What Are the Limitations of ChatGPT?

Last updated: October 2026.

The limitations of ChatGPT are that it states wrong facts confidently, cannot see your business systems, should not be given private customer data, and holds no accountability for what it produces. For an Australian business the practical risks are privacy exposure, misleading claims under consumer law, and content that reads like every competitor’s.

What is ChatGPT good at?

ChatGPT is genuinely good at transforming text you already have: rewriting, shortening, reordering, translating, summarising and drafting from a clear brief. It is also strong at producing structure, such as turning rough notes into a checklist, an agenda or a set of headings.

The pattern is that it performs well when the source material is in front of it and poorly when it has to remember facts. Give it the information and ask it to reshape that information, and the output is reliable. Ask it to recall something and the output is a guess in fluent English.

Holding that distinction is most of the skill of using it. The twelve limitations below are all variations of the same thing: ChatGPT is a language tool, not a source of truth.

What are the 12 limitations of ChatGPT?

Each limitation below follows the same shape: what it means, a real example an Australian business would recognise, and the workaround. None of them are reasons to avoid the tool, and all of them are reasons to keep a person reviewing the output.

1. It states wrong facts confidently

ChatGPT produces fluent text whether or not it has a reliable basis for it, so a wrong answer arrives with the same confident tone as a correct one. This is what people mean by hallucination, and it is more common in narrow, numerical and local facts than in general explanation.

A plumbing business asks for the average cost of a bathroom renovation in Melbourne and receives a precise figure that no one can trace to a source. It reads well enough to publish, and it may be years out of date or invented.

The workaround is to never publish a number you have not opened at its source. Ask where the figure came from, and if the answer is vague, treat the figure as fictional and find the primary source yourself.

2. Its knowledge can be out of date

Even with web search switched on, ChatGPT may answer from an older training position rather than from a current page, and it does not reliably signal which it is doing. Products, prices, platform features and regulations all change faster than any model retrains.

A marketing agency asks for current Google Business Profile features and receives a list including tools that were retired, mixed in with ones that are live. Nothing in the answer flags which is which.

The workaround is to treat any answer about a live platform as a starting hypothesis and confirm it against the vendor’s own documentation on the day you publish. Where a post states a platform feature, run it at the source.

3. It does not know your business, prices or stock

ChatGPT has no access to your job history, your rate card, your stock levels or your service area unless you put that information into the conversation. It will still answer as though it does, which is where invented pricing comes from.

An electrical contractor asks for a quote template and gets one with callout fees and hourly rates that belong to nobody, which then goes out to a customer.

The workaround is to build your own template once, with your real rates, and give it to the tool as context each time. Better still, keep pricing in your quoting system and use AI only for the wording around it.

4. Pasting customer data creates a privacy risk

The Australian privacy regulator is explicit on this point. The OAIC recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools, because of the significant and complex privacy risks involved (OAIC).

A receptionist pastes a customer’s full complaint email, including their address and phone number, into a free chat to get help wording a reply. The customer’s details are now outside the business.

The workaround is a written rule and a rewritten habit: remove names, addresses, dates of birth and account numbers before anything is pasted, and use placeholders such as Customer A. Where real data must be involved, use a business plan with administrative controls rather than a personal account.

5. It cannot verify claims, which creates consumer law risk

Australian consumer law applies to what a business publishes about its products and services, regardless of how the words were produced. A claim must not be false or misleading, and an AI draft that overstates a result is still your claim once it is published (ACCC).

A skincare retailer asks for marketing copy and receives lines promising clinically proven results, with no study behind them and no approval to make that claim.

The workaround is to give the tool only claims you can substantiate, and to have someone check every superlative before publishing. Words like guaranteed, proven, best and number one should be treated as claims, not as adjectives.

6. Generic copy that sounds like every competitor

Unedited output regresses to the average of everything the model has read, which is why AI-drafted pages in a category all read the same. The result is accurate in a general sense and indistinguishable from competitors, which is a conversion problem rather than a ranking one.

