The five mistakes that sink most AI content creation are starting with the tool instead of a brief, confusing volume with taste, letting the brand drift from one output to the next, shipping without a human edit or a rights and disclosure check, and measuring how much was made rather than what it achieved. None of them is a software problem. Midjourney, Runway, Veo, Sora, Adobe Firefly and Higgsfield can all produce genuinely good work in 2026. They will also produce mediocre work at remarkable speed, and they have no opinion about which one you ship.

Audiences, unfortunately, do have an opinion. A 2024 NielsenIQ study that combined surveys of more than 2,000 people with EEG measurement found participants intuitively spotted most of the AI-generated ads and rated them more annoying, boring and confusing than conventional ones; even the ads judged high quality produced weaker memory activation.[1] That is the price of getting this wrong, and it is why the answer is rarely less AI. It is more direction.

Mistake One: Starting With the Tool Instead of the Brief

Most bad AI content begins with someone opening a generator and typing. The prompt becomes the brief, which means the brief is whatever occurred to that person at four o'clock on a Thursday. The model fills every gap with its own defaults, and its defaults are the statistical middle of everything it has seen. That is where the glossy, faintly familiar, instantly forgettable look comes from.

A proper creative brief names the audience and the single thing you want them to feel or do. It says where the asset will live, because a 9:16 story frame and a six-sheet poster are different problems. It sets the references, the colour world and the casting, and it lists what must never appear. The prompt then becomes a translation of decisions already made rather than a substitute for making them.

This is the order we work in on AI content production: creative vision first, tool selection second. Choosing between Veo and Runway for a shot is a craft decision, and a much easier one when you know what the shot is for.

Mistake Two: Confusing Volume With Taste

Generation is now cheap enough that the bottleneck has moved. Making a hundred variations costs almost nothing. Knowing which one is good, and having the nerve to bin the other ninety-nine, is the expensive skill.

Coca-Cola's 2025 holiday campaign is the instructive case. A behind-the-scenes video described a team of five specialists producing and refining more than 70,000 video clips in 30 days, and much of the online reaction to the finished ads reached for words like "soulless" and "creepy".[2] Whatever you make of the work, output at that scale did nothing to settle how people would feel about it.

The fix is to treat selection as a stage in its own right, with time in the schedule and a named owner. Review outputs against the brief rather than against each other, because the best of a weak batch is still weak. And give the final call to someone with real visual judgement, ideally a person who spent years framing shots and grading footage before any of these platforms existed.

Mistake Three: Letting the Brand Drift Between Outputs

Every generation is a fresh roll of the dice. The founder's face shifts between frames, the product's proportions change, the colour grade wanders from warm to clinical within one carousel. Each asset looks fine alone. Together they look like five brands sharing an Instagram account.

The tools have improved markedly here. Midjourney's style references and style codes let a team capture a visual aesthetic and reapply it across new generations.[3] Runway says Gen-4 can keep characters, locations and objects consistent across scenes from a single reference image.[4] Google added Ingredients to Video to Veo 3.1 in January 2026, designed to hold characters, objects and backgrounds steady from scene to scene.[5] Higgsfield's Soul ID trains on 20 to 80 photos of one person, though Higgsfield itself tells users to expect "clearly the same person" rather than pixel-perfect accuracy.[6]

That caveat is the lesson. Consistency comes from the system around the tools: an approved reference library, locked style codes and prompt fragments, a defined colour grade applied in post, and real product photography wherever the product itself has to be accurate. Write it down. An AI brand guide that lives in one person's head is a single point of failure with a nice eye.

Mistake Four: Skipping the Human Edit and the Rights Check

Raw output from any model is a first draft. Beyond retouching and grading, it needs two checks plenty of teams skip: is it labelled correctly, and do we have the right to use it the way we plan to?

Disclosure Has Moved From Etiquette to Obligation

In the EU, Article 50 of the AI Act has applied since 2 August 2026. Deployers must disclose deepfake content, and providers of generative systems must mark outputs in a machine-readable format, with a grace period to 2 December 2026 for systems already on the market. For evidently artistic, creative or satirical work the obligation narrows to disclosure, but it does not disappear.[7]

The platforms got there first. YouTube requires creators to disclose realistic altered or synthetic content and can remove videos or suspend Partner Program access for those who consistently don't.[8] Meta labels ads made with its own generative tools, places the label next to "Sponsored" when an ad includes an AI-generated photorealistic human, and says it will detect third-party AI tools through industry-standard signals.[9] TikTok requires creators to label realistic AI-generated content and automatically labels uploads carrying C2PA Content Credentials.[10] Sora videos launched with a visible watermark and embedded C2PA metadata,[11] and Veo output carries Google's imperceptible SynthID watermark.[5] The label is increasingly going on whether you add it or not, so it is better to decide how it reads.

