Real political reporters use data journalism techniques to catch fakers at “Wedgewood Polls”

Mary Radcliffe, Elliott Morris, Cooper Burton, and Fiftyplusone report:

On Tuesday, Oct. 20, FiftyPlusOne reached out to Wedgewood Polls, a self-professed polling firm, with a request for comment on an investigation we had conducted into their alleged survey research. . . . We reached out to them to ask who they were and confirm the methodology statement on their website. We were given two names, Jack Sims and Andrew Prater. We were able to independently confirm that Sims had previous experience in the polling industry and reviewed their methodology statement in detail. It is not atypical for a pollster to leave an existing firm, form their own, and self-fund some surveys to build reputation; we had no reason at the time to believe this case was any different. Wedgewood operated as if it were a legitimate polling firm, including collecting money from donors to conduct crowd-funded polls, and sharing raw survey data (which we now have confirmed was fraudulent) with FiftyPlusOne on Oct. 19, 2026.

How did Radcliffe et al. figure out Wedgewood data were fake?

First, in FiftyPlusOne’s data from this vendor [Pollfish, an online survey platform], the ID variable for a person appears as a roughly even mix between a standard string of numbers (e.g., “1234567890”) and a string of numbers prefixed by a hyphen (e.g., “-1234567890”). Second, the transaction identifier correlates almost 1:1 with the time a person submitted their survey. We confirmed with Pollfish that this is the expected behavior of these columns, and verified that data from another pollster who has used the platform also produces patterns that match ours. However, the respondent and transaction IDs recorded in survey data that Wedgewood provided to FiftyPlusOne do not match these expected patterns. Wedgewood told us it conducted a recent poll of the Mississippi Senate race using the same survey platform that we tested, Pollfish. However, in Wedgewood’s data, the respondent and transaction IDs do not behave the same way they do in FiftyPlusOne’s data, or in data from the third-party pollster we consulted. Additionally, they do not comport with the descriptions of these columns given to us by the survey platform. These are just two of many factors that make FiftyPlusOne suspicious of the veracity of the polls being published by Wedgewood.

Maybe there’s an innocent explanation . . . OK, maybe not:

We detailed these concerns to Wedgewood and, they publicly responded indicating that their polls were indeed fake.

Where did the suspicion come from? Radcliffe et al. tell the story:

Last week, we received a credible tip from a trusted source that Wedgewood Polls might be faking data. We had received tips about Wedgewood in the past, but this particular tip had something crucial that other reports didn’t: evidence. So we asked Wedgewood to provide us with the raw data from their most recent survey, a Mississippi poll conducted Oct. 12 – 14, 2026 . . . We were already wary of this survey. The poll was released just hours after a prominent race rater, Sabato’s Crystal Ball, had changed its rating of this race from Safe Republican to Likely Republican (somewhat of a surprise to elections watchers) and showed Hyde-Smith up only 4 points over Colom. That’s a notable coincidence. When we got the raw data from Wedgewood, we found that it was surprisingly clean. For example, the raw, unweighted data was 67 percent white and 31 percent Black, fairly close to what might be a reasonable statewide estimate of around 60 percent white and 37 percent Black. The design effect for the poll, a measure of how much the pollster had to weight the data in order to get a representative sample, was 1.18, which is fairly low. . . . It’s possible, however, to achieve lower design effects by using quotas or other techniques to construct a more representative sample at the time of recruitment. . . . We double-checked with Wedgewood that they didn’t use quotas to create their sample, which could explain the low design effect. A quota is a mechanism by which a pollster can constrain the sample, to ensure the panel provider returns a certain minimum number or percentage of respondents in each demographic group, and appears to be the only way Pollfish allows researchers to control the demographics of the sample during fielding. They said they didn’t, which made the low design effect especially unusual. But we were still left with a question: Was there anything that actually proved the data was fabricated? That’s when we decided to try to replicate their survey, to the last detail. . . .

Lots more information at the link.

Radcliffe et al. then step back to survey the bigger picture:

It’s getting easier to conduct low-quality polls — or fake them altogether. . . . the wrongdoing we have uncovered on the part of Wedgewood Polls is just the most recent in a series of cases of pollsters manipulating and fabricating data. . . . AI-powered tools for generating content such as websites, PDFs, and spreadsheets have made it easier for those willing to act in bad faith to produce potentially bogus documentation of results that match the quality of those produced by legitimate research firms.

This reminds me of our recent post, It’s fine to use computer-generated survey responses—if you don’t care about the data anyway!

In this case, “Wedgewood Polls” responded with a social media post describing themselves as “an experimental project to demonstrate the vulnerabilities, lack of oversight, and overall low barrier to entry in the modern political polling industry.” Maybe so, but it’s kind of suspicious that they happened to release this statement on the very day that they got caught.

I continue to think there are too many polls, even just counting the legitimate ones. As I wrote back in 2004, and it’s so much truer now:

The U.S. is over-polled. You might have noticed this during the recent election campaign when national polls were performed roughly every 2 seconds. . . . My complaint is not new, but this recent campaign was particularly irritating because it became commonplace for people to average batches of polls to get more accurate estimators. As news consumers, we’re like gluttons stuffing our faces with 5 potato chips at a time, just grabbing them out of the bag.

There are two problems with all these polls. First, they’re a waste of time and money. Millions of people wasting time answering polls. Instead they could be planting flowers, cooking dinner, enjoying themself in some way. The second problem is that all these useless polls have decreased people’s willingness to participate in surveys (for example, of public health) that could have general benefit.

The problem with horserace journalism is: first, journalism resources are limited, and effort that goes into horserace journalism could crowd out more valuable journalism; second, it can degrade public discourse.

The connection of too many polls and too much horserace journalism is that they feed on each other: the polls provide headlines for horserace journalism, and horserace journalism motivates more polls.

Throwing fake polls into the mix makes things that much worse. Good job by Radcliffe et al. to catch these particularly noxious offenders.