Three Melbourne web design studios publish service pages that open with the same sentence about a strong online presence, because all three used the same tool with the same brief.

The workaround is to supply the specificity the model cannot invent: your own job examples, real customer objections, named suburbs, actual timeframes and your own figures. Generic input produces generic output, and specific input is the only thing that breaks the pattern.

7. US-centric answers, spelling and pricing

ChatGPT defaults to American English, American spelling, US dollars and US legal and tax frameworks, because that is the majority of its training material. Left unchecked, a draft will use American spelling throughout, quote prices in US dollars, and reference tax rules that do not exist here.

A Brisbane accounting firm publishes a blog post referring to IRS filing deadlines and 401k contributions, which destroys its credibility in one sentence.

The workaround is to instruct the tool at the start of every conversation to use Australian English, Australian dollars and Australian law, and then to spell check the output anyway. Instructions reduce the problem; they do not eliminate it.

8. Weak knowledge of Australian suburbs and councils

Local detail is where ChatGPT is thinnest, because it has less material on individual suburbs, council areas and state regulations than on global topics. It will write confidently about local relevance while getting the geography and the rules wrong.

A lawn and garden business asks for local content and receives a draft referring to the wrong council for their postcode and to planning rules from a different state.

The workaround is to never let the model choose your local facts. Name the suburbs, the council and the state in the brief, and check anything about regulation against the council or state authority directly.

9. Small prompt changes give very different answers

Rewording a prompt slightly, or asking again, produces materially different output, which makes results hard to repeat across a team. Without a saved prompt, two staff members asking the same question get two different documents.

A real estate agency has three people writing property descriptions, and every listing reads in a different voice and structure.

The workaround is a saved prompt library: when a prompt produces a good result, keep it with a note on what it was for. Consistency across a team matters more than cleverness in any single prompt.

10. Usage limits and plan restrictions

Free and lower tiers impose limits on how much and how often you can use the strongest models, and those limits changed repeatedly as the product developed. A workflow that works in testing can stall in a busy week for reasons unrelated to your work.

A two person business builds its whole quote process on a free account, then hits a limit on a Friday afternoon with customers waiting.

The workaround is to avoid making any time critical process depend on an AI tool with no fallback, and to check current plan limits in the product rather than trusting any guide. Where a process matters, document how it works without the tool.

11. It cannot act inside your systems without integrations

A chat window produces text. It does not book the job, update the customer record, send the invoice or answer the phone unless something has been built to connect it to those systems. Many businesses expect automation and receive a document instead.

An owner asks ChatGPT to book appointments and gets a helpful set of instructions for booking appointments, which is not the same thing.

The workaround is to separate the two jobs. Use the chat tool for drafting and thinking, and use a purpose-built integration when the task is to take an action. Answering calls and capturing enquiries is a system, not a prompt.

12. It cannot replace legal, tax or medical judgement

Advice that carries professional liability, regulatory consequence or health risk belongs with a qualified person. ChatGPT can explain a concept, but it cannot know your circumstances, hold a licence or carry responsibility for being wrong.

A business uses a chat answer about single touch payroll to decide what to report, and the error surfaces in an audit.

The workaround is a firm boundary. Use it to prepare questions for your accountant, lawyer or doctor, never to replace their answer. The output of that conversation is useful; the tool’s opinion is not advice.

Do other AI assistants share these limitations?

Claude and Gemini share most of these limitations, because they are built on the same kind of language model technology and inherit the same fundamental weaknesses. What differs is the surrounding product: pricing, plan limits, data controls, the tools each one connects to, and how each behaves on a specific task.

Switching tools therefore fixes almost none of the twelve problems. A different assistant will still state wrong facts confidently, still not know your business, and still need a human reviewing the output. Choose on data controls and integration rather than on which one sounds smarter.

Is ChatGPT safe for Australian businesses?

ChatGPT is safe for an Australian business when the rules are written before staff start using it, because the risks are manageable and the obligations are clear. The Privacy Act 1988 applies to any use of AI that involves personal information (Federal Register of Legislation), and the OAIC’s guidance gives a direct starting position for organisations.

