AI creative tools sell a one-step promise. Drop in a product shot and watch it appear in a dozen new environments. Type a detailed prompt and get back something that looks, at first glance, like a finished editorial image. When budgets are tight and timelines are shorter than they used to be, that promise is understandably seductive.
The trouble is that the one-step result has become recognizable. We call it AI slop: work that’s smooth, symmetrical, technically competent and strangely anonymous. Slop is often the first accepted answer, delivered with enough “can’t-quite-put-my-finger-on-it” polish that nobody stops to ask whether it is the right one.
At Level, we don’t think of AI as a creative process in itself. We have more than 30 creative professionals, and they use these tools in different ways: during discovery, concepting, storyboarding, production, iteration, or not at all for stretches of a project. Ask ten artists how they use Photoshop and you will get ten different answers. AI should be no different. The process remains divergent and human. The technology is tools inside it.
When everyone has access to the same tools, what separates the work is taste, craft, experience, the thinking behind it, and the feeling it creates. Increasingly, those are the things that matter most.
What AI slop looks like
The obvious mistakes are disappearing. Hands are better. Faces are more convincing. Text is improving. Yet the work can still carry that hard-to-name synthetic quality: flawless skin, generic typography, balanced compositions, perfect atmosphere, a surface-level beauty that feels as though it has been averaged from thousands of other beautiful things.
That isn’t an accident. Image models are designed to move toward plausible, desirable results. Human feedback helps keep bizarre or unusable outputs in check, but it also nudges the model toward what has already been judged safe and good. The default is therefore not chaos. Increasingly, the default is competence.
And competence has its own sameness. To get somewhere more specific, a creative has to push against the model’s inclinations rather than simply accept them.
You can see the effect working in reverse now. Real photography gets accused of being AI-generated because it looks too perfect. People have learned the visual language. When audiences start reading polish itself as evidence of fakery, the issue is no longer just aesthetic. It becomes a question of trust.
The challenge for brands and creatives is finding that voice and deciding what we want these tools to do for us, rather than letting the tools decide what the work becomes. The goal should always be to create amazing work. AI is simply one of the many ways to get there.
The real process behind one AI-generated image
To show the difference between a generated image and a directed one, we built a demonstration around a fictional Level soft drink. It is intentionally not client work, and we would not generate a realistic person like this for a final client deliverable. The point was simply to test the assumption that a polished editorial image now comes from one very good prompt.
The prompt was:
“Ethereal editorial portrait on a misty beach, pale light and soft horizon, blonde woman in iridescent pink ruffled garment and bonnet, translucent textures. She is sipping from a black LEVEL can, pressed against a black horse.”
It is a perfectly serviceable prompt. It names the subject, wardrobe, mood, light, setting and product. The model gives you all of those nouns. What it doesn’t give you is judgment.
The first result is polished in exactly the wrong way: porcelain skin, hair that feels arranged rather than lived in, a backdrop that reads more like an illustration than a photograph, and a product whose details become unstable under scrutiny. Nothing is catastrophically wrong. That’s precisely the problem.
The stronger image came from treating generation as material, not an answer. The process looked like this:
- Build the product first. We start with a dedicated product prompt and render the can on its own, where its label, typography, proportions and materials can be controlled before a larger scene complicates things.
- Give the product physical credibility. A 3D-style refinement pass establishes believable lighting, shadow and volume so the object can survive compositing later.
- Explore broadly. Midjourney becomes a moodboarding tool: beach, wardrobe, horse, atmosphere, camera, light. This is where abundance is useful. We can look at more directions and reject more of them.
- Choose a base image. The model generates options; a person decides which one is worth pursuing. That selection isn’t a technical step. It is creative direction.
- Composite deliberately. The rendered can is placed into the chosen frame and refined through additional realism passes. Again, the system returns options. Again, someone chooses.
- Upscale before finishing. Tools such as Magnific give the image enough resolution and texture to withstand the final refinement.
- Finish in a conventional design tool. Color, detail, typography and realism are handled with the control expected of commercial work. This is also where we undo some of the machine’s instincts: bring back texture, break symmetry, correct anything that feels too clean or too resolved.
- Deliver the final high-resolution output.
No single tool covers that chain, and that’s the point. The best results often come from what Carly described in the session as ‘Frankensteining’ tools together, using each for a particular strength while keeping human taste and decision-making in the loop throughout.
AI has changed the economics of exploration. It lets a team see more versions, test more directions and throw away more work before committing. But it has not removed the need for someone to know what should be thrown away.
What this looked like on a client project
A more useful example came from a campaign we produced for Hudson Yards, including Edge and Vessel. The original photography and video were created for summer and performed well. When winter arrived, the client didn’t have the budget for another large shoot. The fallback was an older image library that no longer matched the aesthetic or quality of the summer work.
Instead of generating a winter campaign from scratch, we went back to unused frames from the original shoot: real photographs, real people, real source material. AI became a way to extend what already existed rather than replace it.
