In 1984, the Macintosh shipped with MacPaint and a 512×342 pixel screen. Within two years, working designers at newspapers and ad agencies were being told that desktop publishing software — PageMaker, released in July 1985 — would eliminate their profession. The New York Times ran pieces speculating that typographers and paste-up artists were finished. What actually happened: the industry absorbed a hundred thousand new practitioners who never could have entered the field under the old cost structure. The tool changed who could design; it did not eliminate the need for judgment about what to design.
In 2022 and 2023, Midjourney, Stable Diffusion, and Adobe Firefly arrived under an almost identical set of headlines. Creative directors at agencies including Wieden+Kennedy and Pentagram began publicly working through the same question PageMaker raised: not "will AI replace designers?" but "which parts of design does AI do cheaply enough that the profession must reorganize around the parts it cannot touch?" The Adobe MAX keynote of October 2023 demonstrated Generative Fill removing and replacing complex image elements in seconds — tasks that previously required an hour of skilled Photoshop work.
This course maps where those tools sit in a real design workflow today: what they accelerate, where they break, and how a designer who understands their failure modes uses them faster and better than one who simply prompts and hopes. We will be specific about capabilities, honest about limitations, and practical about the decisions that remain irreducibly human — taste, client judgment, and the ability to know when the output is wrong.
If you finish every module, here's who you become:
In September 2023, the brand agency Collins — whose clients have included Spotify, Facebook, and the Obama Foundation — posted a detailed internal case study describing how they had begun using Midjourney v5 during the moodboarding phase of brand identity projects. Creative director Ben Crick noted publicly that the firm was not using AI to generate final deliverables but to collapse the visual research phase from two to three days down to a single afternoon. The client still received handcrafted vector work. What changed was how long it took to align the team's creative direction before that work began. Concepting accelerated; craft did not disappear. This is the pattern that repeats across every documented adoption story in professional design in 2023–2024: AI is being absorbed into the earlier, more exploratory layers of the workflow — not the final production layers that carry legal, brand, and quality risk.
The term "workflow" is used loosely. For our purposes, a design workflow is the ordered set of decisions a project passes through from initial brief to final deliverable. The British design firm Pentagram has described their process in several public interviews as moving through four broad phases: problem definition, visual research, concept development, and execution and refinement. This four-phase model is not universal, but it is representative enough to be useful.
Each phase has a different decision type. Problem definition is primarily verbal and strategic — the designer is asking what the artifact needs to accomplish. Visual research is associative — gathering references, building mood, identifying the visual territory the solution should inhabit. Concept development is generative and evaluative — producing candidate directions and selecting among them. Execution is technical — translating chosen directions into production-ready files using whatever tools the medium demands.
AI tools in 2024 are strong at the middle two phases and weak at the outer two. They can flood visual research with options and they can generate raw concept material quickly. They are poor at interpreting the nuanced strategic language of a brief, and they are unreliable for final production work where pixel-precise output, brand compliance, and legal clearance matter.
If you treat an AI image generator as a final-output tool rather than a research-and-concepting tool, you will consistently encounter problems: incorrect typography, inconsistent proportions, uncleared intellectual property references, and outputs that look plausible on-screen but fail at production scale. The tool is not broken — it is being applied at the wrong layer.
By 2024, design-relevant AI tools had consolidated into three functional categories, each mapping to different workflow layers.
The April 2023 issue of Eye: The International Review of Graphic Design published a roundtable of practitioners that included designers from NB Studio, dn&co, and Browns. Several participants independently converged on the same observation: AI produces outputs that are statistically average — they are plausible because they are composed of the most common visual relationships in the training data. The problem is that strong brand identity work is categorically not the most common visual relationship. It is the precise, unexpected, slightly uncomfortable solution that could only have come from a deep reading of a specific client's situation.
This is not a temporary limitation waiting to be solved by a larger model. It is a structural feature of how generative models work. They optimize for plausibility, not appropriateness. A designer using AI for visual research must therefore know how to identify and discard the statistically average outputs and recognize the edge cases worth developing — a skill that requires the same trained eye it always required.
AI tools have entered the design workflow at the visual research and early concepting layers. They accelerate options generation. They do not replace the strategic reading of a brief, the trained judgment required to select among options, or the technical precision required for final production. Knowing which layer you are at determines which tools are appropriate.