In Australia, the National AI Centre's November 2025 guidance on AI-generated content is voluntary, but it is explicit that the Australian Consumer Law still applies: representations must be truthful and not deceptive, including where AI was involved in making them.[12] An AI-rendered product that looks better than the thing in the box is a consumer law problem before it is a creative one. In the US, the FTC's 2024 rule on reviews and testimonials specifically prohibits AI-generated fake reviews, which should give pause to anyone producing synthetic "customer" testimonial ads.[13]

Ownership Is Thinner Than You Think

The US Copyright Office's January 2025 report concluded that prompts alone do not make someone an author, while human selection, arrangement and modification of AI output can be protected.[14] The courts have held the same line. In Thaler v. Perlmutter the DC Circuit upheld the human authorship requirement in 2025, and the Supreme Court declined to hear the case on 2 March 2026.[15] The commercial consequence is uncomfortable: an asset made purely by prompting may be something a competitor can reuse freely. The human edit is also your IP position.

The input side matters too. Disney and Universal sued Midjourney in June 2025, alleging it trained on their characters and could reliably reproduce them on request.[16] Prompting for recognisable characters, named artists' styles or real people's faces is a risk the brand carries, whatever the vendor's terms say. For assets where legal exposure matters most, Adobe says Firefly is trained on licensed content such as Adobe Stock plus public domain material, and offers IP indemnification on qualifying plans.[17] Match the tool to the risk profile of the asset, put the rules in an AI policy, and bring your own counsel in early, because jurisdiction matters.

Mistake Five: Measuring Output Instead of Outcomes

AI content reporting tends to celebrate the wrong numbers: assets produced, cost per asset, days saved. Those are production metrics. They tell you the machine is running, not whether anyone cared.

Creative is where advertising money is made or lost. NCSolutions' 2023 analysis of nearly 450 consumer packaged goods campaigns found creative drove 49% of incremental sales, well ahead of brand, reach, targeting or recency.[18] If creative quality accounts for roughly half of what advertising achieves, a cheaper asset that performs worse is a false economy with a very tidy spreadsheet.

So measure what the content was for: incremental lift, hold rate on video, recall, cost per acquisition, qualified pipeline. Then point AI's volume at the job it genuinely does well, which is structured testing. Sixty variants organised around clear hypotheses about the hook or the offer will teach you something; sixty random ones mostly teach the algorithm how to spend your budget. That discipline is how we run creative testing in paid media, and every round should sharpen the next brief.

Taste Is Still the Moat

The models will keep improving, probably faster than any article can track. The next release will not fix any of the five mistakes above, because the model did not cause them. They are human decisions with human fixes.

Brands that treat AI as a production capability with a creative director attached will keep making work people remember. The rest will keep adding to the flood. If you would rather be in the first group, talk to us.

Frequently asked questions

What are the most common mistakes brands make with AI content creation?

The five that come up most are starting with a tool instead of a creative brief, confusing volume with quality, letting brand consistency drift between outputs, publishing without a human edit or a rights and disclosure check, and measuring assets produced rather than business outcomes. None of them is fixed by switching tools. They are fixed by creative direction, a documented system and honest measurement.

Do I have to label AI-generated images and videos in my marketing?

It depends on where the content runs. In the EU, Article 50 of the AI Act has applied since 2 August 2026 and requires deployers to disclose deepfake content. YouTube and TikTok require creators to label realistic AI-generated content, and Meta applies AI labels to ads in some cases. In Australia, the National AI Centre's labelling guidance is voluntary, but the Australian Consumer Law still prohibits misleading or deceptive representations, including those made with AI.

Can a brand own the copyright in AI-generated content?

In the United States, the Copyright Office has concluded that prompts alone do not make someone an author, while human selection, arrangement or modification of AI output can be protected. The courts have upheld the human authorship requirement, and the Supreme Court declined to revisit it in March 2026. Rules differ by jurisdiction, so brands should take local legal advice on assets that matter commercially.

Which AI image tool is safest for commercial use?

Adobe states that its Firefly models are trained on licensed content such as Adobe Stock and on public domain material, and it offers IP indemnification on qualifying enterprise plans, which makes it a common choice for higher-risk assets. Whichever tool you use, avoid prompting for recognisable characters, real people's likenesses or named artists' styles, and document the rules in an internal AI policy.

How do you keep AI-generated content on-brand?

Build a system around the tools: an approved reference library, locked style references or codes, trained character models where a recurring face is needed, a defined colour grade applied in post, and real product photography wherever accuracy matters. Then put every output through human review against the brief before it ships.

References