We made 5,000 campaign ideas. Strategists kept the plain ones.

The basin is real, and deeper than you think

Every aligned language model has favorite ideas. Ask for 40 earned-media campaign concepts and 18% of them are the same idea — "humans vs. AI, blind test" — wearing different headlines. 57% sit in one emotional register. Clustered strictly for "same idea," a 40-idea batch holds roughly 14 actual ideas.

This isn't randomness failing you. Preference training teaches models that typical answers are good answers — researchers call it typicality bias — so the model learns to stay near the middle of everything it knows. The 2026 literature is close to unanimous that the collapse lives in the weights, and our measurements agree:

• Temperature does nothing. Max versus low temperature on the same model: 17 vs 14 distinct ideas. Noise.
• Verbalized sampling barely moves a frontier model. The paper-faithful version scored within noise of a random-constraints control, three replicates.
• Self-rewording helps the count, not the reader. Spec-first decomposition — have the model write 40 maximally-different one-line specs, then render each — reliably doubled distinct ideas. It led our scoreboard for two full rounds. Then a human read the output.

Wider isn't better: the viability trap

We had a strategist blind-read whole 40-idea batches and tick every idea he'd actually take to a client. Plain prompt: 33 of 40 viable. The spec-first "winner": 22 of 40. Cluster the survivors and both land near 12 usable, distinct ideas. The diversity machinery bought its spread almost entirely with ideas a human throws out.

Almost every method that widened the set did it by escaping the brief instead of escaping the basin.

A "provocation" instruction — render each idea as the strangest version of itself — made the trap vivid: machine-scored originality soared while human-scored viability collapsed to a 0.28 win rate. Surprising and un-runnable is not a campaign.

Update (September 2026): seven more readers, one conclusion moves

Because this result carried so much weight, we re-ran it as a panel: seven Fractl strategists blind-read the same plain and spec-first batches on both briefs — same sheets, same tick-every-viable instruction, no knowledge of which set was which. The viable-rate gap replicated cleanly: plain wins on usable ideas everywhere (readers averaged 23 vs 20 of 40 on the main brief, 29 vs 22 on the second). The yield conclusion moved. Keep only the ideas a majority of the panel ticked, cluster the survivors:

On the main brief, spec-first's survivors span 15 distinct ideas against plain's 12 — a modest real edge, not the wash a single reader saw, and it holds for six of the seven readers individually; on the second brief it's a wash. The mechanism is arithmetic, not taste: plain's 40 ideas only contain 14 distinct ideas to begin with, so its 25 survivors pile into the same clusters, while spec-first's 20 starting clusters survive a harsher cut with more of them still standing. Spec-first is still the wasteful way to widen a set — a quarter of its ideas die at review — but the spread that survives review is real spread. "Flat or worse" becomes "flat to modestly better."

The judges have their own basin

We tried to automate the human read. Four frontier models from three companies judged ~21,000 blind head-to-head pairs. Checked against 139 decisive blind human picks across two briefs:

The failure has a mechanism. Judges score pairs in isolation and never saturate: a human reading 60 pairs is tired of "the AI writes its own obituary" by pair 12; the judges rewarded it every single time. In one experiment, 13% of all judge-favored ideas on a moving-company brief were the same divorce-data premise. And the judges share the generators' training, so they share the generators' taste — a judge can't see its own blind spot.

One technique helped: showing the judge the candidate pool first and declaring its premises used up lifted originality agreement from ~50% to 63% — and to 77% on the pairs where two judge families agreed. Useful for triage; not good enough to replace a human. Every ranking in this study that matters was made by a person.

The $110 brain surgery

If collapse lives in the weights, edit the weights. Open-weight models ship in pairs — the raw pretrained "base" model (the whole distribution, unaligned) and the aligned "instruct" model — with identical architecture. So you can blend them, parameter by parameter. We interpolated Qwen2.5-72B at 30 / 50 / 70% base on a rented H200 and measured how many regions of idea-space each blend reached:

Withdrawn, October 2026: re-counted with the repaired judge, the 70%-base blend was no wider than the untouched model. The paragraph below is kept as first published; see "We ran it again" and corrections log item 7.

