why. This page is the practical companion. It walks through the checks that actually catch the mistakes, broken down by the kind of question you are asking.

The pattern that holds across all of them: treat AI as the first draft, not the source. A draft is something you read, push back on, and verify before acting on. A source is something you cite. AI is the first. It is rarely the second.

Health explanations

AI is good at the translation job. Plain-English explanations of a diagnosis, a discharge summary, a results sheet, a medication leaflet. It is also wrong often enough on the specifics that you should not act on its medical advice without a second source. Treat AI as a second opinion you got for free, not as the doctor.

For general health information and what symptoms to take seriously, check HealthDirect (the Australian federal-government health information service) or your state or territory health department. Both are written for the public and updated by clinicians. For medications specifically, the source of truth is the Australian Therapeutic Goods Administration (TGA) consumer medicine information leaflet for that drug. You can search for any TGA-approved medication and read the actual leaflet, which covers dose, interactions, and side effects in language you can follow.

Do not act on AI medical advice without speaking to a GP or pharmacist. Pharmacists are particularly under-used here. They are free to talk to, they are clinically trained, and they will tell you when something genuinely needs a GP visit and when it does not. Use AI to write down the questions before you walk in. The conversation goes faster and you remember what you wanted to ask.

Money questions

AI is genuinely useful for framing options in plain English. Fixed versus variable, offset versus redraw, the difference between an industry super fund and a retail one, what an "ongoing fee" actually buys you. The numbers themselves are where the trouble starts. Models confuse rates, mix up periods, miscount months, and forget the fee structure halfway through a calculation. Always get the numbers from a second source.

Use the Australian government's Moneysmart.gov.au for the official calculators on home loans, super, retirement income, and budgeting. ASIC (the Australian Securities and Investments Commission) runs it, the calculators are conservative, and the inputs are spelled out. Use the Australian comparison site Canstar for live comparisons on loans, savings accounts, credit cards, and insurance, but read the methodology, because rankings depend on what is being weighted. Or go directly to the bank, the super fund, or the insurer and look at their own product pages and product disclosure statements.

For super specifically, the product disclosure statement for your fund is the document that matters. It lists the fees, the investment options, the insurance cover, and the conditions. AI is a fine tool for translating it into plain English. It is not a fine tool for telling you whether a fee is reasonable for your circumstances. That is what a financial adviser is for, or in simpler cases the Moneysmart super comparison tool.

Legal and rights questions

AI is fine for "what does this clause mean in plain English". A residential lease, an employment contract, a phone contract, a body-corporate notice, the small print at the bottom of a quote. The translation job is something AI does well, and getting that translation in front of you before you sign is genuinely useful.

For anything you would actually act on, check a real Australian source. Your state's Legal Aid commission runs free advice lines on tenancy, family, criminal, and consumer matters. The Australian Fair Work Ombudsman has the official information on minimum wages, awards, leave, redundancy, and unfair dismissal. Your state or territory consumer affairs body (Consumer Affairs Victoria, NSW Fair Trading, Office of Fair Trading in Queensland, and so on) has the consumer-law side. Each of these is free and authoritative, and most of them have phone lines staffed by people whose job is to answer questions like yours.

One specific trap: AI will sometimes cite a court case, a section number, or an Act by name. About a third of the time, the citation is wrong or the case does not exist. Search for the citation directly before you use it. The Australasian Legal Information Institute (AustLII) hosts most Australian case law and legislation for free.

Quotes, statistics, court cases, and scientific papers

If the AI gives you a specific quote, a statistic with a number on it, a case name, an academic paper, or a book reference, search for it directly before you use it. About a third of these are wrong or invented. The wrong ones are the most dangerous because they sound the most plausible.

Search the exact phrase in quotation marks on Google. If a real source exists, it usually shows up on the first page. If nothing shows up, or if the only hits are AI-generated content scraping each other, the quote or the statistic is probably fabricated. For academic papers, search the title on Google Scholar; if the paper exists, you will find it. For court cases, search AustLII. For statistics, find the original release from the Australian Bureau of Statistics, the Australian Institute of Health and Welfare, or whichever agency would have produced it.

If you cannot find the source, do not use the quote, statistic, or citation. The model will sometimes apologise and give you a different one when challenged. That one might also be wrong. The verifying is yours to do.

Letters and emails on your behalf

This is the use case where the checking feels least urgent and matters most. The AI produces a clean, polished draft. You hit Send. Three days later you realise the tone was off, or the AI invented a fact about your situation, or it has cheerfully said yes to something you meant to say no to.

The recipe is simple. Read the draft slowly once, out loud if you can. Reading aloud catches the tone problems your eye skips over. Then ask yourself: does this say what I actually mean, in a tone I would be happy with if it came back to me? If the answer is anything short of yes, edit it. The model produced a draft. You produce the version that gets sent.

For sensitive messages, an apology, a complaint, a difficult email to a family member, give it more time. Sleep on it. Re-read it in the morning. The same draft sometimes lands very differently to the version of you who is no longer tired or angry.

Forms and applications

Never let AI fill an official form for you unless you read every line. AI fills boxes confidently. It will sometimes confidently put the wrong answer in the wrong box, and the form will go through. Treat the AI version as a draft. Read every field. Compare against your actual records. Sign nothing without checking.

The same applies to anything where you are agreeing to something. Loan applications, insurance applications, government forms, employment paperwork, tenancy applications. The AI does not know your circumstances as well as you do, even after you have explained them. Read every line.

A rule of thumb for the pace of checking

The more it matters, the more the checking matters. AI is fast and frequently wrong, and verification is the cheap part of the workflow. A two-minute fact-check on a number you are about to act on is a tiny tax on the speed of the rest. The version of you who skips it is the version of you who ends up writing a follow-up email to fix something that should not have happened.

The hard rules that go with this page, on what you should never paste in and what you should never decide on AI alone, are on Rules That Matter. The everyday verification habits that go with both are on Important Reminders.