Ask AI for another way to approach a problem, and it will probably just give you a similar suggestion to the last one. At Fractl, we tested whether giving AI perspectives borrowed from other fields would help. We used a biology problem to compare the explanations it produced, then translated the approach into prompts you can try when developing campaign ideas.

For your next brainstorm, instead of asking AI to “be more original,” give it a specific question to explore. Here’s what worked in our test and how you can try it when planning a campaign.

79 Suggestions Amounted to Just Six Ideas

We asked AI to invent a new way of approaching the same problem on each attempt. Its responses used different names, but many described a similar process: something builds up, reaches a threshold, and triggers a response.

Changing the label didn’t change the reasoning. When we grouped the approaches by what they described, the list was much smaller.

We had asked AI to suggest ways of thinking before giving it the actual problem. Even at that first step, it kept repeating itself. For a campaign brainstorm, the same warning applies: 

Ten different headlines aren’t ten ideas if they all call for the same study. 

Give AI a Specific Question To Explore

We tested a question about bacteria that survived an antibiotic treatment. The question specified that the survivors had no resistance genes and didn’t pass resistance to their descendants. We wanted AI to suggest explanations that went beyond repeating the familiar account of dormant cells surviving treatment.

Rather than asking AI to invent another method, we supplied questions borrowed from engineering, research checklists, and creative exercises. 

Could a group average hide differences between its members? Could something we assumed was a cause actually be an effect? These gave AI specific things to look for.

We first had AI explain one of those principles without knowing the biology question. Only then did we give it the problem and ask it to apply the principle.

One answer proposed that cells might move in and out of dormancy faster than the experiment measured them. It also suggested a way to investigate that explanation: observe individual cells throughout treatment, rather than rely on occasional measurements.

These were possible explanations to test, not discoveries about how bacteria actually behave. 

A Longer Question List Helped AI Find More Explanations

Our first comparison reused three outside perspectives. They produced about as many different explanations as the approaches AI invented for itself. We then expanded the collection to 20 perspectives while keeping the problem and scoring approach the same.

The larger collection produced 38 distinct explanations, compared with 20 from the smaller collection.

The AI-generated approaches produced 19 distinct explanations. Each method generated a similar number of answers, between 76 and 79, before we grouped repeats together.

An explanation counted as different only if it proposed a different cause and implied a different experiment. A separate AI model grouped the answers using that definition.

If we’d stopped after trying three questions, we would have missed the improvement from using more. 

When a brainstorm stalls, try a different question before giving up on the topic.

Keep a list of questions you can reuse: who is missing from the data, what varies by location, or why people switch products. We haven’t established an ideal list length or shown that this produces more media coverage.

Ask the General Question First, Then Add Your Topic

We also changed the order of the prompts. In one version, AI explained a general idea, such as how group averages can hide individual differences, before seeing the biology question. In the other, it already knew the biology question when it wrote that explanation.

Both versions produced 13 distinct ideas. But the share of answers our AI grader judged solid enough to test was 64% when AI developed the approach first, compared with 34% when it already knew the topic.

Try two prompts instead of putting everything into one request.

First, ask AI: “How can an average hide differences between groups of people? Give me an example.” Then tell the same AI tool: “We’re planning a study of household costs. Which groups should we compare to find differences the overall average might miss?”

You might get a question such as whether renters face bigger cost increases than homeowners. That gives your team a comparison to investigate, rather than the broad topic “rising costs.” This marketing example illustrates how to use the prompts; it wasn’t part of our biology test.

Ask AI To Reconsider Its Weakest Point

One follow-up asked the model: “What’s the weakest step in that reasoning, and what’s the version that survives it?”

That follow-up increased the number of distinct ideas from seven in the comparison group to 20 with the follow-up. The share of answers judged worth testing stayed at 87%.

Another follow-up asked AI to justify how it had followed our instructions. It did that well, but only 29% of those answers were judged worth testing.

Asking AI to challenge its answer more often led to a changed explanation. Asking it to justify following our instructions often got us the same idea in different words.

