Skip to content

The Nigerian Princes of Polling

commentary

The Nigerian Princes of Polling

by John Nienstedt

You would think every reporter has been trained, at some point, to spot the email from the “Nigerian Prince.” The grammar is a little off, the promise is too good, and the ask is always for something small up front in exchange for something enormous later. Most journalists delete it without a second thought. Political Twitter, it turns out, has no such kill switch.

That’s the real lesson of the Median Strategies affair, and of its October sequel, Wedgewood Polls. It’s not “polling is broken,” but that the ecosystem surrounding polling — the aggregators, influencers, campaigns, and cable-news chyrons that immediately turn a number into a narrative — has almost no antibodies left. And, encouragingly, that the parts of the ecosystem with purpose-built antibodies worked as designed. Both times, the fakes were exposed by people who vet polls for a living.

The setup

Stripped of spin, here is what happened. An anonymous outfit calling itself Median Strategies registered a domain, wrote a three-page PDF dressed up with the trappings of legitimacy — sample size, margin of sampling error, methodology language — and posted it to a website with zero organic search presence. The subject was the Los Angeles mayor’s race: incumbent Karen Bass supposedly leading challenger Nithya Raman by double digits, with third-place finisher Spencer Pratt’s voters breaking overwhelmingly for Bass. The perpetrator then opened a brand-new X account with 21 followers and personally messaged the release to high-volume poll-tracking accounts. That was the entire distribution strategy.

There was no data collection. No respondents. By the American Association for Public Opinion Research’s (AAPOR’s) account, “the responses were simply made up.” Median Strategies, it later emerged, had also fabricated polls in a Wisconsin gubernatorial primary and a Nevada governor’s race, so the Los Angeles caper wasn’t a one-off.

A little digging into Median’s own footprint tells you how thin the operation was. The domain’s registration lists a North Carolina address, but that’s not evidence anyone involved is based there. It’s the address of Porkbun’s WHOIS privacy service, a placeholder any domain buyer can opt into for a few dollars a year. The real registrant information sits privately with Porkbun, unpublished. In other words: even the “paper trail” was a rental.

It turned out to be an apt metaphor for the whole operation, because there was no operation. The Guardian has since identified the person behind Median Strategies as Rahil Prakash, a 21-year-old recent college graduate who built the site himself, using AI, and ran it alone. There was no funding, no campaign, nothing more sophisticated than a laptop and a hunch. He told the Guardian he’d been struck by how badly public polling missed the Michigan Senate primary, where Abdul El-Sayed’s lead over Haley Stevens evaporated on election night, and wanted to “test my theory” about how easily fabricated numbers could travel. “I think my overall takeaway was that if you just make it look a bit pretty, it’s rather easy for it to spread,” he said. He denies trading on any prediction market and says no one involved was compensated. If that holds up, the most unsettling part of this story isn’t a shadowy operation, it’s the absence of one. One person with a free AI website builder did what would have once required actual disinformation apparatus.

Who bit, and who didn’t

The accounts first amplifying the numbers, @PollTracker2024 and @OpenSourceZone, pushed them to a combined audience of more than 50,000 followers. A spokesman for @OpenSourceZone later told the Washington Post the account “relies on publicly available polling information and does not independently verify the underlying data of every survey,” and that Median “presented itself as a polling organization” convincingly enough to pass. That single, thin layer of trust, which amounted to a properly labeled margin of sampling error, was the entire con. From there, virality mechanics took over: the 85.2%-of-Pratt-voters number was tailor-made for the attention economy, the California Post wrote it up as a rare bright spot for Bass, Pratt mocked the story to his massive following, the Bass campaign turned the datapoint into a momentum graphic, and Fox 11 Los Angeles cited the campaign’s own announcement in a broadcast update.

But notice who didn’t bite. The Los Angeles Times’ Noah Goldberg said he found the poll “suspect immediately because the firm was so new, had so few followers, and did not have an actual human pollster listed on its website.”  That constitutes a baseline the Times requires of any polling firm it covers. The New York Times, RealClearPolitics, and FiftyPlusOne (three major poll-tracking sites) never added Median to their averages at all. Both the Times and FiftyPlusOne said they reached out to Median for basic methodological information and simply never heard back. It’s a reasonable guess that those routine questions are precisely what caused Prakash to abandon the ruse rather than answer them. As longtime pollster Mark Blumenthal, a co-founder of pollster.com, put it: “Bravo. The fact that common aggregators have put transparency standards into place helped keep this bogus data out of their aggregates.” His sharper concern was aimed at the journalists and influencers willing to share clickbait numbers.

