AI can widen your idea space or quietly collapse it into the average of everyone else's — the difference is sequencing and adversarial use. This module covers throwaway concept generation, workshops where AI never anchors the room, red-team critique panels, and competitive teardowns whose claims you can actually defend.
Rohan leads design for a Pune team building Trellis, a project-tracking product, and he runs a tight ideation workshop. For the Q3 "reduce status-meeting overhead" initiative he tries something new: to save time, he has FigJam AI generate fifteen seed ideas before the session and pastes them on the board so the team "doesn't start from a blank canvas."
The workshop feels productive. Stickies flow. But watch what actually happens: every human idea generated in the first thirty minutes is a variation on one of the AI's fifteen. The AI suggested digest emails, auto-generated standup summaries, and status dashboards — so the room produces flavors of digests, summaries, and dashboards. Nobody suggests the weird idea. Six weeks later Trellis ships an auto-summary feature that four competitors shipped the same quarter, because they were all drawing from the same model's sense of the average answer. The one genuinely differentiated concept — killing status updates entirely by letting stakeholders subscribe to live project events — surfaces three months later in a hallway conversation, from the quietest designer on the team, who admits she'd thought of it during the workshop and assumed it was out of scope "because it wasn't on the board."
That is anchoring. It is the defining failure mode of AI-assisted ideation, it is invisible from inside the room, and it has a two-word fix: humans first.
Used correctly, generative tools are a divergence machine. Prompt-to-UI tools — Figma Make first drafts, v0- and Lovable-class generators, Midjourney for moodboards — will produce three to five distinct concept directions for a problem in the time it used to take to sketch one. The value is real, and so is the trap, so the boundary gets written down before the first prompt:
A model's unprompted output gravitates toward the median of its training data — which means your competitors, prompting the same models with the same problems, are receiving approximately your ideas. Differentiation lives in what you bring: the constraint the model doesn't know, the user insight from Module 2, the deliberate rejection of the obvious answer.
Rohan's failure wasn't using FigJam AI — it was the order. The facilitation rule that fixes it is simple enough to fit on a sticky note: humans diverge first, AI joins after.
AI ideas go up first. Every subsequent human idea is unconsciously a variation — the room converges on the model's median before true divergence happens. Feels efficient, produces what four competitors also shipped. The quiet designer's weird idea never surfaces because the board's gravity says it's off-topic.
Silent individual ideation first, everything on the board. Then AI enters with a specific brief: gap-fill against what humans produced — "here are 40 ideas; generate 10 that share no mechanism with any of them." Then FigJam AI sorts stickies into themes and drafts the summary. AI amplifies the room's range instead of replacing it.
Beyond sequencing, workshop AI has two legitimately great jobs. Clustering and summarizing: sorting eighty stickies into themes in seconds and drafting a summary you edit — the drudge work of facilitation, gone. HMW seeding, on demand: if the room stalls, AI can reframe the problem statement six ways; you choose which reframe unlocks people. Both keep the human material primary.
Solo designers and small teams have a critique problem: not enough perspectives in the room, and the perspectives present are polite. AI solves the quantity and the politeness at once — if you configure it as an adversary rather than an assistant.
The red-team panel is four separate critique passes, each with a distinct lens and a mandate to attack:
Run each lens as its own prompt or project with your concept (screens and a one-paragraph intent statement) as input. Then the part that makes it professional rather than theater: you log every finding with an accept/reject decision and one line of rationale. The panel attacks; you decide. Rejected findings matter as much as accepted ones — the log proves you saw the objection and had a reason, which is exactly what a design lead or a client wants to see.
The decision log is the artifact. It shows curation, not obedience.
Lab 5 runs a provided concept through this exact panel and grades the decision log, not the concept.
Competitive analysis is now a two-engine job, and both engines require verification because both hallucinate differently.
Engine one: deep research. Deep-research modes (ChatGPT, Claude, Gemini, Perplexity-class) will crawl a competitor's public surface — product pages, docs, changelogs, pricing, reviews — and produce a structured report in minutes. It arrives confident, occasionally stale, and sometimes wrong about the one detail your strategy hinges on.
Engine two: your own eyes, multiplied. Walk the competitor's actual product yourself, screenshotting as you go, then run the screenshots through a vision model against your comparison rubric — onboarding pattern, empty states, pricing presentation, AI-feature entry points. This grounds the teardown in what the product does today, not what the training data remembers.
Both engines feed one artifact: a claims table, where every row carries the claim, the source URL, a freshness date, and a verification status. Top claims — anything that will influence a roadmap or a pitch — get verified by hand.
| Claim | Source | Fresh as of | Verified |
|---|---|---|---|
| Competitor X ships AI standup summaries on the $12 tier | pricing page URL | 2026-04-01 | Yes — screenshot |
| Their mobile app lacks offline mode | deep-research run | 2026-03-28 | No — verify before pitch |
| Onboarding requires calendar OAuth before value moment | own walkthrough | 2026-04-02 | Yes — recorded flow |
Two more critique workflows round out the module, both governed by the same rule: AI findings are a draft for your judgment.
Multimodal pre-critique on your own work. Before human critique, screenshot your screens and run them past a vision model with your team's rubric — hierarchy, contrast, alignment, copy tone, state coverage. Fix the lintable issues so critique time goes to judgment calls. Anchor the model with an explicit rubric; un-anchored vision critique regresses to generic advice ("consider improving visual hierarchy") that wastes everyone's time.
AI-assisted heuristic review as a pre-research pass. A vision model walking screens against Nielsen's heuristics and platform HIGs produces a plausible findings table with severities. Research on synthetic heuristic evaluation shows partial overlap with human evaluators — useful, not trustworthy. Treat it as prep: you validate every severity, discard false positives (there will be several), and treat the surviving findings as hypotheses for real usability data — never as a substitute for it. The false positives have a recognizable flavor: the model flags a deliberately minimal navigation as "insufficient wayfinding," calls an intentional friction point a usability violation, or invents a heuristic breach on a screen state it misread. Each discard teaches you the model's blind spots; keep a running list, because they're consistent enough to pre-empt in the next prompt. A heuristic review that skips human validation isn't a discount evaluation; it's a liability with a table format.
Visual exploration gets the same two-sided treatment as everything else in this module. The craft side is straightforward and genuinely liberating: generate wide — twenty moodboard directions instead of three, style mashups you'd never assemble by hand, reference imagery for domains you can't photograph — then art-direct hard, because the model's default aesthetic is everyone's default aesthetic. Saved style references keep exploration coherent across a project instead of resetting per prompt.
The governance side is not optional, and it starts on day one, not at delivery:
End every exploration cycle by having AI draft the decision trail — "we explored A, B, C; chose B because…" — from your file history and critique notes, then edit it. Ten minutes per cycle buys you the case-study narrative, the onboarding doc, and the answer to "why does it work this way?" eight months from now, for free.
Sequence beats speed: humans diverge before AI generates, AI attacks before humans critique, and every AI finding — idea, critique, or competitive claim — passes through a logged human decision. That's how AI widens your thinking without replacing it with the average of everyone else's.