For your brainstorm, ask: “What could be wrong with this idea? What would you change to address that?” 

A useful revision gives you a better question to research or a way to check a doubtful claim. A more confident explanation of the same idea isn’t enough.

Try Different Questions in Your Campaign Brainstorm

When AI keeps suggesting the same campaign, give it one of these questions to explore. Its answers can suggest what to research next; they aren’t findings you can publish without checking:

Do these customers want the same thing? Ask AI why people in the same age group might buy a product for different reasons.	Does the answer change by place or season? Ask AI what a national average might hide about people in different cities.
How are people solving this already? Ask AI what people might be doing when existing products don’t meet their needs. Then research whether those habits are common.	Could a drawback be a selling point? Ask AI whether a feature that annoys one group might attract another.
Whose experience is missing? Ask AI what you would miss by surveying only current customers, rather than people who stopped buying.	Why do people switch? Ask AI which questions to put to customers who recently changed products or services.
What’s making this harder for people? Ask AI how a tight budget, limited time, or lack of access might change people’s choices.	Does the usual advice hold up? Ask AI which common recommendations you could check against real data.

Try one question at a time with the same campaign brief. Compare the answers and keep the suggestions that lead to a study your team could actually carry out.

Check the Ideas and the AI Reviewing Them

We used another AI model to group repeated ideas. To check its work, we gave it identical copies of one answer. It correctly counted them as one idea. It also found fewer different ideas in answers produced without the supplied questions.

Those checks were useful, but they weren’t independent human verification. Each idea count in this article comes from one AI review, so treat the counts as estimates from this test, not unquestionable totals.

Other parts of our testing failed. One follow-up asked AI to reconsider an answer without including the answer it was supposed to assess. The model generated a response anyway. We removed 93 fabricated answers caused by that error and reran the scoring.

A believable score can still come from a broken test.

When you assess an AI brainstorm, check more than the final rating:

• Compare the claims. Group suggestions that would investigate the same question, even when their headlines differ.
• Check what AI received. Confirm that each revision or review includes the idea and evidence it’s supposed to assess.
• Test an obvious repeat. Give your reviewer identical suggestions and check that it recognizes the duplication.
• Review the shortlist yourself. Decide which questions you can research and which findings could interest a journalist.

We tested AI’s answers to a scientific question, not marketing campaigns, and used AI graders to assess them. The results suggest a way to get more varied suggestions, not proof of better campaigns or more media coverage.

Give Your Next Brainstorm a Different Starting Point

Start with a campaign brainstorm that has stalled. Pick one question from the list above, ask AI to explain it, then give it your campaign brief. Ask what could be wrong with its suggestion before choosing what to research. The goal isn’t a longer list of titles. It’s a question your team hadn’t considered and can now investigate.

We use this research to improve how Fractl Agents develops and reviews campaign ideas. Your team’s judgment still determines which suggestions deserve further work.

If you want help turning those questions into original research and stories worth pitching, reach out and partner with Fractl. We’ll develop the campaign, give it to relevant journalists, and help boost your brand authority.

Methodology

We ran this study in September 2026 and collected about 3,100 AI answers to one biology question: why some bacteria survived an antibiotic treatment without resistance genes and without passing resistance to their descendants. We compared answers built from approaches the AI invented itself with answers built from three, and then 20, perspectives borrowed from other fields. Each method produced 76 to 79 answers, and we also ran every method with its perspective removed as a control.

A separate AI model grouped the answers, counting an explanation as distinct only if it proposed a different cause and implied a different experiment. When we gave it 79 copies of one answer, it counted one idea. An AI grader judged which answers were solid enough to test.

We found seven errors in our testing setup. One sent follow-up questions without the answer they referred to, which produced 93 fabricated answers out of about 1,000. We removed them and reran the scoring, and the figures in this article exclude them.

We tested one scientific question, not marketing campaigns. A second test question failed because our list of its known explanations was incomplete. Each idea count comes from a single AI grouping pass.