Skip to today
Recode

Thursday, November 5, 2026 · Graphic Design

AI Can Draft the Layout, but Someone Still Has to Judge It

Noted by Orana Velarde · 666 reads

AI can generate a full layout from one prompt, but finishing it still depends on knowing what to fix. Here's where human design judgment matters most.

AI Can Draft the Layout, but Someone Still Has to Judge It
Visual depicting YouTube videos that teach graphic design skills to finalize AI outputs

Better AI drafts haven't made design knowledge less useful. They've moved the work somewhere harder to see. A tool can now produce a full layout from a single sentence, but the distance between that first pass and something you'd actually publish is covered by people who know what to look for. Figma CEO Dylan Field put it simply: "the better cogeneration gets, the more design matters."

The first draft is rarely the finish line

AI tools now cover a wide range of work: prototypes, presentation slides, one-page documents, marketing graphics, web pages, interactive HTML and applications. AI-native platforms like Lovable, along with general assistants such as Claude and ChatGPT, can make a complete design from a prompt. Visme's AI works the same way. The catch is that almost nobody ships the first result. Corrections come back through chat, inline comments and direct canvas edits, and each one requires you to know which problem you're solving.

Victor Dibia, PhD, tested Claude Design for his newsletter Designing with AI and concluded that the sensitive decisions around generated output still belong to the person publishing it. That's a useful way to frame the job. The tool assembles options quickly, but deciding what your company puts in front of customers remains a human call.

Several designers have landed in the same place. Product designer Aurélie Radom argues in a Medium article that "AI did not democratize design judgment. It democratized output." In her Hey Designer newsletter, she goes further: "Taste cannot be delegated." That distinction matters. Anyone can generate a page now. Knowing whether it works is a separate skill, and it's the one that's getting scarce relative to demand.

Field told a16z in 2023 that moving from an AI first draft to a finished product is "kind of hard and usually takes a team." Models have improved since then, but the sentiment still holds. The remaining work is judgment-heavy, and judgment rarely comes out of a single prompt.

Vocabulary is the real bottleneck

Most changes in AI tools happen through language rather than manual editing. That shifts the problem. If you can't name what's wrong, you can't ask for a precise fix, and you may not notice the problem in the first place. A UC Berkeley study on AI design calls this the "Vocabulary Gap," the situation where users lack the precise terms for their intent.

The fix is less mysterious than it sounds. Words like hierarchy, alignment, contrast and white space let you describe a problem in a way the model can act on. Compare two requests. "Make it look better" gives the tool nothing concrete. "Increase the headline size, left-align the body text and add more space between sections" names three changes you can check when the result comes back. The second prompt also gives you a yardstick. You can see whether each instruction was followed, which is more than you can say for the first.

This is also why design vocabulary pays off even if you never open a design program. It turns vague dissatisfaction into a fix you can request, verify and reuse.

Why so many AI designs look alike

Generative tools learn from existing work, so their outputs tend to drift toward patterns that already appear often. Over time, a particular layout or type pairing becomes the default, and it shows up everywhere. The Berkeley researchers describe this as a "Convergence Trap," where the system "defaults to generic, 'polished' styles." Changing the requested fonts or colors doesn't always break the pattern, which is why so much AI-generated visual content shares a recognizable look.

Figma's Oscar Dumlao makes a related point on the Figma blog: "AI trained on the past can't invent the future." Models remix what they've seen. They're good at recombining familiar pieces, and weaker at deciding what a specific brand should look like that nobody has seen before.

The practical implication is that brand character has to be pushed into the output on purpose. Specify the palette, the type scale, the grid and the spacing rules, and treat any unrequested stylistic choice as suspect. In Visme's State of B2B Sales Content 2026 report, 31.1% of B2B leaders said generic-looking output was a reason they use AI less. Teams that solve that problem tend to be the ones who gave the model more specific direction, not the ones who waited for a better model.

More people are making visuals now

Design work is spreading beyond design departments, and that's a big reason the skills gap matters. Among marketers who use AI, the share using it for image creation rose from 19% to 43% in one year. According to Figma's 2026 AI Report, the share of developers doing design work rose from 44% to 60% in the same period. In the same report, 90% of respondents said design is at least as important as it was before AI, and nearly six in 10 said it's more important.

Visme's Brand Execution Gap 2026 report points to a skills problem that predates AI. 47.4% of brand managers said colleagues know the brand basics but lack the design skills to apply them. Half of respondents (50.3%) reported non-designers making spacing, layout and alignment mistakes, and 49.3% saw font mistakes. When everyone creating business content understands the basics, there's less back and forth between departments and production moves faster. That's the strongest argument for teaching these skills broadly rather than leaving them to a specialist.

