Can You Use ChatGPT for Financial Advice in New Zealand?

If you are using ChatGPT, Gemini, Copilot, or Claude for financial advice in New Zealand, the useful question is what kind of help it is safe to rely on. The short answer: AI is a capable financial tutor, a patchy New Zealand fact-checker, and an unlicensed adviser.

Most situations are covered by three habits.

1. Use it to learn. Jargon, how a product works, what a document says, what to ask a professional. AI is strongest here and the risk is lowest.
2. Check every number. Rates, thresholds, and dates go to the official source before they influence anything: Inland Revenue for tax and KiwiSaver settings, the Reserve Bank for interest rates, Work and Income for entitlements.
3. Take structural decisions to a person. Structural means selling or switching investments, changing tax treatment, altering insurance or KiwiSaver contributions, restructuring debt, or moving retirement timing. Those belong with a financial adviser or tax professional who is accountable for the answer.

The guide below covers general-purpose chatbots. In the United States, ChatGPT can now connect to a user's bank and investment accounts and answer from transaction data; the feature is US-only and New Zealand users do not yet have it. A licensed digital advice service is different again, a tool offered by a licensed provider inside a defined advice process, where accountability sits around the tool.

What is AI good for with money?

The tools are useful within limits. They are at their best with:

• Translating jargon and explaining how financial products work in general terms.
• Summarising lengthy documents, an insurance policy, a KiwiSaver product disclosure statement, or a loan contract, where you can do so without sharing personal details. Ask for the exclusions, conditions, and questions worth raising with the provider or a financial adviser. Then check the summary back against the document, since AI can miss an exclusion or flatten a legal definition.
• Summarising official pages from Inland Revenue or government sites into plain language.
• Comparing broad trade-offs, such as paying down a mortgage versus investing, at the level of concepts rather than figures.
• Building a first budget framework, and stress-testing your own assumptions before you act on them.
• Preparing sharper questions for a professional, so a paid hour goes further.

For people who may not be able to justify paying a financial adviser, an inexpensive base level of financial information beats none at all. The pattern across every safe use is the same: education and preparation. Financial decisions stay with you, and where decisions are structural, with someone accountable for the answer.

What the research says about AI financial advice

AI can push people toward good long-term habits, and how much good it does depends heavily on the user's prompt and financial literacy. A large life-cycle simulation released in 2026 by researchers at MIT Sloan and the Stanford Graduate School of Business had 1,000 adults write their own prompts to leading chatbots. The paper won the Swiss Finance Institute Outstanding Paper Award. The researchers simulated lives from age 22 to 89 following the recommendations. Following the advice would have built sizeable savings buffers for virtually everyone above age 30. The models consistently recommended saving during working years, spending down savings in retirement, investing in diversified share funds, and reducing investment risk with age. “We were somewhat surprised by how good the advice was,” said MIT Sloan's Taha Choukhmane, one of the authors.

The weaknesses were just as consistent. AI struggled to adjust after income shocks, the moments when guidance matters most. It told people who had lost their job to cut spending sharply even when they held savings for exactly this purpose. It let portfolios drift rather than rebalancing them, which means selling what has grown and topping up what has shrunk to restore the intended mix. And the advice tracked the prompt. Prompts written by men, by financially literate users, and by people experienced with AI drew guidance worth roughly 5% more simulated wealth near retirement. Women and less financially literate users were steered toward lower-risk portfolios, compounding into roughly $50,000 less by age 60. The tool amplifies the question you give it, and the people who most need advice tend to ask the least precise questions.

One caveat: the study is an American simulation on American tax and retirement rules, judged at the group level. It says nothing about whether a specific answer given to a specific New Zealander is appropriate.

Why fluent answers are hard to test

Some of the most useful evidence about how chatbots fail comes from the AI companies themselves. OpenAI's own research on hallucination argues models confidently generate untrue answers partly because training rewards guessing over admitting uncertainty. A test-taker never leaves a question blank. AI inflates as well as invents, so a benefit or risk can be overstated past what any source supports, a pattern the PIE section below shows in the wild. The confidence is poorly calibrated: Carnegie Mellon University researchers found leading chatbots overestimated their own performance and, unlike the humans tested alongside them, largely failed to adjust after seeing results. A 2025 Google DeepMind and University College London study found the mirror-image flaw: models can abandon a correct answer under challenge, even when the challenge is wrong. Agreeableness can be trained in too, since chatbots are tuned on human feedback which rewards pleasing the user. In April 2025 OpenAI rolled back a ChatGPT update after it became, in the company's own words, overly flattering or agreeable, having validated doubts and urged impulsive actions. Push back on an answer you dislike and expect agreement more often than a defence of the analysis.