We changed summer outfits into winter ones, adjusted light, tone and shadow, and replaced ice cream cones with Edge-branded cups. A rough version let us show the client that the idea could work before anyone invested in a painstaking finish.
Once the concept was approved, the shortcut ended. Every person in the frame was isolated and handled individually. Clothing was refined one subject at a time. Details were corrected in Photoshop. Commercial work has to stand up to scrutiny; there is a meaningful difference between making a quick generated joke for a colleague and making an advertisement that has to carry a client’s brand.
The campaign existed because the reshoot didn’t have to. That’s the kind of speed we care about: speed that makes good work possible, not speed that becomes the reason to lower the standard.
The photographer was compensated again, and so were the models, with their consent obtained for the AI-assisted use of the original work. AI created efficiency without erasing the value of the original artists. If a project’s economics only work by cutting out the people whose work made the source material valuable, the math is wrong.
What we tell clients before they sign
We are explicit about AI use from the beginning. It comes up in proposals, in contracts and in the conversation before work starts. There should be no reveal later.
We also label AI-generated material when it is being used for concepting or storyboarding. That isn’t just disclosure; it helps people give the right kind of feedback. A pencil sketch announces itself as unfinished. A generated comp can look deceptively complete, which can pull a room into debating a sleeve, a face or a piece of lighting when the real question is whether the idea works.
Transparency keeps the lens on the concept instead of the artifact.
We are equally deliberate about people. In general, we avoid generating realistic human subjects for client deliverables. When people are required, we prefer to begin with real photography, stock or footage and edit from there. The same principle applies more broadly: use AI where it solves a real problem, not simply because the tool can technically do the job.
How to evaluate AI-generated creative
The risk with synthetic-looking work isn’t merely that somebody recognizes the tool. The deeper risk is that it makes the brand feel as though it is trying to get away with something.
Advertising depends on trust. A brand is asking someone to believe the product, the service, the promise and, eventually, the exchange of money for value. When an image feels false or a product is shown behaving in a way that doesn’t feel real, the audience notices. They may not know exactly what is wrong, but the doubt arrives before the message does.
Research is beginning to support that intuition. In an Ipsos study conducted with faculty from Syracuse University’s Newhouse School, 20 ads across 10 brands were rebuilt as fully AI-generated versions and tested with 3,000 consumers. Many people could not confidently tell which ads were AI-made, yet the human-made work performed better on Ipsos’s measures of short-term impact and long-term brand health.
So before AI-assisted work ships, the questions should be painfully simple:
- Is this the best result we can get, or merely the first acceptable one?
- Is it actually better than standard stock or a conventional production solution?
- Has it been directed to this brand, or could the same image belong to anyone?
- Does it carry emotion? Does it tell a story?
- Is AI contributing something beyond speed?
Speed is real value when the alternative isn’t making the work at all: a campaign that would otherwise be unaffordable, more concepts tested before production, more useful versions in market. Speed becomes a warning sign when it is the only thing the technology contributed, because that often means the thinking was removed along with the labor.
AI is a tool. The work is still the work.
The machine doesn’t make the work and then hand a creative person the good version. A person decides the idea. A person decides what the product should look like. A person rejects most of the moodboard, chooses the frame, corrects the details and knows when the image has become too perfect.
That’s why the tooling conversation can become a distraction. We have tested dozens of AI products, and the stack keeps changing. Some tools disappear. Others are replaced in months. We recommend starting small: a few curious people, a pressure-tested set of tools, and enough flexibility to change course when the market moves. Don’t roll out a platform because it is trending. Pressure-test it against a real need. Keep the stack small enough to understand. The tools will churn; the team’s understanding of composition, typography, story, emotion and craft will carry across them.
This matters for junior creatives too. AI can generate hundreds of possibilities before someone has developed the visual history or confidence to judge them. Traditional creative education, references and hands-on craft still matter because they create the internal standard against which an output can be measured. You need a point of view before you can tell when a model is giving you somebody else’s average.
The most useful way to look at this moment is as an opportunity to reinvent ourselves. To question processes and assumptions we haven’t questioned before, experiment with new ways of making things, and imagine different futures for creative work.
What to ask for
Ask an agency for great work, not great AI work.
The second phrasing makes the technology the subject and quietly lowers the bar. It encourages everyone to admire how the thing was made rather than ask whether it says anything worth remembering.
Most audiences don’t care whether AI summarized the research, helped a storyboard become photorealistic or accelerated a special-effects pass. They react to the finished product. Does it feel legitimate? Does it build trust? Does it make them feel something?
That’s the standard the work has to survive.
Direction at the front. Craft at the end. Generation somewhere in the middle.
Learn more about our commitment to creative excellence in our on-demand session
Level’s creative leaders recently walked through how AI fits into an actual creative process: where it adds leverage, where it creates risk, and why taste and judgment still have to sit inside the work. The virtual session is available on demand. Watch the replay and if your team is wrestling with similar questions, let us know.