Keep a person in the loop

Teams that publish AI-assisted work well treat the output as a first draft and review it before it goes live. According to Ahrefs' 2026 survey, only 7% of AI-using marketers publish mostly unedited AI content. The Designer Fund's 2026 AI in Design report found that only 5% of design leaders are putting less emphasis on execution quality. Read together, those figures suggest that most people still see the final review as the part that justifies the whole process.

Visme's Brand Execution Gap report offers a clear picture of where the friction is. Brand owners who pulled back from AI cited too much cleanup (33.5%) and output that didn't match their brand (19.5%). Only 1.3% of creators said a template is ready to use as it is, and 83.2% still apply brand elements by hand. Those numbers say less about whether AI works and more about how much finishing it still requires.

The larger risk isn't a bad font choice. Errea, a design agency from Spain, put it this way in its Substack: "The problem is not about using AI, but letting it think and decide for us." If a team stops developing its own ideas and simply accepts the first layout it's given, the output drifts toward sameness no matter how polished it looks. Visme's own position is that AI features should work alongside manual control, and its drag-and-drop editor is built for the final adjustments a generated draft still needs.

Seven checks to run on any AI draft

You don't need to be a trained designer to catch most of what AI gets wrong. A repeatable review pass helps, and each check maps to a skill you can learn in a few minutes.

Hierarchy. Squint at the page or blur it. The first thing that stands out should be the message you intended. Scale, weight and color decide what dominates, and AI often gives every element similar weight. Ask the tool to rank elements and make one of them clearly dominant. The basics of visual hierarchy cover how proximity and figure-ground contrast reinforce that ranking.

Balance. Look at where the eye lands. Generated layouts often pile weight on one side. Decide whether you want symmetry or an asymmetric balance, then name it in your prompt so the model knows the target.

Alignment and spacing. Check shared left edges, consistent margins and equal gaps between repeated items. Small drift is one of the most common AI problems, and it's easy to fix once you name it. Ask for a single column grid or a shared baseline, then verify each element against it. Our layout design guide covers grids in more depth.

Color and contrast. Generated palettes often look alike, and text contrast is frequently weak. A useful accessibility baseline is a 4.5:1 contrast ratio for body text against its background, which is the level commonly used as a minimum for normal-size text. Check each text and background pair, then adjust the failing ones rather than eyeballing them. Save the final palette as a brand kit so future designs reuse it.

Typography. Limit yourself to two typefaces, one for headings and one for body. Set a clear type scale, comfortable leading and a reasonable line length. For body text, you'll often see recommendations near 16 pt with about 1.5 line spacing and roughly 65 characters per line. Check every text element for sizes and weights that fall outside the scale, since that's where brand drift usually starts.

Decoration and white space. AI loves rounded cards, gradients and filler shapes. Keep only the elements that guide the eye or explain an idea, and remove the rest. Often the fix is subtraction. If a layout feels crowded, delete the least important element first and see whether the page improves.

Data. Charts are where AI can mislead you. In his review, Victor Dibia found that the model chose a visually interesting chart over an accurate one. Check the chart type against the question, confirm the axes and values, and make sure the takeaway is stated plainly. Start a bar chart's value axis at zero, and label the key comparison directly.

Two more things belong in every review: whether the visual type fits the message (photo, illustration, diagram or chart), and whether any AI-generated image contains errors in text, hands or perspective. Those problems are easy to miss at a glance and obvious to anyone who's looked at a few hundred generated images.

Where to start learning

Visme's Digital Design Foundations Course is one structured option. It consists of 12 short videos, each covering one skill and ending with prompt examples you can adapt. The timed videos run just under an hour in total, so the time commitment is modest. Signing up through the course page also gives access to practice resources and a certification you can add to your LinkedIn profile. The videos are also available on YouTube if you only want the lessons.

The lessons cover purpose and visual style, balance, hierarchy, alignment, color theory, typography, shapes and icons, visual communication, data visualization, consistency, white space and type as a design element. Work through them in order if you're new to the material. If you already have some experience, jump to the lessons that match the problems you see most in your own drafts.

It's also worth weighing the job market honestly. The US Bureau of Labor Statistics projects graphic designer employment to fall 2% between 2025 and 2035. Entry-level roles are likely to feel the most pressure, while senior work that depends on judgment and direction still needs people. That's a reason to build skills that go beyond production, such as critique, brief writing and knowing when a draft should be rejected.

What to do with your next AI draft

Pick a draft from any AI tool you already use and run it through the seven checks. Write three prompts that name the specific changes you want, rather than asking for general improvement. Compare the revised result against your original requests. You'll quickly see which problems the model handles well and which ones need your hands on the editor.

My read is that the valuable skill is shifting from making every element by hand toward auditing and directing machine-made work. The people who do that well will be the ones who can say exactly what's wrong and why it matters for the audience. The tools will keep getting better at the first draft. What they can't do yet is decide what your brand should sound like, and that decision still rests with you.

Source: Orana Velarde · visme.co

Also on the calendar