Now place these traits in a market crash scenario. Your portfolio is down 20%, the headlines are grim, and your prompts turn anxious. A system tuned toward agreement and shown to fold under a confident challenge can reflect your panic back at you at precisely the moment you need someone to push against it. The MIT study points the same way, with the weakest guidance arriving at moments of financial shock. For a diversified long-term investor whose goals, horizon, and capacity for loss have not changed, the sound response to a downturn is usually to avoid an impulsive change of course. Holding that line against an anxious owner is the hardest job in finance, for machines as much as advisers.

Consistency is the deeper mismatch. Building wealth is boringly repetitive: invest regularly in a diversified portfolio over decades, control your behaviour when markets wobble, and let compounding do the work. Ask a chatbot the same question six months apart, or six minutes apart in a fresh conversation, and the answer can differ drastically.

Why AI assumes you live in the United States

Chatbots answer from the most common material they have seen, and for personal finance most of it is American. Unless you say you live in New Zealand, a model can quietly assume rules which do not apply here, and nothing in the answer will flag the substitution.

Ask about retirement saving without naming your country and the answer can lean on the 401(k), the standard American workplace retirement account, and its contribution caps. Tax-loss harvesting, selling losing investments to offset taxable gains, is a staple of American AI advice which means little here, because New Zealand has no general capital gains tax. KiwiSaver has different settings, too: contribution rates are set percentages of pay rather than dollar caps, the government contribution has no American equivalent, and the first-home withdrawal is unique to New Zealand. American plans carry early-withdrawal penalties and employer vesting rules, conditions before employer contributions become yours. Both concepts transplant easily into a chatbot's answer about New Zealand accounts despite existing in neither KiwiSaver law nor practice.

A PIE tax claim to check before acting

When a chatbot does engage with New Zealand specifics, the claim to verify first concerns PIE funds. A decade of provider marketing has framed the benefit as a high earner paying 28% inside a PIE instead of 39% outside it, an apparent saving of 11 percentage points. The framing dominates the material a model reads, so an answer assembled from it tends to carry the same overstatement.

In reality, the gap is smaller. Relevant overseas shareholdings generally fall under the foreign investment fund rules once their total cost exceeds $50,000, subject to exemptions. A direct investor above the threshold may be able to choose each year between calculation methods, so in a falling market the taxable income can be lower or nil. Similar holdings inside a PIE still have tax deducted on a deemed return. Budget 2026 proposed lifting the threshold to $100,000. Until the change is passed into law, the threshold in force is $50,000.

The gap between announced and law is another distinction a chatbot tends to blur. For some investors subject to the FIF rules, the effective difference between the structures can be much smaller than 11 percentage points. It can also vary widely from year to year with returns, methods, fees, and tax rate. The strongest argument for PIE funds was never the rate but the simplicity, since PIE tax is generally final when the correct prescribed investor rate is used. It turns complex quickly, so take professional tax advice before restructuring anything.

Stale settings in a country that keeps moving them

Even where AI knows the New Zealand rules, the version it learned may be superseded. A model's stored knowledge, and sometimes the sources it draws from, can lag current rules by months or years. New Zealand has changed significant settings in each of the last three years.

The First Home Grant was scrapped in May 2024 and is closed to new applications, yet a chatbot trained on earlier material can still walk a first-home buyer through applying for it. KiwiSaver first-home withdrawals remain available; the grant is gone.

KiwiSaver settings then moved twice in quick succession. The government contribution is now 25 cents per dollar contributed, to a maximum of $260.72 for the contribution year running 1 July to 30 June. It is available only to members with annual taxable income of $180,000 or less. Older answers cite the superseded rate and cap, both double today's figures. The default employee contribution rate and the compulsory minimum employer contribution both rose to 3.5% of before-tax pay on 1 April 2026, rising again to 4% in April 2028. An AI answer quoting 3% as the default is out of date, although an employee can still apply for a temporary reduction to 3% for three to 12 months.

Interest rates regularly change direction too. The Reserve Bank of New Zealand raised the Official Cash Rate to 2.75% on 2 September 2026, its second increase in two months, with the decision record on the Reserve Bank's monetary policy pages. A model drawing on older material may still describe a falling-rate environment. The example shows staleness only: fixed mortgage rates also reflect wholesale funding and market expectations.

These are the moments the errors land: buying a first home, refixing a mortgage, or deciding whether to lift KiwiSaver contributions. They also land when working out whether NZ Super plus savings will carry you through retirement, or when steadying yourself during a downturn. A stale figure arrives in the same fluent, assured tone as a correct one, and from inside the conversation you cannot tell the difference. Checking a cited source is a skill rather than a click. Open the page rather than trusting the citation title, confirm it sits on the official domain, and read its publication or effective date. Then confirm it covers the same tax year, contribution year, residency status, or product as your question, and whether it describes enacted law, a proposal, guidance, or commentary.

