Anyone Can Make an AI Image. Using One Well Is the Real Skill.

Every article image on this site was made with AI, and every one of them says so. Labelling them took seconds. Deciding which tool to use, how to brief it and what we would never generate took longer.

When AI imagery first reached marketing teams, it meant learning one specialist tool and its prompt codes. Now image generation sits inside software most organisations already pay for. The hard questions have moved from syntax to choice, brand and trust.

Start With the Tool You Already Have

Most New Zealand organisations already have an AI image generator, whether they know it or not. Microsoft Copilot can create and edit images. So can ChatGPT and Gemini. Canva builds generation into the design tool many marketing teams open every day.

A tool your organisation already licenses has been through procurement, sits under an enterprise agreement, and has data terms someone has read. A specialist tool signed up for on a personal card has none of that. Many banking, government and professional services teams are Copilot-only for exactly this reason.

So the first question to ask is “what are we allowed to use, and does it do the job?” If the answer is yes, start there. When a specific job needs a specialist, put that tool through the same approval process as any other, using the same data and privacy checks.

Choose by Job, Not by Brand

If a model learned from images used without permission, the reputational fallout can land on the brand that published the output, not just the vendor. Adobe says Firefly was trained on licensed content such as Adobe Stock and on public-domain material, and it offers IP indemnification on qualifying business plans. Microsoft extends its Customer Copyright Commitment to output from paid commercial Copilot services, provided you keep its guardrails switched on. Read the terms for the plan you actually have; indemnity is rarely as broad as the headline.

Is Midjourney still worth using?

Midjourney put AI imagery on most marketing teams’ radar, and it still produces some of the most distinctive aesthetic results available. What has changed is that it is no longer the only serious option, and for everyday business visuals it is rarely the first one to reach for.

Two practical points if your team uses it. Midjourney’s plans page says a company making more than US$1 million in gross revenue a year must buy the Pro or Mega plan. And Stealth Mode, which keeps your images private, is only available on those two plans. Unreleased campaign concepts made on a cheaper plan may be more visible than you think.

From Prompt Formulas to Conversation

Early image tools rewarded people who learned their codes: aspect-ratio flags, version numbers, stylisation settings. Most of that has gone. Current tools understand plain language, edit an existing image when you ask, and can work from a reference photo or your brand assets.

The skill that matters now is briefing, the same skill that gets good work from a designer or a photographer.

Write a brief, not a prompt

• Purpose: Where will it appear, and what must it do? A LinkedIn header, a slide that explains a concept, a campaign mood board.
• Audience: Who needs to recognise themselves or their world in it?
• Subject and setting: What, where, and in what light?
• Composition and format: Wide or close, landscape or square, space left for a headline?
• Brand constraints: Colours, style, and what must never appear.

Then refine in conversation

First brief:

“A header image for a LinkedIn post about hybrid work at an Auckland professional services firm. A modern office with floor-to-ceiling windows overlooking the Waitematā Harbour, morning light, two people collaborating at a shared table. Natural, documentary photography style. Wide landscape format with clear space on the left for a headline.”

Then:

“Make them look mid-conversation, not posed. Warmer light.”

“Swap the laptop for printed plans, and remove any visible logos or text.”

Three short rounds usually get further than one perfect prompt. Keep the briefs that work in a shared library, so the next person starts from a proven version rather than a blank box.

Where AI Imagery Earns Its Place

• Campaign concepts and mood boards: Test ten visual directions in an afternoon before committing budget to a shoot.
• Presentation and report visuals: An image that explains your idea beats a stock photo of a handshake.
• Product and packaging visualisation: Show stakeholders a concept before investing in prototypes or photography.
• Editing existing photos: Extend a background to fit a new format or remove clutter, while keeping the edit honest (see below).
• Localised variants: Adapt one approved visual for different regions or audiences without starting again.

Don’t Replace Designers. Make Them Faster

The teams getting the most value use AI for range and speed early, then hand the chosen direction to a designer for brand compliance and finishing. AI makes twenty options. People decide which one is right, and make it good.

AI imagery does not belong anywhere your audience reasonably expects a real photograph of a real thing. Staff profiles, customer stories, event coverage and product shots that say “this is what you will receive” should stay real.

Rights, Rules and Disclosure

This is where most AI image advice goes quiet, and where the real risk sits. None of it is a reason to avoid the tools, but it is a reason to agree a simple policy before someone needs one under pressure. What follows is general guidance, not legal advice.

Copyright: New Zealand is unusual

Most countries require a human author before copyright exists. New Zealand’s Copyright Act 1994 is one of the few that provides for computer-generated works: where there is no human author, the author is the person who made the arrangements necessary for the work to be created (section 5(2)(a)).

Whether that person is the user who wrote the brief or the company that built the model has not been tested in court, and the Act is under review. Treat ownership of AI images as possible but uncertain. Don’t build an asset you will need to defend, such as a logo, purely from AI output.

Commercial terms differ by tool and plan

Free tiers and personal plans often carry different rights from business plans, as the Midjourney revenue rule shows. Before images go into paid campaigns, someone should check the terms for the exact plan your team is on, not the one in the vendor’s marketing.

Consumer law: an image can mislead as easily as a claim

The Fair Trading Act 1986 prohibits misleading conduct and representations, and the Commerce Commission judges an advertisement by the overall impression it creates, images included. An AI-generated product shot that makes a meal look bigger, a room look newer or a result look better than customers will get carries the same risk as an over-retouched photograph. Fine print does not fix a misleading image.

Real people need real consent

Never generate or alter an identifiable person’s likeness without their consent, whether they are staff, customers or public figures. AI-generated “people” in testimonials, reviews or case studies imply endorsements that never happened. That is deceptive, and it is a consumer-law risk as well as an ethical one.

Māori imagery needs cultural authority

Image models reproduce Māori motifs without understanding them. Do not generate campaign imagery depicting tā moko, korowai, koru or other taonga without express cultural authority. The wider question of who controls Māori data and representation is covered in our piece on AI and Māori data sovereignty.

Keep customer and staff photos out of unapproved tools

Uploading a photo of an identifiable person to an AI tool is sharing personal information, and the usual privacy rules apply. Before anyone uploads a customer photo, a staff headshot or an unreleased product, work through our five-question data checklist.

One disclosure rule

The test we use and teach, and the same one we apply to AI video: if your audience would feel deceived when they find out, disclose it before they do.

• Always disclose: Images that could pass as real events, real people or real products, and any AI imagery in news-style or editorial content. Synthetic people in testimonials should not be used at all.
• Usually disclose: Campaign and hero images realistic enough to be mistaken for photography.
• Low risk: Obviously illustrative or abstract images, though a small label costs nothing.

We label every article image on this site, including the one at the top of this page. It takes seconds, and it is the kind of habit our AI ethics guide recommends building from a team’s first day with AI.

Provenance labels help, but don’t rely on them

Some generators now attach provenance signals. Adobe applies Content Credentials, an open standard for recording how a file was made, to images fully generated by Firefly. Google embeds an invisible SynthID watermark in Gemini images, and OpenAI has announced it is adding SynthID alongside the Content Credentials metadata its images already carry. Metadata can disappear with a screenshot, so treat these as useful signals, not proof, and keep your own disclosure in a visible caption.

Frequently Asked Questions