Drawing with Ghost Hands: When GenAI Helps Architects — and When It Quietly Undermines Them

A sketch begins with a blank surface. That is the romantic version, anyway.

In a real design studio, the blank surface is rarely blank. It is crowded with precedent images, studio habits, tutor expectations, client language, spatial constraints, and the designer’s private suspicion that the first idea will be embarrassingly ordinary. Now add DALL-E 3 to the desk. Suddenly the first idea does not have to be drawn. It can be summoned.

That is the obvious promise of generative AI in architectural concept design: less friction, faster visual exploration, fewer students stuck staring at nothing. It is also the easiest way to misunderstand what is happening.

A recent controlled experiment on GenAI in architectural conceptual design does not support the clean story that AI simply “boosts creativity.” It shows something more useful and less brochure-friendly: DALL-E 3 helped lower-performing novice designers improve their revised design performance, but it did not produce a broad overall performance advantage over a conventional precedent-search workflow using ArchDaily. It also produced a significant negative effect on general creative self-efficacy, while cognitive load only improved for certain prompting behaviors rather than for GenAI use as a whole.¹

So the question is not whether GenAI helps architects.

The better question is: which part of the designer is it helping — and which part is it quietly replacing?

The paper tests GenAI against precedent search, not against doing nothing

The study design matters because it avoids a common fake victory in AI evaluation. The authors did not compare DALL-E 3 with an empty room, a tired student, and no external support. They compared it with a realistic design-learning baseline: browsing architectural precedents on ArchDaily.

The experiment involved 36 university students. Participants first created an initial architectural concept sketch without external tools. Then, in a revision phase, the control group used ArchDaily for precedent study, while the GenAI group used DALL-E 3 through ChatGPT 4. The task was to design the massing and geometry of a research laboratory building at WPI’s Gateway Park. Each phase produced a sketch on a digital drawing tablet. The outputs were judged by eight evaluators — four architectural professionals and four senior architectural students — using a rubric focused on clarity and legibility, complexity and detail, and overall visual communication.

The study then measured three things:

This is already a better frame than most “AI in design” arguments. Design quality alone is not enough. A tool can improve an output while weakening the learner’s sense of authorship. It can reduce manual effort while increasing evaluation effort. It can produce more images while making the designer think less clearly. The ghost hand may draw beautifully; the question is whether the human hand learns anything.

Mechanism one: GenAI scaffolds missing visual schemas

The clearest positive result appears only after segmentation.

Across the full sample, GenAI did not create a statistically significant performance advantage. In the revision phase, the GenAI group’s design performance mean was slightly higher than the control group’s — 2.903 versus 2.785 on a five-point scale — but the difference was not significant. The difference-in-differences estimate for design performance was positive at 0.583 standard deviations, but its 95% confidence interval crossed zero, from -0.476 to 1.643. In plain English: the direction looked favorable, but the study cannot treat this as a robust overall GenAI win.

Then the moderation analysis changes the story.

When participants were divided by initial design performance, tool condition interacted significantly with initial competency level for revision-phase design performance: $F(1, 32)=4.303$, $p=0.046$, partial $\eta^2=0.118$. Novice participants using GenAI performed significantly better than novice participants in the control group, while experienced participants showed no comparable GenAI advantage.

That pattern is the article’s first mechanism: GenAI helps most when the user lacks internal visual scaffolding.

Novice designers often do not just lack polish. They lack a mental library of forms, typologies, spatial analogies, and visual moves. They may understand the assignment verbally but struggle to convert that understanding into a workable massing sketch. In that moment, DALL-E 3 behaves less like a genius collaborator and more like an externalized visual memory. It supplies forms to react to. It gives the student something to accept, reject, combine, simplify, or redraw.

That is useful. It is also narrower than the marketing version.

For experienced designers, the same tool may not provide the same value because they already have internal schemas. They can sketch, search precedents, and evaluate form through trained judgment. A generic AI output can become friction: another image to inspect, another mismatch to correct, another “almost but not quite” proposal that interrupts the designer’s own workflow. The paper’s discussion makes this point directly: experienced learners may experience generic or context-insensitive outputs as distracting rather than beneficial.

This is not anti-AI. It is anti-average.

Averaging novices and experienced designers together hides the operational lesson. The tool is not equally good for everyone. It behaves like a scaffold when the user lacks structure, and like noise when the user already has a stronger structure than the model output.

For firms, that distinction matters. The business case for GenAI in design should not begin with “all architects will become faster.” It should begin with workflow diagnosis:

The point is not to deny senior designers access to GenAI. Please, spare us another policy written by someone who thinks creativity can be fixed by a software procurement rule. The point is that the expected value is different by user and task. A junior designer facing the blank page and a senior designer refining a spatial logic problem are not doing the same work, even when both are “designing.”

