Pinned post
If you know me, you know my line: LLMs lie. A lot.
In 2025, Columbia's Tow Center proved AI search engines botch citations.
I re-ran their test in 2026: 800 answers, and I clicked every link. About 70% wrong, and half the links broken or fake.
My addition: a checkpoint that clicks every link before an answer reaches you, checks the quote is on the page, the site, the date. Fabricated links reaching the reader fell to zero. Wrong answers fell from about 7 in 10 to about 1 in 8.
The engines lie just as much; the lies just can't get through.
The receipts, case by case: marcus-ai.dev/on-the-record/…
The code: github.com/painted-porch/…
In 2025, Columbia's Tow Center proved AI search engines botch citations.
I re-ran their test in 2026: 800 answers, and I clicked every link. About 70% wrong, and half the links broken or fake.
My addition: a checkpoint that clicks every link before an answer reaches you, checks the quote is on the page, the site, the date. Fabricated links reaching the reader fell to zero. Wrong answers fell from about 7 in 10 to about 1 in 8.
The engines lie just as much; the lies just can't get through.
The receipts, case by case: marcus-ai.dev/on-the-record/…
The code: github.com/painted-porch/…