The Research You Don't Quite Trust: How to Get Findings You Can Actually Check

The afternoon that turns into a trust problem

It's 4 p.m. Someone asked you for a background brief — a read on a new market before Monday, a comparison of three vendors before the team commits, a plain summary of what's actually known about a rule change landing next quarter. Nobody wants a novel. They want two pages, the key numbers, and enough confidence to make a call.

So you open ten tabs. You lift a figure from a report, a stat from a press release, a line from an analyst note. By 5:30 you have a document that reads well. And then, right before you send it up, you stop on your own footnotes and go cold. Where did that 34% come from? Was it this year's number or last year's? If someone in the room asks "says who?", can you answer without scrambling?

That hesitation — not the typing — is the real cost of research work.

Why this eats the afternoon

The writing was never the slow part. The slow part is gathering and reconciling. McKinsey's Global Institute pegged the time knowledge workers spend just finding and pulling together information at roughly 1.8 hours a day — about 9.3 hours a week The Social Economy, McKinsey Global Institute. Close to a full working day, every week, spent before a single sentence of analysis gets written.

AI was supposed to collapse that. Ask a chatbot, get a tidy synthesis in seconds — and for a first draft of your thinking, it does. But it quietly moved the cost instead of removing it: from gathering to verifying. When a model hands you a paragraph with three citations, you inherit a new job — checking whether those citations are real.

They often aren't. In a controlled comparison of chatbots asked to produce references for systematic reviews, hallucination rates ran from 28.6% for GPT-4 to 91.4% for Bard, with GPT-3.5 at 39.6%, measured across 471 references Hallucination Rates and Reference Accuracy of ChatGPT and Bard, PMC. The setting is medical, but the failure mode is universal: a confident-sounding source that doesn't exist, or a real paper with the wrong author, year, or figure bolted on. The output looks finished. That's exactly the trap — it looks finished whether or not it's true.

So you end up worse off than when you started. Instead of not having the answer, you have an answer you can't tell apart from a good guess.

What "checkable" actually means

This is the one place a chat window can't compete, and it's where dwork is built to win. dwork doesn't hand you a paragraph and wish you luck. It takes the whole job — pull the sources, reconcile them, write it up — and hands back a deliverable where every claim keeps a thread back to where it came from.

Concretely, that looks like:

• Every figure traces to its origin. A number in the brief isn't just a number. Click it and you land on the exact sentence in the exact source it was pulled from — the report page, the filing, the article. You aren't asked to trust the 34%; you're shown the line it came from and left to decide.

• Conflicts get surfaced, not smoothed over. When two sources disagree — one says the market is $2.1B, another says $2.8B — dwork doesn't silently average them or quietly pick the prettier number. It flags the disagreement and shows both, with dates, so the judgment call stays with you. That's the opposite of a single confident answer that hides how shaky it is underneath.

• It comes back as something you can hand up. Not a chat log to copy-paste from, but a finished report or a comparison table — the vendor matrix, the market-sizing one-pager — laid out to send onward. When someone asks "says who?", the answer is already sitting in the document.

And because it's a general office assistant, the research isn't the end of the chain. Ask for the brief, the summary table, and a short deck to present it — in one plain-language request — and it runs the whole sequence, handing back files rather than instructions for building them yourself.

The honest part

dwork will not tell you it's never wrong, and you should be wary of any tool that does. Sources go stale, publishers disagree, and no system has read every page on the internet. What changes isn't the existence of error — it's your ability to catch it. When the work shows its sources, a wrong number becomes something you can spot in the thirty seconds it takes to click through, instead of something that ships buried in a paragraph and detonates in a meeting.

A few things it deliberately doesn't do: it's not a live market feed, so it won't quote you a real-time price, and it's not an investment adviser — it assembles background you can read, not calls you should act on. It gives you the checkable groundwork. The decision stays yours.

The afternoon, given back

The goal was never to take the human out of research. It was to take out the part where you re-do the machine's homework. When every figure carries its own receipt, the afternoon stops draining into footnote-chasing and goes back to the work the brief was for — reading the landscape, weighing the trade-off, forming the view only you can.

Trust in a research deliverable shouldn't come from how polished it reads. It should come from being able to check it. That's the deliverable dwork is trying to hand you: not the fastest answer, but the one you can defend line by line.

Built with dwork.ai