This is the part AAPOR’s statement should say louder: the vetting desks, the aggregator standards, a reporter who knows to ask for a human pollster’s name provided a real, working safeguard in this situation. The “polling establishment” worked; it just didn’t work as fast as Karen Bass’ political operation.

The sequel: Wedgewood Polls

Then came the sequel. On October 20, Wedgewood Polls, a small firm that had launched in June, announced its polling was fake, a “social experiment” to expose the “low barrier to entry in the modern polling industry.” The alibi was nearly identical to Median’s, and so was the sequence: the confession arrived only after outsiders closed in. According to researchers at the University of Washington’s Center for an Informed Public (CIP), Wedgewood’s statement was allegedly a response to FiftyPlusOne, the polling aggregator, which had been publicly flagging inconsistencies in Wedgewood’s methods and data. Within the hour, FiftyPlusOne co-founder G. Elliott Morris was questioning the “social experiment” excuse and pointing to reporting that documented how his team had spent recent days closing in on the fabrication, including its conclusion that the data was likely AI-generated.

The rest of the polling community pitched in. Three minutes after the confession, @PollTracker2024, the same account that had pushed Median’s numbers in August, was calling Wedgewood a bogus firm, and others said they had harbored suspicions. FiftyPlusOne’s digging also reportedly led another fledgling pollster, SoCal Strategies, to scrub its online presence; people alleged to be behind Wedgewood claimed a prior affiliation with it.

The safeguards were not airtight, and that deserves candor. Wedgewood touted that the New York Times, VoteHub, and FiftyPlusOne included its polls in their aggregators. The Times replied that it never used Wedgewood data in its own polling and forecasting; it merely listed a few of the firm’s polls on its tracking pages. MS NOW had reported on the polls and the damage was real: Mallory McMorrow, a Michigan state senator who had been a Democratic candidate, asserted Wedgewood’s “numbers” were likely a catalyst for suspending her campaign as money and support dried up. “It killed us,” she wrote.

Two kinds of poll, one kind of caution

The distinction between the professionals who caught this and the amateurs who spread it maps onto a distinction the public rarely makes: there are two fundamentally different kinds of political polling, and they carry different warranties. Campaign polling commissioned by a candidate or party exists to inform internal strategy and is not designed for public consumption. Sure, those results are often selectively leaked when favorable, and no responsible reporter should automatically treat a campaign’s internal polling as ground truth. Public-consumption polling — the kind many AAPOR member firms, a university center, or a media outlet’s in-house poll conducts — exists to inform the public itself and comes with an obligation to transparency, including disclosed methodology, named researchers, disclosed sampling and weighting, and a track record. Median/Prakash borrowed the visual grammar of the second category to distribute something worse than the first — not self-aggrandizing spin, but pure invention.

The working test involves asking: Is there a survey report with real methodological detail like question wording, sampling, mode, weighting? Is it attributed to named researchers with credentials? Was it conducted independently of a campaign or advocacy group? And is the firm cited by aggregators like the Times, FiftyPlusOne, or RealClearPolitics? A “no” to any of those doesn’t mean the results are wrong, but it indicates there’s a reason to pause and investigate before sharing.

Buyer beware, user beware

That has to become the operating assumption for anyone consuming political data online. And “user beware” applies just as much to the platforms and aggregator accounts functioning as de facto gatekeepers now. An account that auto-posts “publicly available” releases without asking who released them is laundering information, not aggregating it. A candidate’s own tweet isn’t corroboration; it’s an echo chamber with (possibly) better production values. And a broadcast segment built on a campaign’s celebratory press release should probably be sponsored by the campaign, because it’s functionally campaign communication.

When a fake becomes a weapon

Wedgewood also showed what happens when one fake is treated as an indictment of every real poll. Per the CIP’s timeline, the story moved within hours from polling wonks to partisan influencers. On the right, it was framed as proof that polls cannot be trusted, especially polls showing Democrats ahead. The next morning, FCC Chairman Brendan Carr warned of a “surge in entirely Fake Polls” and threatened broadcasters who shared them. House Speaker Mike Johnson argued that Republicans were being under-sampled and urged Republican voters to ignore the polls and vote. President Trump accused the New York Times of being “ARTIFICIAL” for posting the fake polls. On the center-left, the worry was that fake numbers had distorted Democratic primaries. Lost in the noise, the CIP noted, was FiftyPlusOne’s finding that the data was likely AI-generated.