The dial is real. Every step toward base widened the ideas, monotonically, and the model still followed instructions at 70% base. It's a lever no API will ever expose. And it wasn't enough. The open model's basin was so much deeper than the frontier model's that maximum blend never caught a plain Sonnet prompt on the main brief.

Two more findings from the open-weights lane:

• Division of labor beats either model alone. Let the blend emit raw one-line premises and a frontier model render each into a finished idea. This harvest-and-render pattern beat direct generation for every configuration — and on the second brief the 70%-base harvest (27.9) beat the best pure-frontier method (22.4).
• You cannot predict which models are wide. One community model family is wide at 405B and collapsed at 70B; its successor is the exact reverse. Same lineage, opposite outcomes. Screening each candidate empirically costs about $1 with this instrument — guessing costs you the campaign.

We ran it again. Here's what survived.

After we found the counting problem (corrections log, item 6), we re-ran the open questions from this study from scratch, with the rules for calling a winner written down before we started. Eleven ways of generating ideas, each run several times against the same plain prompt, each batch counted by the repaired judge, and one full batch of forty from every method read blind by the same strategist, who marked each idea no, usable, or I'd pitch it.

Three things held up

Different models have different good ideas. Mix three unrelated models and you get about ten ideas per forty that none of our earlier runs had produced. Seven models, about the same. The catch is what the strategist saw: nearly every portfolio idea was usable, but only seven in forty were worth pitching. Mixing models buys range, not quality.

Seeds first, then write-ups. Ask a model for forty one-line seeds, hand each seed to a strong model and have it written up for the brief. With DeepSeek writing the seeds, almost every idea in the batch was new to us, and the strategist pitched 23 of 40. With Claude writing its own seeds, same pitch rate. Writing the seed up did not pull it back toward the usual ideas; we checked by writing the same seeds up two different ways and got the same spread.

You can't talk a model out of its habits by editing the brief. We hid our company name during generation and restored it afterwards: two or three new ideas per forty. We deleted the through-line requirement: no change at all, on either brief.

Two things moved

The "tensions" trick is a trap. Ask the model to name the five tensions in the brief, then to write seeds from the tensions, then to write them up. On paper it is the widest batch in the study. The strategist would run 29 of the 40 and pitch 2. More new ideas is not the goal; more ideas you'd pitch is. This is the cleanest example we have of why the last step of this pipeline is still a person.

The model-surgery result from August did not hold. When we re-counted the base-and-instruct blends with the repaired judge, the 70%-base blend was no wider than the untouched model. The lever may exist; our measurement can't show it. We've withdrawn the claim instead of defending it.

Cheap and real, but not enough on its own

If you change nothing else, change how you ask:

The pitch column is one blind read of one batch each, and the differences in it are inside the reader's own day-to-day swing; the new-ideas column is the finding.

Both changes gave us far more ideas we hadn't seen. Neither raised the share worth pitching. Use them as the plain arm of a portfolio, not as the answer.

What this still doesn't tell you

One brief for most of it, one strategist's reads, one sitting. The strategist's pitch calls moved on 4 of 20 cards when re-read a day later, always toward stricter, so small differences in pitch rate between methods mean nothing and we haven't called any. "New to us" means new against everything we generated in August, not new to the world.

What to do with this on Monday

Everything above reduces to one working method. It costs about a dollar in model calls per brief and about twenty minutes of a strategist's time.

1. Fix the brief first (ten minutes). Three sentences: the goal, permission to be tangential, and the through-line rule ("a mention of the client should not feel out of place in the finished piece"). This is what moves the pitch rate; nothing below does.

2. Get your plain batch the better way (one call). Give your main model the brief and ask for all forty at once, as a list of title, two-sentence concept, and the proof point that makes it credible. If the model cuts off before forty, ask in four rounds of ten and paste the earlier rounds back in each time. Don't bother asking it to avoid a list of "ideas already used"; the memory alone does the work.