Mechanism two: better output can still weaken creative agency

The most uncomfortable result is not performance. It is confidence.

The paper distinguishes task-specific creative self-efficacy from general creative self-efficacy. This distinction is important. Task-specific confidence asks whether the participant felt their response to the task was creative. General creative self-efficacy asks broader questions: whether they feel good at coming up with new ideas, whether they have many good ideas, whether they have imagination.

Those are not the same psychological object.

The control group improved in general creative self-efficacy from the initial to the revision phase, while the GenAI group declined. The resulting difference-in-differences estimate for general creative self-efficacy was -0.543 standard deviations, with a 95% confidence interval from -1.068 to -0.017. Unlike the performance estimate, this interval did not include zero.

That means the negative general-confidence effect is not just a decorative caveat. It is one of the central findings.

The mechanism is subtle. GenAI may improve a student’s visible output while making the student feel less personally creative. Why? Because the locus of creativity becomes ambiguous.

When a student improves a sketch through precedent search, the process still feels recognizably human: look, interpret, adapt, draw. The precedent may inspire, but the student performs the transformation. When a student improves a sketch through DALL-E 3, part of the transformation becomes machine-generated. Even if the student curates the result, the emotional accounting changes. The output is better, but the student may ask: Was that mine?

This is where “human-AI collaboration” stops being a slogan and becomes a problem of authorship.

The paper’s discussion suggests that GenAI can shift the user’s internal benchmark for originality upward. When learners work beside a system that can instantly generate polished visual alternatives, their own ideas may feel smaller. Creativity becomes distributed across the human-AI system rather than located inside the individual. That may be philosophically defensible. It is not always pedagogically harmless.

In business settings, the same mechanism appears in a different costume. Junior designers, marketers, analysts, consultants, and product teams may use AI to produce stronger first drafts. Managers may see the improvement and declare victory. But the user may experience the workflow as a quiet demotion: “I am no longer the person creating the thing; I am the person selecting among machine proposals.”

That can matter for retention, learning, and long-term capability. People do not build expertise merely by shipping better artifacts. They build expertise by forming internal models of why a choice works. If AI assistance repeatedly improves outputs while reducing perceived authorship, the organization may get short-term polish and long-term dependency. Very efficient. Also very silly, in the expensive way.

The practical lesson is not to ban GenAI from early creative work. It is to design the workflow so the human contribution remains legible.

For example:

The distinction is not cosmetic. It determines whether AI becomes a prosthetic imagination or a training wheel that never comes off.

Mechanism three: GenAI does not remove cognitive load; it moves it

A common claim about GenAI is that it reduces effort. In some tasks, it does. In this study, however, overall cognitive load did not significantly differ between the GenAI and control conditions. The difference-in-differences estimate for cognitive load was -0.069 standard deviations, with a wide 95% confidence interval from -0.742 to 0.604.

So, no, the paper does not show that DALL-E 3 generally made conceptual design feel easier.

The prompt analysis explains why.

Sixteen GenAI participants had usable screen-recording data for prompt analysis. Their prompt behavior varied substantially: they produced between 3 and 26 prompts, with an average of 9.38 prompts, and word counts ranged from 20 to 295 words, averaging 110. Prompt count itself did not significantly correlate with cognitive-load change. Word count did not either.

That is the first useful correction. More prompting is not automatically better prompting. A longer prompt is not necessarily a more intelligent interaction. Anyone who has seen a 600-word prompt produce a beige blob already knows this, but it is nice when data joins the meeting.

The significant correlations came from prompt type. Two categories were associated with larger reductions in cognitive load:

Other prompt types, including simple generation, information gathering, general feedback, asking for feedback, and regeneration, did not show significant associations.

This is the third mechanism: GenAI reduces cognitive load only when the interaction has structure.

The difference is between using the model as a vending machine and using it as a design partner under supervision. “Draw a building” gives the model control and leaves the human with the task of interpreting whatever arrives. “Make the building shorter while maintaining the gabled roof” keeps the human in the loop as a director. “Too spiky, try again” may sound blunt, but it is operationally useful: it tells the system what visual communication failed and what needs correction.

The cognitive work does not disappear. It shifts from low-level production to interaction management: prompt formulation, output inspection, mismatch detection, and iterative refinement. For some users, that shift reduces burden. For others, it creates a new layer of work on top of the old one.

This is why prompt training is not a cute add-on. It is part of the tool.

A firm that gives junior designers access to GenAI without teaching interaction patterns is not “empowering creativity.” It is handing them a stochastic image machine and hoping taste, judgment, and authorship survive the interface. Hope is not a workflow.

The evidence is a three-part mechanism, not a single headline

The paper’s findings are easiest to misread if each result is treated separately. Performance: mixed. Self-efficacy: negative. Cognitive load: conditional. That sounds like a messy paper.

It is actually coherent.

