Dr. Robert Epstein on politics: what the evidence says · JRE #1768

What the evidence says 01 / RECORD

Epstein and colleague Ronald Robertson published peer-reviewed experimental work in PNAS (2015) showing that biased search rankings can shift undecided voters' preferences in controlled lab settings. The specific 2.6-to-10.4-million-vote figure for the 2016 election, however, comes from a separate, non-peer-reviewed white paper and later congressional testimony, not from a published, peer-reviewed study; it extrapolates from a small 2016 field observation of 95 subjects (21 undecided) over 25 days, using an unpublished computational model. When the same 2.6-to-10.4-million figure was cited (via a distorted "up to 16 million" framing) by then-President Trump in 2019, PolitiFact's review found outside experts identifying major methodological problems: the study lacked a clear definition of "biased" search results, relied on crowdsourced subjective bias ratings, excluded Gmail users based on a speculative theory that Google was intentionally showing them unbiased results, and never publicly disclosed the formula used to extrapolate from a few dozen subjects to a multi-million-vote national estimate. Independent researchers said there was insufficient information in Epstein's materials to verify that his model validly translated observed search-result bias into vote totals, and Google called the claim inaccurate and previously debunked. As of this writing, Epstein's vote-shift estimates for 2016 have not been independently replicated or published in a peer-reviewed venue, and the methodology remains contested.