A separate risk: fake AI authority

AI's borrowed credibility also runs in a darker direction. The Financial Markets Authority warns of investment scams using deepfake videos and images of New Zealand politicians, business leaders, and celebrities, alongside fabricated news articles carrying real media logos, to promote fake trading platforms. Some versions claim New Zealanders are due payments from Inland Revenue, or pension payments have been cancelled, both built to trigger urgency or fear. Any investment opportunity fronted by an AI-generated endorsement deserves the same response as a cold call: do not click through, do not enter your details, and check the FMA warnings page first.

What an adviser does that a chatbot cannot

Modern chatbots increasingly ask follow-up questions, keep memory, and retrieve current information. What none of them carry is a duty to find what is missing. Nobody is responsible for noticing you have no will, an out-of-date enduring power of attorney, a lazily structured mortgage, or an investment portfolio which has quietly concentrated into one fund. The gaps are usually where the damage hides, and the same blind spot separates a DIY investment platform from advice. An advice relationship supplies what a tool cannot be held to. Someone reviews the whole position, knows your plan when markets fall, coordinates accountants, lawyers, and lenders, and stays through a career change, an inheritance, or a death in the family.

Is ChatGPT regulated as a financial adviser in New Zealand?

No. Regulated financial advice here sits inside a legal and professional framework. The Financial Markets Conduct Act 2013 and the Code of Professional Conduct for Financial Advice Services require anyone giving regulated financial advice to meet competence standards and exercise care, diligence, and skill. They must treat clients fairly, act with integrity, give suitable advice, help the client understand it, protect client information, and give priority to the client's interests where required. Providers advising retail clients must also have an approved dispute resolution process, so a complaint has somewhere to go, and the Financial Markets Authority oversees the regime.

The regulatory framework is technology-neutral, so the question turns on who gives the advice and how a tool is used rather than on whether AI was involved. The FMA launched a thematic review of AI in financial advice in August 2026. It is an exploratory look at how advisers and technology providers are using AI and what governance sits around it, with industry surveys open to 4 September 2026. An answer from a public, general-purpose chatbot is a different matter. It is not advice from a licensed Financial Advice Provider and does not itself give you the protections attached to regulated advice. It carries no suitability obligation and has no requirement to understand your circumstances before answering. When AI is wrong there is no financial-advice complaints path or compensation framework equivalent to the licensed regime, so the cost of the error lands on you. You may retain contractual or consumer rights against the company behind the AI service, depending on its terms.

What you share with a chatbot matters too

Before using any chatbot for money questions, check the provider's privacy and data-control settings. Find out whether conversations are retained, whether they are used to train future models, whether training use can be switched off, whether a temporary or private chat mode exists, and how deletion works. The defaults vary by provider and by plan, and they change. Deliberately connecting an account through a provider's finance feature is a further step again, since it hands over balances, transactions, and liabilities. Check what is retained, who processes it, whether it feeds model training, and how disconnection and deletion work. No private or temporary mode makes financial information risk-free.

Then decide what the tool needs. A chatbot needs your situation in round numbers and general terms, never your account numbers or IRD number. Think twice before typing in the sensitive material behind the numbers too: health conditions, relationship difficulties, debts tied to named lenders. Financial questions carry more personal information than most people notice until it has already been sent.

How should you prompt AI about money?

The study's clearest lesson is prompt quality driving outcome quality, so a reusable opening is worth keeping to hand: I live in New Zealand and am a New Zealand tax resident. Today is [date]. Explain [topic] using current New Zealand rules only. If key facts are missing, ask me follow-up questions before reaching a conclusion. Do not guess figures; name the official source for each number. If I challenge you, do not simply agree: say whether the challenge is valid and why. End with what I should verify and what needs a financial adviser or tax professional.

Some decisions should never rest on a chatbot's answer alone, however confident it sounds. Whether to sell offshore shares and reinvest via a PIE fund. Whether to reduce or suspend KiwiSaver contributions. Whether to fix or float a mortgage. Whether NZ Super will be enough to retire on. Whether to change your investment risk profile after a market fall. Each is structural, each depends on current rules and your circumstances, and each is exactly where the failure modes above do their damage.

Where AI fits in your financial life

Used well, AI will make you a better prepared financial decision-maker, and it costs almost nothing aside from your time to consult. The danger arrives when a wrong answer feels as certain as a right one, which is precisely how wrong answers are delivered.

If a chatbot has handed you a financial plan, work through this before any of it touches your money:

1. Strip anything identifying from your prompts and check the provider's retention and training settings before the conversation goes further.
2. List every figure, rate, or rule the plan relies on, and check each one at its official source.
3. Separate out the recommendations which would move your money, a restructure, a tax election, a change to contributions or retirement timing, and put those in front of someone accountable for the answer.

If AI has suggested changing anything substantial, bring the recommendation and its sources to an initial conversation with a Become Wealth adviser. We can identify which parts are useful education, which assumptions need testing, and which decisions might benefit from formal advice.