You will interrogate an AI assistant about where AI tools fit in a real design workflow. The goal is to practice the kind of critical, layer-aware thinking the lesson introduced — not to accept whatever the AI says, but to push back, ask for specifics, and test the boundaries of its claims.
In early 2023, the design director at Huge Inc. — the digital experience agency — described in a public LinkedIn post a controlled internal experiment. Two senior designers were each given the same brief for a fintech company's visual identity moodboard and asked to generate twenty reference images using Midjourney v4. Designer A produced outputs that were largely generic: dark palettes, abstract data visualizations, generic glass-and-steel corporate photography aesthetic. Designer B produced a cohesive set of references that the creative director described as "actually usable." The only difference was that Designer B had spent forty-five minutes writing what amounted to a visual specification document before touching the tool — breaking the brief into tone words, explicit exclusions, reference artists, compositional parameters, and color temperature language. The AI was identical. The output quality was not.
A prompt is a specification. It is not a magic spell, a search query, or a casual request to a colleague. The confusion around prompting largely stems from the fact that AI tools respond to casual, vague language — they produce something — which creates the illusion that the tool has understood you. It has not. It has made a statistical inference about what you probably wanted based on the most common associations with the words you used.
When Midjourney processes the prompt "logo for a coffee brand", it draws on the statistical distribution of what coffee brand logos look like in its training data. The result is almost always a brown, circular, serif-typography arrangement — because that is the center of mass of the training distribution. If your client makes specialty cold-brew targeting urban cyclists, that center-of-mass output is almost certainly wrong. The prompt needed to specify: what it should feel like, what it should not feel like, what visual references are adjacent to the target, and what constraints apply.
Research from the Runway ML team (published in their 2023 model documentation) and practical analysis by designers at Fantasy Interactive and Instrument converges on four components that separate useful from useless visual prompts.
Adobe's internal research team, writing in the October 2023 Adobe MAX documentation for Firefly, noted that professional designers using the tool were averaging six to twelve prompt iterations before arriving at a usable output direction. Novice users were averaging two iterations before abandoning the tool with the assessment that it "doesn't work." The difference is not the tool. It is the expectation. Professional designers understand that the first output is diagnostic information — it tells you which parameters you under-specified, not that the tool has failed.
This maps directly to how experienced art directors write briefs: the first draft of a brief is not a final document, it is a probe. You send it out, see what interpretations it generates, and refine based on what the responses reveal about your own unspoken assumptions. Prompting an AI image tool is the same process, compressed into minutes.
Prompt quality determines output quality more than any other variable. A functional visual prompt contains a subject with context, style anchors, technical parameters, and negative constraints. First outputs are diagnostic, not final — the iteration process is where skill expresses itself. Designers who understand this are consistently more productive with AI tools than those who treat prompting as guesswork.
You'll bring a design brief to the AI — real or hypothetical — and work through building a strong visual prompt for it. Start with a weak, vague prompt. Ask the AI to critique it and suggest improvements across all four components: subject/context, style anchors, technical parameters, and negative constraints. Then iterate.
On May 23, 2023, Adobe released Photoshop 24.6 with Generative Fill to all Creative Cloud subscribers. Within seventy-two hours, the Photoshop subreddit — which has over 3.5 million members — was flooded with demonstrations of complex compositing tasks completed in seconds: a car moved from a driveway to a mountain road, a product photographed on a white background extended to a full environmental scene, a portrait background replaced without masking. The obvious takeaway was speed. The less obvious takeaway — noted by photographers and retouchers in the same thread — was that every single output contained AI-generated content that was not photographed, not licensed, and potentially not legally clearable for commercial use under the terms of Adobe's then-current content credentials framework. By August 2023, Adobe had updated its Content Authenticity Initiative guidance to include explicit recommendations for disclosing AI-generated content in commercial work. The speed was real. So was the liability.
Embedded AI features differ from standalone generative tools in one critical way: they operate within the context of an existing file. When Photoshop's Generative Fill extends a background, it synthesizes new pixels that must match the existing image's color profile, lighting model, and perspective. When Illustrator's Generative Shape Fill creates a pattern from a text prompt, it operates within the vector constraints of an already-established design system. This contextual integration is genuinely valuable — the outputs are better constrained than a standalone model prompt — but it does not eliminate the fundamental issue that the AI is synthesizing content it did not photograph or draw.
