TL;DR
When you ask an LLM to come up with something new, there's a huge chance it'll suggest linking two existing things together. This isn't a coincidence: research shows that 47–64% of ideas from any top-tier model boil down to the format "there's a gap between X and Y, let's bridge it." For humans, this idea type accounts for just 12%. The model isn't being lazy — it literally doesn't see other moves.
The problem is that there are at least seven distinct ways to justify a new idea: explain a puzzling phenomenon, close an evidence gap, reduce failure risk, or eliminate a resource bottleneck. The LLM uses one trick — the "bridge" — and calls the job done. You get an idea that looks clever but is really just a mashup of two known things.
The research gives you a taxonomy of idea types: 7 flavors of "why this idea is needed" and 7 flavors of "how to build it." If you explicitly tell the model to use a specific type, it snaps out of default mode and generates fundamentally different ideas.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: Give context (what already exists in the niche) → list 3-5 sources/competitors/approaches
STEP 2: Ban the default mode → explicitly exclude "bridge/synthesis" from allowed idea types
STEP 3: Specify the required type from the taxonomy → the model generates an idea using that template
[All in one prompt]
Example in Action
Task: You're launching an online financial literacy school for a 25–35 audience. You want to come up with a new course format — not yet another "budgeting course," but something unexpected.
Prompt:
Here's what already exists in the personal finance online education market:
— "From zero" courses: teach basics (savings accounts, budgeting, credit)
— Kiyosaki-style bootcamps: motivational content about investing
— Excel templates and planners on Gumroad
— Podcasts with case-study breakdowns
Come up with a new course or product format.
Important constraint: DO NOT propose an idea in the format "combine X and Y"
or "build a platform that unites A with B."
I don't want that type of idea.
Instead, pick ONE of the following types and generate an idea based specifically on it:
TYPE 1 — UNEXPLAINED PHENOMENON: There's a user behavior
that nobody can explain. What is it? Why does it happen?
Build the idea around explaining this phenomenon.
TYPE 2 — EVIDENCE GAP: Everyone claims X, but nobody has actually
tested it. The idea is to test it and report back.
TYPE 3 — FAILURE RISK: People do something standard,
not realizing it leads to a common mistake.
The idea protects against this failure.
TYPE 4 — RESOURCE BOTTLENECK: There's a specific resource
(time, money, attention) that's become a chokepoint.
The idea removes exactly that bottleneck.
Pick one type, name it, explain the reasoning, and give a concrete product idea.
Result: The model will name the chosen type, explain the specific phenomenon/risk/gap it found in the niche, and propose a product built around it — not another "podcast meets course" mashup. On the next request, you can ask for a different type and get a fundamentally different idea.
Why This Works
LLMs are trained on text patterns, not on the logic of reasoning. When you give one a few sources and ask it to "come up with something new," it looks at the context and searches for connections between objects. This is the single most common pattern in text: "study A + study B → conclusion C." The model reproduces what it's seen most often.
"Bridge" ideas aren't bad ideas — they're just one of seven possible moves. The problem is that without explicit instructions, the model only uses that one. It's like a chess player who knows exactly one opening and plays it in every position.
An explicit ban plus an explicit type changes the generation. You remove the default template ("don't build a bridge") and hand over a different one ("use the FAILURE RISK type"). Now the model has no choice — it generates according to the pattern you specified. This isn't a vague "be more original" request — it's a concrete rewiring of the generation mechanism.
Control levers: - Number of types in the prompt → give it all 7 and ask for one idea per type — you'll get a set of genuinely distinct ideas - How the ban is worded → the more specific ("don't propose 'combine X with Y'"), the stronger the effect - Sequential requests → specify a new type each time and build a diverse idea bank - Add "why this type" → ask it to explain the reasoning behind the choice — the model develops the idea more deeply
Prompt Template
Here are the existing approaches / competitors / solutions in the {field} niche:
— {source 1}
— {source 2}
— {source 3}
Come up with a new idea for {task}.
BAN: do not propose an idea in the format "combine/unite/integrate X and Y."
Use THIS thinking type: {type from the list below}
THINKING TYPES:
— UNEXPLAINED PHENOMENON: find a behavior that nobody explains
— EVIDENCE GAP: find a claim that nobody has actually tested
— SCALE MISMATCH: a method works in one context but fails in another — why?