This mechanism-first reading gives a better interpretation than the usual question: “Does AI improve design?”

That question is too blunt. It compresses different effects into one artificial verdict. GenAI can improve one user’s design performance, weaken another user’s confidence, and leave overall workload unchanged because the burden has merely moved from drawing to directing. The same tool can be scaffold, mirror, critic, distraction, and confidence trap — sometimes in the same 35-minute session. Conveniently, reality has once again refused to fit into a vendor slide.

What design firms should infer — and what they should not

The paper directly studies students, not professional firms. Still, the business implications are real if we translate carefully.

The first implication is segmentation. GenAI adoption should not be rolled out as a flat productivity layer. It should be mapped to user maturity and task type. Early ideation, rough visual exploration, and novice onboarding are plausible high-value zones. Late-stage spatial reasoning, technical coordination, and expert-led concept refinement may require more controlled, domain-specific tools.

The second implication is that AI training should not focus only on prompt syntax. The deeper issue is decision architecture. Users need to know when to generate, when to refine, when to stop, and how to convert outputs into human-owned design moves. A training module that teaches “better prompts” without teaching evaluation is just teaching people to ask for prettier noise.

The third implication is agency preservation. In creative organizations, confidence is not a soft decorative variable. It affects exploration, critique tolerance, initiative, and learning velocity. If AI-assisted workflows make junior staff produce better images while feeling less like designers, the firm has not solved the talent problem. It has hidden it under more polished concept boards.

A practical GenAI design workflow should therefore include three checkpoints:

This is not bureaucratic overhead. It is how the organization keeps the human skill from dissolving into the tool.

Tool builders should stop optimizing only for output abundance

For AI product teams, the paper suggests a less obvious product lesson: more generations may not be the highest-value feature.

The prompt analysis found that cognitive-load reduction was associated with iterative refinement and visual-communication feedback, not with total prompt count, word count, or generic generation. That points toward interfaces that help users steer and evaluate rather than merely produce more images.

For architecture and design tools, useful features might include:

• revision histories that preserve the human design path;
• explicit controls for massing, circulation, façade logic, and visual communication;
• side-by-side comparison modes that ask users to articulate selection criteria;
• “critique before regenerate” workflows;
• prompts or templates tied to design intent rather than style keywords alone.

The current text-to-image interface often encourages a slot-machine rhythm: prompt, inspect, regenerate, repeat. It feels productive because something is always appearing. But visual abundance is not the same as design progress. The user may be moving through images without moving through reasoning.

A better interface would make the designer’s judgment visible. It would treat generation as one step inside a design loop, not the loop itself.

The boundary: this is pilot evidence, not a universal law of AI creativity

The study has important limits, and they matter for how far we can generalize.

The sample size was 36 participants, with 17 in the GenAI group and 19 in the control group. The prompt analysis used 16 GenAI participants after excluding two because of missing screen-recording data. The task lasted one short design session, not a semester-long studio or a professional project. The tool was DALL-E 3 through ChatGPT 4, not a specialized architecture model with controllable spatial constraints. The outcome was a conceptual sketch judged on visual communication, not buildability, client fit, technical coordination, cost, sustainability, or code compliance.

These boundaries do not weaken the article’s main lesson. They prevent the wrong lesson.

The right interpretation is not: “GenAI helps novices, hurts confidence, and only good prompts reduce load everywhere forever.” The right interpretation is narrower and more useful: in a short architectural concept-design task, GenAI’s value depended strongly on user competency and interaction strategy, while its psychological effects were not uniformly positive.

For business use, this should inform pilot design. It should not be used as a universal procurement justification or a universal warning label.

The real lesson is not “AI creativity.” It is allocation of creative labor

The phrase “AI creativity” keeps dragging the conversation into the fog. It invites people to argue about whether the model is creative, whether the human is still creative, and whether the final image “counts.” Fascinating, perhaps. Operationally, often useless.

This paper points to a better frame: allocation of creative labor.

In architectural conceptual design, someone or something must generate visual possibilities, evaluate them, refine them, and communicate intent. GenAI changes who performs each part and how visible that contribution feels to the human user. For novices, offloading some generation can unlock revision. For experienced users, the same offloading may interrupt judgment. For everyone, the tool can quietly shift effort from making to managing. And if the workflow is poorly framed, the user may leave with a better sketch and a weaker belief in their own imagination.

That is not a failure of GenAI. It is a failure of lazy integration.

The business lesson is therefore simple, though not easy: do not deploy GenAI as a magic creativity layer. Deploy it as a conditional scaffold, a structured feedback instrument, and a carefully bounded visual exploration tool. Teach users how to direct it. Make their authorship visible. Measure confidence and learning, not just output quality.

The ghost hand can help draw. But if the designer stops feeling the pencil, the studio has bought itself a very elegant problem.

Cognaptus: Automate the Present, Incubate the Future.