AI image generators turn a sentence into a picture in seconds. This starter guide explains how they work, where they go wrong, and the copyright and privacy checks to make before your first prompt.
This guide is general information, not technical or legal advice. If you’ve lost money or personal information, contact your bank, IDCARE on 1800 595 160, or report it through ReportCyber at cyber.gov.au.
Quick summary
- AI image generators are trained on millions of captioned images and create new pictures using a noise-to-picture process called diffusion.
- Free tiers come with limits such as generation caps, watermarks or lower resolution, so check what you get before paying for a plan.
- Australian copyright requires a human author, so images a tool generates on its own are generally unprotected, and there is no exception permitting AI training on copyrighted works without consent.
- The OAIC says privacy obligations apply to any personal information you input into commercially available AI products, so keep faces, documents and contact details out of prompts.
- Regulators including the eSafety Commissioner and the OAIC have flagged deepfake images and videos as a key risk of generative AI.
Type a sentence into an AI image generator, click a button, and a picture appears seconds later. The results can be genuinely handy for cards, blog illustrations and rough design ideas, but the tools fail in predictable ways and they raise copyright and privacy questions worth settling before you type anything. This guide explains what these programs actually do, how to write your first prompts, what free and paid access involves, and the checks to make about ownership and your personal information.
What an AI Image Generator Actually Does
An AI image generator is a program trained on enormous collections of pictures, each paired with a written description. The datasets behind this kind of training are vast: ImageNet, a well-known research dataset, holds more than 14 million links to images, and commercial models are trained on hundreds of millions of captioned images gathered from the web. The machine does not see pictures the way you do. An image is broken into pixels, the pixels are stored as numbers, and the model works with patterns in those numbers.
Most current tools rely on a technique called diffusion. During training, the system is shown images with random visual noise added, and it learns to strip that noise away step by step. Run in reverse, the model can start from pure noise and gradually shape it into a picture that matches your description. Older approaches, known as generative adversarial networks or GANs, could also produce images but were fairly limited in what they could create, and computer-generated art has a longer history than people assume, with rule-based experiments dating back to the 1960s.
Four terms cover most of what you will read:
- Model: the trained program itself.
- Training: the process of feeding it millions of captioned images so it learns patterns linking words to visuals.
- Prompt: the description you type.
- Diffusion: the noise-to-picture method most tools now use.
If you reach these tools through a phone or tablet app rather than a website, run the usual checks first. Our guide to downloading apps safely covers what to look for before you install anything.
Your First Prompts: Describing What You Want
A prompt is just a description, and vague descriptions produce generic pictures. A workable pattern for beginners has three parts: the subject, the setting or action, and the style.
| Prompt part | What it does | Example |
|---|---|---|
| Subject | the main thing in the picture | a kelpie puppy |
| Setting or action | where it is and what it is doing | asleep on a sunny verandah |
| Style | the overall look | watercolour illustration |
Stack the parts into one sentence: “a kelpie puppy asleep on a sunny verandah, watercolour illustration.” Then regenerate a few times. Each run produces a different result, so the third or fourth attempt often beats the first, and changing one element at a time lets you see what actually made the difference. If the tool offers built-in styles or presets, start with those before writing long descriptions, because they do a lot of the work for you.
Free Access Versus Paid Plans
The well-known tools can all be tried for free, and free access is the right place to start. Expect limits of some kind: a cap on how many images you can generate in a day, lower resolution, slower speeds, or a watermark. The details differ between services and change often, so check four things before investing time: the generation cap, whether output carries a watermark, how long the service keeps your prompts and images, and whether the terms allow commercial use of what you make.
Paid plans lift many of those restrictions, but a subscription deserves the same scrutiny as any other recurring charge. Our walkthrough of buying an AI subscription sets out what to compare first, and the guide to free trials and subscriptions explains how to avoid charges you never meant to sign up for. If you prefer easing in slowly, our timeline for trying an AI chatbot for the first time takes a staged approach that works just as well for image tools: read the policies, try the tool on real tasks, and decide about paying after a week or two of genuine use.
What These Tools Are Good At, and Where They Fail
These generators are strong at turning a rough description into a plausible picture quickly. They produce variations on a theme, handle stylised illustrations well and are useful for exploring visual ideas before you commit money or time to anything. Because they blend patterns from millions of images, they excel at imagery that looks right for its genre.
