Most people ask AI for ideas the same way they ask a search engine for results: describe the problem, receive the output. This works fine when you want a competent answer in the center of the distribution. It fails badly when you want something that departs from it. The problem isn't that AI can't generate surprising ideas — it's that the default prompt gives it no instruction to try. It optimizes for plausibility, not novelty.
Divergent thinking is the discipline of deliberately expanding the idea space before collapsing it. The goal is not to produce more ideas — it's to produce ideas from more directions, so that when you eventually converge, you have genuinely different options to choose between. Two structured techniques make AI useful for this: SCAMPER, which systematically forces every standard category of transformation, and random-stimulus prompting, which uses an unrelated input to interrupt habitual framing.
This module teaches you to use both. You'll run a real problem through SCAMPER with AI as your operator, and then test what changes when you inject a random concept into the same problem. By the end, you'll have a repeatable method for generating idea diversity rather than idea volume.
A marketing team at a mid-sized software company has run the same basic campaign structure for three years: product spotlight in Q1, customer story in Q2, feature launch in Q3, year-in-review in Q4. Each year the campaigns perform reasonably well. Each year the team says they want to try something different. Each year the output looks like a refined version of the previous year.
This year, a team member decides to use AI to generate new campaign concepts. She opens a chat and types: "Give me five creative campaign ideas for a B2B productivity software company targeting mid-market finance teams."
The AI produces five ideas. They're polished and well-reasoned. Two are variations on customer storytelling. One is a thought-leadership content series. One is an interactive ROI calculator campaign. One is a webinar series. All five are things the team has either done before or consciously rejected in prior planning sessions. None would surprise anyone who'd worked in B2B SaaS marketing for two years.
This isn't a failure of the AI — it's a failure of the prompt. The AI was asked for "creative" ideas without any structure forcing it off the expected path. It produced the statistical center of B2B SaaS marketing ideas because that's what the prompt optimized for. The team is disappointed but doesn't understand why. They asked for creativity and got competence.
What the team needed wasn't a better AI — it was a better prompt structure. SCAMPER would have forced the AI to ask: "What happens if we eliminate the content entirely? What happens if we put the campaign format to a completely different use? What if we reverse the audience relationship?" Those questions lead somewhere different. An unconstrained "give me ideas" prompt does not.
Divergent thinking requires a method, not a mindset. Telling yourself to "think creatively" produces the same ideas with more confidence. Systematically applying a checklist of transformations forces your reasoning into territory it wouldn't reach on its own. AI is an excellent operator for SCAMPER precisely because it can apply each transformation quickly and without the self-editing that makes human divergence collapse back to the familiar.
Divergence fails when it stays inside the category. If you ask "what else could this be?", your brain searches within the same conceptual neighborhood it already knows. Structured divergence works differently — it forces operations that cut across categories rather than traversing them. SCAMPER is the most durable of these structures because its seven operations are genuinely distinct transformation types, not variations on a theme.
Each letter is a distinct operation. Running all seven forces the AI (and your thinking) into spaces that no single "give me ideas" prompt reaches:
After SCAMPER, the random stimulus provides a second divergence pass with a completely different mechanism. You give the AI a random noun or concept — ideally from an unrelated domain — and ask it to find connections to your problem. "You're designing a campaign for finance software. Now incorporate the word 'migration.'" The AI's job is to find those connections; your job is to notice which ones feel like genuine reframes and which are just wordplay.
Research on creative problem-solving shows that random stimuli produce ideas with measurably higher originality scores than structured brainstorming alone — not because the random word is directly useful, but because it forces a new entry point into the problem space. The randomness is the mechanism, not a side effect.
SCAMPER produces more ideas. That's not the goal. The goal is to produce ideas from different directions so you have genuinely distinct options when you converge. These three distinctions determine whether you actually get that.
A session that produces 40 variations on "a content series with a finance angle" has low divergence despite high volume. Genuine divergence means the Substitute result looks nothing like the Reverse result — they don't belong to the same category. When evaluating SCAMPER output, the question isn't "how many ideas?" but "how many different directions?" If most of your seven operations produced ideas that feel like siblings, the operations weren't really distinct — you applied the same frame to each one.
The temptation is to choose a random word that feels relevant. "Finance software" → "efficiency" or "data." That's not random — that's the same category with a label change. The point is to use something genuinely unrelated: a physical object, an animal, a natural phenomenon, a historical event. "Finance software + glacier" forces genuinely different processing than "finance software + optimization." The AI's job is to bridge the gap; your job is to let the gap exist rather than closing it in advance.
The most common failure mode is editing SCAMPER output as it arrives — dismissing the Reverse result because it seems impractical, tightening the Eliminate result back toward the status quo. Divergence requires suspension of judgment during the generative phase. You write everything down. You evaluate nothing until all seven operations (plus the random stimulus) are complete. Convergence happens after you have the full spread in front of you, not while you're still building it.
You'll apply all three principles in the lab — the AI will run your problem through SCAMPER one operation at a time, then run a random-stimulus pass, and then ask you to evaluate which directions produced the most genuine divergence.