How to Use AI for Research Without Fabricated Sources

The failure mode that ends careers is not the model being wrong. It is the model being wrong in a format that looks exactly like being right — a citation with plausible authors, a real journal, a volume number, and a paper that does not exist.

The defence is not a better prompt. It is understanding one structural fact: a language model cannot verify anything. It can tell you what a claim would look like if it were true, and it can tell you what to go and check. Those are different jobs, and conflating them is where research work goes wrong.

Fabrication is not a bug you can prompt away

A model generates text that is probable given what came before. A citation is highly patterned — author, year, title, venue — so producing a well-formed one that happens to be fictional is not a malfunction. It is the system doing exactly what it does, applied to a case where form and truth come apart.

Only cite real sources does not fix this, because the model has no mechanism for checking whether a source is real. It will comply with your instruction the only way it can: by producing citations that look more convincingly real.

This holds even when the model is retrieving. Purpose-built research tools that fetch documents still misattribute what they fetched, because summarising a source and citing it accurately are separate operations and only one of them is grounded.

The distortion we hit in our own work

Two examples from writing this site, both caught only because we re-checked.

The first: the widely repeated claim that it takes 23 minutes and 15 seconds to recover from an interruption. This is always cited to Mark, González and Harris, No Task Left Behind? (CHI 2005). We ran the actual PDF through pdftotext — the string 23 min appears nowhere in it. The paper reports that 57% of working spheres are interrupted, and that resumed work typically has more than two intervening activities first. The 23-minute figure comes from a 2006 Gallup Business Journal interview with Gloria Mark, not the paper. Every article repeating it inherited the misattribution from another article.

The second: we asked a fetch-and-summarise tool for the author list of an arXiv paper. It returned Yupei Jia, Zhaohan Shao, Yuqi Liu. The real list is Yuqi Jia, Zedian Shao, Yupei Liu — the given names had been shuffled across the surnames. Every name was real. Every name was attached to the wrong person. Nothing about the output looked wrong.

Neither error would have survived a source check. Neither would have been caught by a better prompt.

Make the model identify claims, not verify them

The single most useful reframing: stop asking is this true? and start asking what would I have to check for this to be true?

A model is genuinely good at the second. It can read a draft and produce a list of every load-bearing factual claim — every number, date, attribution, and causal assertion — ranked by how much damage a wrong one would do. That list is a research plan. You then check it against sources yourself.

Read this draft and list every factual claim that would need verification before publication. For each: the claim, why it matters to the argument, and what kind of source would settle it. Do not tell me whether the claims are true.

That last sentence does real work. Without it, the model will grade the claims, and its grades are exactly as reliable as the claims themselves.

Work from supplied material, not from memory

The reliability difference between summarise this document I'm giving you and tell me about this topic is enormous, and it is the main lever you control.

When the source text is in the context window, the model is doing comprehension — a task it is good at, and one where errors are checkable because you have the source in front of you. When it isn't, the model is doing recall from training, which is where fabrication lives.

So: paste the paper. Paste the transcript. Paste the ten pages of notes. The context window is large enough for most real research inputs, and the reliability gain is worth more than the convenience of not pasting.

One caveat on long inputs. Lost in the Middle (Liu et al., TACL 2023) found that models use long contexts unevenly — accuracy is highest when the relevant information sits near the beginning or the end, and degrades noticeably in the middle. A big context window is not the same as even attention across it. Put what matters most at the edges, and for long documents ask about sections rather than the whole thing at once.

Synthesis is the real strength

Where models genuinely excel is the work between reading and writing: taking fifteen sources you have already gathered and finding the shape in them. What do these agree on? Where do they conflict, and is the conflict substantive or terminological? Which claim is load-bearing for the conclusion and supported by only one source?

That last question is the valuable one, and it's the one a tired human reading fifteen papers is worst at. It's also safe, because it operates entirely on material you supplied and can check.

The research and analysis prompts on this site are built on that constraint throughout: work only from provided material, distinguish evidence from inference, never invent a statistic or a source, and — because the model cannot browse — identify what needs verifying rather than claiming to have verified it.

A working method

1. Gather sources yourself. Search engines and databases find documents; models do not.
2. Paste the material in. Comprehension beats recall by a wide margin.
3. Ask for structure, not verdicts. Themes, conflicts, gaps, and which claims carry the argument.
4. Get a verification list. Every factual claim, ranked by consequence.
5. Check the list against sources. This step is not optional and cannot be delegated.
6. Re-check anything that travelled. If a fact reached you through a summary, a secondary citation or another article, go to the primary. Both of our errors above were inherited, not generated.

That last habit is the one that separates research you can stand behind from research that merely reads well. A summary is a lossy copy, and the losses do not announce themselves.

What this costs

More than the marketing suggests, and less than doing it unaided. You still read the sources. You still check the claims. What you save is the synthesis pass, the first-draft structure, and the tedium of listing what needs checking — real savings, on the order of hours.

What you must not save is the verification. The moment you skip it, the tool's fluency stops being an advantage and becomes the mechanism by which a fabricated citation reaches print looking exactly like a real one.