That same quality explains the failures. Anatomy is a well-known weakness, with people and animals sometimes gaining extra limbs or misplaced features, and flaws of that kind persist even in the newest models. More broadly, a generator produces images that look plausible rather than images that are accurate. It has no understanding of facts, so asking for a specific real place, product or person tends to return something approximate. If accuracy matters, use a photograph instead. Researchers and regulators have also raised concerns about bias and disinformation, since a model reflects patterns in whatever it was trained on.
Misuse deserves a mention too. The Digital Platform Regulators Forum, whose members include the ACCC, the ACMA, the eSafety Commissioner and the OAIC, has published a working paper on the type of generative AI that produces images, audio and video, and it names deepfake images and videos, meaning realistic fabrications of real people or events, as a widely cited risk. Australian creator organisations have also pointed out that using a person’s name, image or likeness without permission likely breaches other Australian laws. The practical rule is simple: do not generate pictures of real people without their agreement, and never make images intended to mislead anyone.
Who Owns the Images You Generate?
Two separate questions sit here: whether anyone owns the output, and what you are allowed to do with it.
Under Australian law, a work only attracts copyright when a human author has put their own independent intellectual effort into creating it. A picture that a tool produces on its own generally sits outside copyright altogether, which means nobody can claim exclusive rights over it. What you may do with the output is governed by the tool’s terms of service, so read those before using generated images commercially, because restrictions differ between services. The Attorney-General’s Copyright and Artificial Intelligence Reference Group is looking at how the law might change, but there is no settled answer yet.
The training side is more contentious. Generative AI models are built on very large datasets scraped from the web, which often includes copyrighted material collected without the rights holder’s consent. Australian copyright law offers no exception that would let someone copy protected works for AI training or automated data analysis without that consent. If a system was built on infringing material and your output can be traced back to it, you may receive a takedown request or an infringement notice.
Change may arrive through licensing as much as through law reform. An alliance of 22 creator organisations, representing a creative economy it values at $90 billion and 5.9 per cent of the Australian workforce, argues that models should be trained only on lawfully obtained material used with permission. In one sign of how permissions are being organised, Copyright Agency extended its Business Licence from May 2025 to cover the use of Australian newspaper articles in generative AI tools in workplaces. None of that changes what an individual hobbyist can do today, but it explains why the rules keep moving.
Privacy: What Happens to Your Prompts and Uploaded Photos
Prompts are not private thoughts; they are data sent to a company’s servers. The Office of the Australian Information Commissioner, Australia’s privacy regulator, makes clear in its guidance that privacy obligations apply to any personal information input into commercially available AI products, and it has published separate guidance for organisations using AI products and for developers who train generative models. That guidance is written for organisations, but it points to habits individuals should copy.
In practice:
- Keep personal details out of prompts. Addresses, phone numbers, tax file numbers and children’s names have no business in a prompt box.
- Be careful with photo uploads. Many tools let you upload an image to restyle or edit it. That upload hands the company a copy, and depending on the settings it may be used to train future models. Photos of other people deserve particular caution.
- Check the default training setting. Look for a toggle covering whether your prompts and images improve the model, and switch it off if you would rather they did not.
- Check retention. A prompt deleted from your history may still sit on a server for some time.
Be careful
Uploading a photo of someone else, especially a child, gives the company a copy of that image and may feed it into training systems. Before you upload a face that is not yours, consider whether that person would agree.
A Short Checklist Before Your First Prompt
- Read the privacy policy for three things: whether prompts and images train the model by default, how long they are kept, and whether you can opt out.
- Read the terms of service to see what you may do with the output.
- Set your own upload rule: no faces without permission, and no documents containing personal information.
- Treat your output as unprotected. Assume no copyright attaches and check the terms before any commercial use.
- Start free, and only pay once you know you will use the tool regularly.
One small step to begin: open a free tool in your browser, spend five minutes with its privacy policy and terms of service, then write a single three-part prompt with a subject, a setting and a style, and see what comes back.
Your questions answered
Q: Can I use AI-generated images for my small business?
A: Read the tool’s terms of service first, because permitted uses differ between services. In Australia, output generated purely by the tool generally will not attract copyright protection, so you cannot stop others from using similar images.
Q: Do image generators keep my prompts?
A: Usually yes, at least for a time, and some services use them to train future models. Check the privacy policy for retention periods and look for a setting that lets you opt out of training.
Q: Why do generated faces and hands often look wrong?
A: Anatomy is a known weakness of these models, and people or animals can end up with extra limbs or misplaced features. Regenerating the image a few times often produces a better result.
Q: Is it legal to generate images of real people?
A: It is best avoided. Regulators have flagged deepfakes as a significant risk of generative AI, and creator organisations point out that using a person’s name, image or likeness without permission likely breaches other Australian laws.