3. Get forty seeds from a different model (one call, cents). Pick a model from another company — DeepSeek was the best seed-writer we tested, and the point is that it has different habits, not that it is smarter. The prompt is short:

Here is the brief. List 40 one-line campaign premises — just the core concept of each, one sentence, no titles, no explanation. Cover as many different subjects, mechanisms and emotional registers as you can; do not repeat a premise in different words.

4. Have your main model write each seed up (forty small calls). One seed at a time:

Here is the brief, and a raw premise from a brainstorm. Turn it into one campaign idea for this brief. Keep what is surprising about the premise, but anchor it: shape it so a mention of the client would not feel out of place in the finished piece. If the premise can't be made to work for this brief, say so and skip it.

In our runs almost nothing got skipped, and the write-ups kept the seeds' spread.

5. Dedupe with a model, pick with a person. Clustering near-duplicates is the one judging job the models do reliably; ask your model to group the eighty ideas into "same idea" clusters and keep one from each. Then a strategist reads the survivors on three levels: no, usable, I'd pitch it. Forty ideas takes a few minutes. Keep the "I'd pitch it" pile; that is the deliverable.

What to expect. In our one blind read, the strategist pitched about one seeded idea in two, and almost none of them were ideas the plain batch already had. Expect twenty to thirty pitchable ideas per brief from the two batches together. (One reader, one batch each; the reader's own calls moved on one card in five between sittings, so treat the ratio as a guide, not a measurement.)

Measure the right thing. If you track this, count pitchable ideas per brief, not distinct ideas. Every method that looked best on "distinct" and worst in the room (the tensions trick, the heavy decomposition prompts) is a lesson in what happens when you optimise the wrong count.

Don't spend time on these. Temperature. Verbalized sampling. Persona and provocation prompts. Hiding your brand name from the model. Deleting the through-line sentence to "free it up". Asking for the tensions first. Blending open-weight models. Each one was tested; none earned its place.

Try it on your own brief: give us your brief and Fractl Agents writes both batches for you, then shows where each one spent its ideas. It is free, two runs a day.

The lever nobody was testing: the brief

Halfway through, our strategist said the wide-model ideas were "cool, but I'm not sure what they're trying to be about." Our brief said ideas "must be something the company could produce and pitch to journalists" — and every generator read that differently. So we rewrote it the way he actually judges: name the goal, license tangents explicitly, but require a through-line — a CTA or mention of the product must not feel out of place in the finished project.

No new model. No technique. A clearer assignment nearly doubled the usable yield — more than anything else we tested, machinery included. A vague brief collapses the model into its safest cluster; a brief that names the goal and licenses tangents gives it permission to spread and a tether that keeps the spread useful.

No single generator wins. The portfolio does.

The final surprise. Under the rewritten brief, four very different generators each produced roughly 12–21 distinct viable ideas — call it a tie. Then we pooled everything the human had ticked viable and clustered the pool:

Different model lineages mine almost completely different veins of good ideas — same brief, same bar, near-zero overlap. Three times the swath of the best single method. The machinery earns its place not by beating the frontier model but by being uncorrelated with it.

The playbook

• Fix the brief first. Name the goal. License tangents. Require a through-line. The highest-leverage twenty minutes in the pipeline.
• Run a portfolio, not a prompt. Ask your main model for all forty at once (or in rounds that see each other), have one or two unrelated models write seeds, and have your main model write the seeds up. The API-only version costs about a dollar per brief.
• Anchor everything. Render every raw premise for the brief with a frontier model — keep what's surprising, draw the through-line.
• Dedupe the pool, then let a human pick. Clustering is the one job AI judges do reliably. Ranking originality is not — don't automate the last step.
• Skip: temperature, verbalized sampling, personas, provocation instructions, the tensions trick (wide on paper, unpitchable in the room), and any judge that claims to score creativity.

First step, today: take your next ideation brief and add three sentences — the goal, permission to be tangential, and the through-line rule. Then ask two unrelated models, not one model twice.