A Research Workflow That Holds Up

AI research assistance is genuinely useful and it fails in a specific way: it produces a coherent, well-organised, confidently sourced summary in which some of the sources do not support the claims attached to them.

The workflow below is built entirely around catching that.

The order

One: use a tool that retrieves. Anything answering from memory alone is unusable for research. If the tool did not search, you have a plausible essay rather than a finding.

Two: ask for sources with every claim, not a bibliography at the end. Claim-level attribution is what makes checking possible.

Three: read the sources. Not all of them, and every one you will rely on. This is the step that gets skipped and it is the entire point.

What you are checking: does the source exist, does it say what is claimed, is it the kind of source that supports that claim, and is it current.

Four: check what is missing. Retrieval finds what matches the query. Ask explicitly what the counter-arguments are, who disagrees, and what the weaknesses in the evidence are. The default output is one-sided because the search was.

Five: go to primary sources for anything load-bearing. A summary of a summary of a study is three steps from the data, and errors compound at each step.

Six: check dates. Both the source's date and whether the position has changed since.

The specific failures to watch for

Fabricated citations. Plausible author, plausible journal, no such paper. Common enough that every citation needs checking.

Real source, wrong claim. More insidious. The paper exists; it does not say what the summary says, or says it with qualifications that were dropped.

Misattributed findings. A result from one study attributed to another.

Dropped qualifications. "Some evidence suggests, in a small sample, under specific conditions" becomes "studies show".

Confident synthesis of contested material. Where the literature disagrees, a summary that presents one position as settled.

Stale positions presented as current.

Low-quality sources treated as authoritative. Retrieval does not weigh credibility well. A content farm and a peer-reviewed journal look similar to it.

What it is genuinely good for

Orientation in an unfamiliar field. Getting the vocabulary, the main positions, the names to look for. Enormously faster than starting cold.

Finding search terms you would not have known to use.

Summarising material you supply. Give it the papers and ask questions about them. This is reading comprehension rather than recall and it is far more reliable.

Explaining a difficult passage in something you are reading.

Finding contradictions between documents you provide.

Drafting a literature summary from sources you have already read and verified.

Notice the pattern: it is most reliable when working on material you gave it, and least reliable when finding material for you.

The habit that makes the difference

Separate finding from verifying, and do them in different sessions.

Use the tool to produce a list of leads. Then go and read them yourself, in the actual sources, before writing anything. Do not write from the summary and check afterwards — by then the summary's framing has shaped what you wrote and you will read the sources looking for confirmation.

Where the stakes justify more care

Anything published under your name. Every citation checked in the original.

Anything legal, medical or financial. Primary sources only, current versions verified.

Academic work. Institutional rules apply and fabricated citations are a serious matter.

Anything a decision rests on. The cost of being wrong sets the verification budget.

The honest summary

These tools compress the orientation phase of research from days to hours, which is a large gain. They do not compress verification, and treating the output as verified is how people end up citing papers that do not exist in documents with their name on them.

Use it to find things. Verify them yourself. That division is the whole discipline.

Checking a citation quickly

The single highest-value verification habit, and it takes under a minute per reference once you have the routine.

Search the exact title. If nothing comes back, it does not exist. This catches most fabrications immediately.

Check the author actually wrote it. Fabrications frequently attach a real author to a title they did not write.

Check the journal and year exist together. A journal that began publishing after the claimed date is a clear tell.

Open it and search for the claim. This is the step that separates real verification from confirming a reference exists. A real paper that does not support the claim attached to it is the more common failure and the harder to catch.

Check the claim survived its qualifications. Sample size, population, conditions. Summaries routinely drop these.

For anything load-bearing, read the abstract and the conclusion properly.

Five checks. On a piece with twenty references it is an hour, and it is the hour that prevents the failure that ends careers.