Why AI Research Feels Mediocre
The Observation
Even the best AI models often give mediocre research answers. Not because they lack intelligence, but because of how they use it.
The Issue
When an AI researches a topic, it pulls from many sources and blends them into one summary. This sounds reasonable, but it causes three problems:
Sources aren't actually independent. Much of the internet is written to be found (SEO), not to be right. Five different pages often just repeat the same shallow idea. So the AI isn't really averaging five opinions — it's averaging five copies of one opinion.
Averaging destroys disagreement. If one source says "go left" and another says "go right," the average is "go straight" — a position nobody actually argued for. Blending conflicting views can produce an answer worse than either original view.
Too many hard problems at once. To do this well, the AI would need to judge which sources are trustworthy and reason through each argument and produce a balanced answer — all in one pass. When too many hard decisions are bundled together, any reasoner (human or AI) takes a shortcut and settles for "good enough" instead of "best." This is what Herbert Simon called bounded rationality.
Underneath this is a deeper issue: judging a source's quality is not purely a fact-check. It depends on values — what counts as reliable, what counts as important. Those values can't be verified by the sources themselves. This is what Simon called the value premise problem: at some point, you're standing on a choice, not a proof.
The Solution
Separate the tangled problem into distinct, manageable steps:
Trace, don't blend. Take each source's line of reasoning and follow it to its own conclusion, separately. Don't mix them together.
Compare outcomes, not opinions. This produces several distinct possible answers, not one mushy average.
Choose with a stated reason. Pick the best outcome using a clear, visible criterion — and say what that criterion is. This won't remove the underlying value judgment, but it makes it visible and arguable instead of hidden inside a paragraph.
This is simply separation of concerns — a well-known idea from good design. Instead of asking one process to weigh sources, reason through them, and decide all at once, split it into stages. Each stage becomes easier to do well, and easier to check when something goes wrong.
The Takeaway
AI research isn't mediocre because the model is weak. It's mediocre because a hard, coupled problem is being solved all at once, with no separation between judging sources, reasoning through them, and choosing an answer. Separating these steps won't make every judgment call disappear — some are genuinely just values, not facts — but it will stop the silent, low-effort averaging that currently masquerades as balanced analysis.