A one page internal rule covers most of the exposure, and it needs to be specific enough that a new staff member can follow it without asking.

A practical second step is to tell customers how you use AI, in plain language, on your website. It costs nothing and it removes the suspicion that comes from finding out by accident.

Does ChatGPT content hurt your SEO?

Using ChatGPT does not hurt your SEO by itself, and unreviewed, repetitive content does, whatever produced it. Google has stated that using automation or AI to generate content is not against its guidelines when the purpose is to manipulate rankings rather than to help people, and that its focus is on the quality and usefulness of the page (Google Search Central, 2023).

The practical test is the same one Google publishes for any page: is this content helpful, reliable and created for people, and does it demonstrate first hand experience (Google Search Central).

In practice that means treating the draft as a starting point and adding what only you can add: your own jobs, your own numbers, your own photographs and your own opinions. Pages that do that are indistinguishable in quality from any well written page, because that is exactly what they are.

When should you use a specialist tool instead?

Use a specialist tool whenever the job is to take an action, to answer a customer in real time, or to work with data you would not paste into a public chat. ChatGPT is the right tool for drafting and reshaping text, and the wrong tool for anything that has to happen without a person driving it.

• Answering the phone after hours is a phone system with AI, not a chat window, which is what AI phone answering is built for
• Following up leads automatically belongs in your CRM and email system, wired to fire on a trigger
• Working with customer records belongs in a system that already holds them, with access controls
• Drafting, rewriting, summarising and structuring is exactly what a general assistant is good at
• Quoting from your own rates is best handled by a template or quoting tool you control

The businesses that get value from AI are the ones that decide this split deliberately. What the benefits look like in practice is covered in the guide to the benefits of AI for Australian businesses, and the plain English explanation of the tool itself is in what is ChatGPT.

What do business owners ask about ChatGPT limits?

Can I use ChatGPT for customer emails?

You can use ChatGPT for customer emails if you remove identifying details before pasting anything in and have a person read the reply before it is sent. The tool is well suited to rewording, toning down an angry response or turning notes into a clear message. The risk is not the writing, it is the customer data that gets pasted alongside it.

Is ChatGPT accurate enough for business information?

ChatGPT is accurate enough for structure, wording and general explanation, and not accurate enough for facts you intend to publish without checking. Names, dates, statistics, prices and legal references are the categories where it most often fails. Treat any specific figure as unverified until you have opened the source that states it.

Does ChatGPT store the information I type?

Conversation handling depends on your settings and plan, and the controls have changed repeatedly since launch, so check the current data controls in your own account rather than relying on any guide. For business use, choose a plan with the administrative controls your obligations require, and record the decision. Do not assume a setting is set the way you expect.

Will ChatGPT replace my staff?

ChatGPT replaces tasks rather than whole roles in most Australian small businesses, and the tasks it removes are the repetitive ones. What changes is where the hours go, with more of them on judgement, client contact and work that needs a licensed or experienced person. Businesses that try to remove a role entirely usually find the customer experience gets worse.

Can I trust the sources ChatGPT gives me?

You should not trust a source ChatGPT gives you until you have opened it, because invented references and misattributed quotes are common. Even when a link is real, the claim attached to it may not appear on the page. Open every citation, find the sentence you are relying on, and cite the source you actually read rather than the one you were given.

How do I stop staff pasting customer data into ChatGPT?

Stop staff pasting customer data by writing a short, specific rule, moving the team onto a plan with administrative controls and giving them a safe alternative they can actually use. A blanket ban tends to push the behaviour out of sight. The most effective version names exactly what must never be entered and shows the anonymised substitute to use instead.

The limitations of ChatGPT are not a reason to avoid it, and they are a reason to keep a person accountable for everything it produces. Write the one page rule, teach the anonymising habit, verify every number at its source, and the tool becomes useful without becoming a liability. More practical guides are on the DigiFix blog, the wider range sits under digital marketing services, and you can read about the team behind the work.