— FAILURE RISK: find a common mistake that people don't notice
— BOTTLENECK: find a specific resource that's holding everything else back
— MEASUREMENT: find something everyone discusses but nobody has actually measured
— ARTIFACT: find a tool or object that doesn't exist but should
Name the chosen type, explain the reasoning, and give a concrete idea.
What to fill in:
- {field} — your niche or market: "EdTech for adults," "B2B SaaS," "wellness products"
- {source 1-3} — competitors, existing approaches, known solutions in the niche
- {task} — exactly what you're brainstorming: a product, content format, feature, or strategy
- {type} — pick one type, or leave the choice up to the model (it'll then name what it picked)
🚀 Quick start — paste into chat:
Here's a template for generating unconventional ideas.
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask about your niche, existing solutions, and idea type — because without that, it can't apply the ban on the default "bridge" template. It will take the structure and plug in your data.
Limitations
⚠️ Synthesis is sometimes the right answer: The taxonomy isn't saying "bridge" ideas are bad. If your task genuinely requires combining two approaches, the ban will hurt you. Use this constraint deliberately, when you specifically want to break out of the obvious.
⚠️ Extended reasoning mode makes it worse, not better: Turning on o1/o3/thinking mode in Claude and DeepSeek amplifies the lean toward synthesis rather than fixing it. Bridge ideas jump from 49% to 71% with thinking mode enabled. If you want diversity, standard mode works better.
⚠️ More context doesn't help: Feeding the model full texts instead of short descriptions doesn't change the picture. The bias persists regardless of input volume.
⚠️ The taxonomy is built for research ideas: The seven types were derived from academic and grant-funding requirements (NSF, NIH, DARPA). For business or creative tasks, you'll need to adapt the wording to your context.
How It Was Studied
A team from Yale and the University of Chicago took 11,683 real research papers from top conferences (ICLR, ICML, NeurIPS) and the journal Nature Communications. For each paper, they worked backward to reconstruct the 4–8 sources that apparently inspired the authors' idea. Then they fed those same sources to nine top-tier LLMs and asked each to propose a new idea. The human idea is what the researcher actually wrote in the paper. The model idea is what the LLM came up with from the same source material.
Next, they tagged each idea along two taxonomy axes and compared the distributions. The result was unambiguous: human ideas were spread evenly across all types, while model ideas were crammed into one corner. The most telling part: the researchers checked whether giving the model full paper texts instead of abstracts would change anything. It didn't. Then they tested thinking mode. It got worse. This suggests the bias is baked in deep — it's not a matter of insufficient information, but a pattern the model reproduces consistently.
Adaptations and Extrapolations
🔧 Technique: Sequentially cycling through all types → a bank of diverse ideas
Run 7 separate requests with the same context, changing the type each time. You'll get 7 fundamentally different ideas instead of 7 variations on the same "bridge."
[Same context about the niche]
Type for this request: FAILURE RISK
Find a common mistake that users in {field} make
without realizing the consequences. Build a product idea around protecting against this mistake.
🔧 Technique: Meta-audit of already-generated ideas
If an LLM already gave you a list of ideas, ask it to classify them using the taxonomy. You'll immediately see whether everything landed in the "bridge" zone.
Here's a list of ideas you suggested to me earlier:
[list of ideas]
Classify each one by type:
"bridge/synthesis" — connects existing things
"explanation" — explains an unknown phenomenon
"risk" — protects against failure
"measurement" — measures something for the first time
"artifact" — creates a new tool
"scale" — transfers a method to a new context
"bottleneck" — removes a specific constraint
Which types are missing from my list?
🔧 Extrapolation: the same principle for content, not just ideas
The "bridge" bias also shows up in content ideas. Ask for YouTube channel topics and you'll get "X vs. Y comparison," "X meets Y." The same ban plus the same taxonomy (adapted) yields fundamentally different formats.
Resources
Measuring the Gap Between Human and LLM Research Ideas Ziyu Chen (University of Chicago), Yilun Zhao, Arman Cohan (Yale University)
Dataset: IdeaLand/IdeaSeed on HuggingFace Code: ziyuuc/TasteGap on GitHub