A successful fraud can provide fodder to people who want to discredit every real poll. CIP researchers urge keeping three categories apart to avoid conflation: fake polls built on fabricated data, like Median’s and Wedgewood’s; low-quality polls that were actually conducted but rest on weak methodology; and legitimate polls conducted using sound methods. Only the first is fraud. The second is a quality problem that disclosure standards expose. The third is polling doing its job. Blurring the categories could let gaps between earlier polls and election results be framed as evidence that the results were suspect.

The part that should worry regulators more than editors

When the Times’ Goldberg finally pinned Median Strategies down, the outfit volunteered an odd unprompted disclosure: no one involved, it said, “held or traded any position in an election prediction market, betting market, or other financial instrument tied to the races covered.” Nobody had asked about trading. That’s the kind of denial you issue when you know your handiwork looks incriminating.

What happened in those prediction markets? The California Post’s story ran the night of August 13, and Kalshi’s Bass contract ticked from 64.3 cents to 65.9, peaking near 67.2 before the fabrication was exposed. Polymarket’s move was sharper, climbing from roughly 61 cents to a peak of 70 cents the morning before the Times’ exposure, with the AP counting twenty separate Bass trades within fifteen minutes of the poll’s release. In plain terms, the prediction markets believed the fake numbers before anyone caught the lie.

That, folks, is the anatomy of a pump-and-dump scheme, a penny-stock con where a promoter seeds false “inside information,” rides the price up on other people’s credulity, and cashes out before the truth catches up. Whether Prakash profited or not, as he claims to The Guardian, the episode should be regarded as a proof of concept: a fake poll, dropped at the right moment, can move real money in thin markets where the traders (i.e. “gamblers”) are anonymous and starved for information. Nobody had to hack anything. They just had to post a PDF and get it to “go viral.”

Wedgewood raised the same question. A Bluesky user claimed to have found a Kalshi account tied to Wedgewood showing a $4,764 profit, but no one has confirmed the link, and Kalshi said that Wedgewood’s releases and confession did not “meaningfully” affect its markets. CIP researchers see limited evidence so far that any exposed fake has moved a betting market. The incentive, though, is obvious.

About that “social experiment”

Which brings us to the other unanswered question: if this really was, as Median/Prakash claims, a “short-term social experiment to examine how purported polling information could enter and spread through the political information ecosystem without independent verification,” where are the experiment’s results? A genuine experiment has a stated purpose, a methodology, observations, conclusions, and researchers willing to put their names on the findings. Now that Prakash has put his name to it, the premise is a little clearer. He says the Michigan Senate miss is what got him wondering how easily bad numbers travel, but the rest of the scaffolding is still missing: no methodology, no data, no debrief, nothing beyond his own account of being surprised at how fast it spread. A name attached to the theory isn’t the same as evidence for it. Wedgewood recycled the alibi two months later, adding a call for “major reform” of an industry whose own watchdogs were, by then, closing in on it. Two “experiments,” two identical alibis, no published results. The American Association of Political Consultants put it more bluntly before Prakash was identified: “Publishing fabricated survey data as independent research is not a social experiment. [It’s] a deception directed at voters, at the press, and at the campaigns that rely on public data to make decisions.”

Where this leaves the profession

The record shows a profession policing itself. Aggregators with transparency standards kept Median out of their averages and asked for methodology that never came. An LA Times reporter caught Median within a week. FiftyPlusOne chased down Wedgewood and published what it found. And when the frauds surfaced, the professional associations drew the line quickly.

“What happened with Median Strategies wasn’t research; it was fabrication dressed up to look like research,” said Anita Watkins, CEO of the Insights Association, the market research industry’s leading trade group, in a statement calling for continued vigilance and transparency. AAPOR’s own statement draws the same line: “It is important that we don’t conflate fraudulent activity like this with legitimate election polling.”

They’re right to insist on the distinction, and the record backs them up. Both frauds were exposed by people who vet polls for a living. That doesn’t make the system foolproof: Wedgewood ran for months, and a few of its polls reached aggregator tracking pages and news coverage. But what Median Strategies and Wedgewood exposed was a flaw in the modern information chain, not professional polling. The right response is more vetting, not less trust in polling. The researchers who vetted, investigated, and spoke up did their jobs. The rest of the chain still needs work.

John Nienstedt is president of Competitive Edge Research & Communication, Inc., a national polling firm he founded in 1987. More information at www.cerc.net.

Photo Attribution: Sanfranman59 — Own work, copyleft: Multi-license with GFDL and Creative Commons CC-BY-SA-2.5 and older versions (2.0 and 1.0)