Divergent Ideation with AI

Good creative work swings between two modes. Divergent thinking opens the space - generating many, varied, even wild possibilities without judging them yet. Convergent thinking closes it - evaluating, selecting, combining and refining toward a decision. Most weak ideation fails at the first: we converge too early, on the first workable idea, and never see the better one two directions over.

AI is a divergence engine. Where you might sketch five options before running dry, an LLM will list thirty spatial strategies and an image model will render fifty, tirelessly and without the ego-attachment that makes us defend our first idea. That volume and variety is exactly what divergent thinking needs, and it is emotionally easier to react to a wall of options than to conjure them from nothing.

The crucial discipline is to keep the modes separate. Use AI hard in the divergent phase - generous prompts, high variety, no judging - and then switch it off and converge with your own judgement. Blending them, or letting the AI both generate and decide, is how you end up with un-owned work. The human-in-the-loop from Module 0 reappears here in creative dress: AI floods the middle; you own the framing at the start and the selection at the end.

There is a real psychological benefit to offloading divergence to a machine. Our own idea generation is throttled by ego and fear: we hesitate to voice a strange option, we get attached to the first decent one, we tire. The AI has none of that - it will cheerfully produce the obvious and the absurd side by side, at no emotional cost, which lowers the stakes of exploring. That detachment is exactly what fixation needs. But it is a double edge: the AI has no taste either, so it cannot tell the brilliant option from the banal one. The division of labour is clean once you see it - the machine supplies fearless quantity, you supply the judgement it lacks.

Most people prompt AI to converge - 'give me the best layout' - which wastes its real strength. To break fixation, prompt for range: ask explicitly for many, deliberately different directions, and forbid it from settling.

Weak:   suggest a layout for a small art gallery
Better: Give me 12 genuinely different organising concepts for a
        small art gallery - vary the parti each time (linear,
        courtyard, spiral, mat, tower, etc.). For each: a one-line
        idea, the visitor spatial experience, and one risk.
        Do not converge; make them as different as possible.

The moves that force real variety: name the number (ask for 12, not 'some'), demand difference ('as different as possible', 'vary the parti'), and ask for the reasoning per option so you can judge, not just admire. You can also assign roles or lenses - 'now give me five that a landscape architect would propose', 'five that prioritise carbon', 'five that a child would love' - to push the model into corners it would not reach by default.

Image tools diverge differently: raise Midjourney's --stylize, use wildly varied prompts, or feed the model random pairings of references. The principle is identical across text and image - the more you push for difference, the more the AI earns its keep as a fixation-breaker. And keep judging switched off while you do it; a bad-sounding option often contains the seed of a good one, so collect first and criticise later.

Divergence is not only about generating fresh options; it is about operating on them. Two moves are especially powerful with AI in the loop.

Combination - genuinely new ideas often come from fusing two unrelated ones. Ask the AI to cross concepts: 'combine the courtyard scheme's calm with the tower scheme's compactness', or 'what would this cafe look like if it borrowed the circulation logic of a bazaar?'. The AI is a tireless combinatorial partner, happy to try pairings you would find awkward to force by hand.

Mutation - take one direction you like and ask for many variations that push a single dimension: 'keep this parti but make it progressively more introverted', 'the same house at five levels of material heaviness'. This explores the neighbourhood of a good idea, where the best version often hides just past the first. It is the antidote to settling: the first workable version of an idea is rarely its strongest, and mutation cheaply walks you to the better one two steps over.

These map onto real creative techniques - morphological combination, controlled variation - that AI simply makes faster. Crucially, the interesting outputs are usually the ones you steer toward with an unusual instruction, not the model's defaults. That is the whole trick to staying original with AI: the raw generation is commodity; the direction you push it is where your voice enters. A pointed, strange, brief-specific instruction pulls the model away from its averages and toward something that is recognisably a response to your problem.

A third move worth practising is constraint injection - deliberately handing the AI a limitation and asking it to design within it, because constraints are famously generative. 'Solve this with a single continuous roof plane', 'no room deeper than one window's worth of daylight', 'every space must borrow light from a courtyard' - odd, specific rules force the model off its comfortable path and often surface the most interesting directions. This is the opposite of the vague prompt that yields the average: the tighter and more particular your provocation, the less generic the result. It also keeps you in the driver's seat, because the constraints are your design thinking, expressed as instructions - the AI is merely exploring the space they define.

Now the honest warning, because it is the most important thing in this lesson. AI models are trained on vast averages and tuned to produce pleasing, probable output. Prompt them plainly and they return the mean of everything they have seen - which is competent, fashionable, and identical to what everyone else prompting plainly receives. Lean on that uncritically and two things happen: your individual work drifts toward a generic 'AI look', and, across the whole field, design starts to converge on a narrow band of model-flavoured sameness. This is homogenisation, and it is the real cost of careless AI ideation.

Worse, the trap is comfortable. The AI's output is polished enough to feel finished, so it is tempting to accept the average and move on - and each time you do, you practise editing instead of designing, and your own voice gets a little quieter. Fixation was a trap of too few ideas; this is a trap of too many samey ones.

The defence is active, not passive. Start from your references and obsessions, not the model's defaults. Push deliberately away from the first, most obvious result. Judge against the brief, not against what merely looks good. And - the step that matters most - reinterpret rather than copy: take the AI's provocation and redraw it in your own hand, on your own site, for your real client, so the final move is authored by you. Used this way, AI is a fixation-breaker that makes you more original. Used lazily, it is a homogeniser that makes you disappear. The difference is entirely in the judgement you bring - which is the whole argument of this course.

It helps to remember where your voice actually comes from, because AI cannot supply it. A designer's voice is built from the buildings you have stood in, the materials you have touched, the places you know, the obsessions you cannot shake - a lifetime of specific, embodied experience the model has never had. That is why the antidote to homogenisation is not a clever prompt but a rich inner library: the more you have genuinely seen and studied, the more distinctive the instructions you can give and the sharper your judgement of what comes back. So keep drawing by hand, keep visiting buildings, keep reading and looking - not despite using AI, but precisely so that when you use it, there is a real you steering. The tool amplifies whatever voice you bring; it cannot lend you one you have not built.