Verify before you cite
The full case study, and the verification workflow that prevents it.
Read the guide →By task
Research and fact-checking
The goal is never an answer you trust. It is an answer you can verify, with the unverifiable parts clearly marked.
Four things that get blurred together
Most unusable research output mixes these into a single confident paragraph. Ask for them separated and the whole thing becomes checkable.
The finding
What a source actually says, ideally quoted rather than paraphrased into agreement with your question.
The source
Where it comes from, specifically enough to open. A title with no locator is not a source.
The inference
What the model concluded by combining findings. Legitimate and useful, but not the same category as a finding.
The uncertainty
What could not be determined from the material available. This is the part that gets silently dropped.
Research [question].
For every substantive claim, label it as one of:
[SOURCE] the source says this — include a direct quotation
[INFERENCE] my reasoning across sources — show the step
[UNKNOWN] not determinable from what I could find
Include findings that contradict the apparent answer. If the evidence is thin or mixed, say so rather than picking the tidier story.
End with the three things I would need to check before relying on this.The citation problem is not theoretical
Fabricated references are the characteristic failure of this task, and they are convincing because a citation is a highly patterned object. Reproducing the shape requires no knowledge that the work exists.
This is not a rare edge case. A public database tracking court filings that relied on AI-fabricated citations grew from roughly two hundred entries in mid-2025 to over fifteen hundred a year later. In one 2026 appellate case a fake authority passed through five sets of professional hands, including a judge's signature, without anyone opening it.
The rule that prevents all of it. If you did not open the source, you did not read it, and you cannot cite it. A model's citation is a lead to check, never a reference to reuse.
Search tools change the failure mode, not the discipline
A model with live search grounds its answer in pages it actually retrieved, which removes much of the fabrication risk. It introduces a different one: the retrieved page may be wrong, outdated, promotional or itself AI-generated. Grounding tells you where a claim came from, not whether the source deserved trust.
So the check shifts from “does this exist” to “is this any good”. Who published it, when, with what interest, and does anyone independent agree.
Where community sources fit
Forums, Reddit and Q&A sites are genuinely valuable for discovering what problems exist, what vocabulary real people use, and which failures are widespread rather than anecdotal. They are not evidence that something is true. Use them to find the question, then answer it from a primary source. That is the same rule we apply to ourselves.
Why confidence is not evidence
The mechanism behind fluent, well-formed wrongness.
Read the guide →Method and sources
How this site decides what counts as a source.
Read the guide →