Figma's AI features, which began beta rollout in June 2023, operate differently. Rather than generating visual content, Figma AI assists with layout logic — auto-populating design components with variable content, suggesting responsive layout adjustments, and renaming layers based on content analysis. This is a narrower intervention: it is automating the mechanical execution of decisions the designer has already made, not generating new visual content. The professional risk profile is correspondingly lower.
Adobe's Content Authenticity Initiative (CAI), co-founded with Twitter and the New York Times in 2019 and now encompassing over 2,000 member organizations, developed the C2PA standard (Coalition for Content Provenance and Authenticity) for embedding metadata about how an image was created. As of 2024, Photoshop can attach C2PA credentials to any file that contains AI-generated content. For commercial work — advertising, packaging, editorial illustration — clients and publishers are increasingly requiring this metadata. A designer who has been using Generative Fill without understanding the disclosure implications may face client disputes retroactively.
Setting aside the liability questions, three documented embedded AI use cases have demonstrated consistent professional value across multiple agencies in 2023–2024.
The retouching and compositing skills that defined professional Photoshop work have not been eliminated by Generative Fill — they have been reoriented. A designer who understands frequency separation, luminosity masking, and perspective matching can direct Generative Fill with precision and identify when its output is wrong. A designer who does not understand these concepts will not be able to tell when the AI has produced a technically plausible but geometrically incorrect result — a shadow falling in the wrong direction, a reflection that doesn't correspond to the new environment, a perspective mismatch between original and generated content. The AI produces the pixels; the trained eye catches the errors.
Embedded AI features are genuinely valuable for background extension, asset variation, and prototype comping. They carry real professional risk in commercial contexts where content provenance matters — risks that Adobe itself has formalized via C2PA metadata requirements. The skills required to use them correctly are not eliminated; they are reoriented toward directing and quality-checking AI outputs rather than producing every pixel manually.
You'll work through a real-world scenario where a client is asking you to use Generative Fill for a deliverable. The AI advisor will help you assess the content provenance risk, identify whether the use case fits one of the three low-risk categories, and draft language to clarify the situation with your client.
In November 2023, Pentagram partner Marina Willer gave an interview to Dezeen in which she described what her studio was and was not using AI for. The studio had integrated AI into early visual research; it had not integrated it into any final deliverables. When asked whether she worried about AI displacing designers, Willer made a distinction that has since been widely cited in design education: "AI is very good at producing things that look like design. It is not yet able to produce things that mean like design." She was pointing at the gap between visual competence — the ability to produce an aesthetically coherent image — and semantic intelligence — the ability to understand what visual choices communicate within a specific cultural, client, and competitive context. A Pentagram identity for a cultural institution is not just visually sophisticated; it carries a precise argument about what that institution is and wants to become. That argument requires understanding the institution deeply. AI can decorate an argument it has not understood.
Analysis of practitioner interviews, agency case studies, and design education discourse in 2023–2024 converges on four competencies that AI tools do not replicate — and that define a designer's professional positioning in an AI-augmented field.
Equally important is an honest account of what AI is devaluing. The ability to produce a technically competent stock illustration — an isometric scene, a flat-style icon set, a generic infographic — is now accessible to anyone who can write a prompt. The rate card for this category of work dropped substantially between 2022 and 2024. This is not a crisis; it is a signal. Designers who have been competing primarily on technical production speed in mid-tier visual categories are facing real market pressure. Designers who have been competing on the four competencies above are not.
The AIGA Design Census 2023, conducted in partnership with Google, found that in-house design teams were most interested in hiring for strategic design thinking and cross-functional collaboration — both of which require the irreducibly human competencies listed above. Production skills were listed as a hiring factor by significantly fewer respondents than in the 2020 census. The market is adjusting faster than most designers' self-assessments.
Every lesson in this course is oriented toward making you a better director of AI tools, not a more skilled AI operator. The practical skills here — workflow mapping, prompt construction, embedded tool assessment, system design thinking — are all in service of the four irreducibly human competencies. AI proficiency is not the destination; it is what clears the path to work that requires judgment.
The designer's competitive position in an AI-augmented field is defined by strategic brief reading, evaluative selection, cultural translation, and systems thinking — none of which AI replicates reliably. The skills being devalued are generic production competencies. This course is oriented toward developing the judgment to direct AI tools, not just operate them.
You will map your own design practice against the four irreducibly human competencies. The AI will help you identify where you are strongest, where your work currently overlaps with what AI can now do cheaply, and what specific moves would shift your positioning toward higher-judgment work.