Source laundering: how weak sources come out of AI looking strong

Ask an AI about a health claim, a market size, or a legal rule, and the answer arrives in one seamless voice. Behind sentence two sits a methodical government dataset. Behind sentence four sits a page written to rank, by no one in particular, sourced from nothing checkable. In the answer, both sentences sound exactly the same.

That flattening is the laundering. On the open web, a junk page at least looks like a junk page: the ads, the padding, the anonymous byline all warn you. Inside an AI answer, those signals are stripped and the content is rewritten in the assistant's calm, authoritative register. The weakness survives; only the warning labels are gone.

Retrieval systems rank sources by relevance and extractability, not by epistemic quality. A page that answers the query directly, in clean structure, with quotable sentences, gets pulled, and content farms have optimized for exactly that shape. Columbia's Tow Center measured the downstream result: AI search engines misattributed or fabricated sourcing on more than 60% of test queries, with even the best engine wrong 37% of the time.

Judging whether a source deserves belief requires knowing who stands behind it, what evidence it rests on, and what it was written to achieve. Those are questions about the world, not about the text, and the pipeline never asks them. Sentence-level citations tell you where a claim came from; they cannot tell you whether where it came from is any good.

The check that undoes laundering takes about thirty seconds per source, and it asks the questions the machine skipped.

Who wrote it? A named expert, an institution with a reputation to lose, or nobody findable. Anonymity is not disqualifying, but it removes the accountability that lets you borrow confidence.

On what evidence? Primary data, named studies, and linked documents, or just assertions with the grammar of evidence. Follow one citation deep: pages that cite pages that cite nothing are a genre.

With what incentive? Written to inform, to rank, or to sell. Incentive does not automatically falsify, but it tells you where the author stopped checking.

Thirty seconds per source is cheap, but not free, so spend it where laundering costs the most: claims you will repeat, numbers that drive decisions, and anything that surprised you. Surprise is the highest-value trigger, because a surprising claim from a weak source is the single most viral form of laundered content.

Skip the check for the free tier of your work, the same boundary the team verification flow draws. And notice the compounding trick: sources you have checked once, your own library of vetted documents and trusted references, do not need re-checking every time. That is the quiet argument for doing serious work over a curated corpus instead of the open web: the source-quality check gets done once, upstream, instead of per-answer, forever.

This is exactly why Tatsulok is built around libraries and curated collections rather than open-web retrieval. Your own documents are sources you already stand behind. Curated collections, like the Philippine law corpus of statutes, codes, and Supreme Court decisions, are vetted at the corpus level, so every answer inherits provenance instead of laundering it. Each claim still links to its exact passage, and gaps are stated honestly, so both layers of the check, the passage and the source, are one click deep.

This week's drill, no signup needed: take one AI answer with sources and run the three questions on each source it cites. Score them: how many were written by someone findable, on evidence you can follow, to inform rather than to rank? That score, more than the answer's fluency, is what the answer was worth.