Fraudulent Data Influencing Decision-Making: AI, Fake Respondents, and the Future of Public Opinion Research

 

By Mariana Meneses

Society increasingly relies on online behavioral data to understand what people think, but we are entering an era in which distinguishing humans from machines is becoming systematically harder.

The withdrawal of a widely publicized 2024 survey claiming a rise in church attendance in Britain became a warning sign for the polling industry, exposing how AI-assisted fraud, bogus respondents, and organized “survey farms” can distort online research and shape public narratives before the data is fully scrutinized. Experts cited in The Guardian’s report warned that paid respondents increasingly use AI tools to generate convincing but unreliable answers, particularly in opt-in surveys vulnerable to fake identities and coordinated manipulation. Rapidly advancing AI systems can evade detection, imitate demographic profiles, and even infer researchers’ hypotheses, raising concerns about the future reliability of online public opinion research.

Polls influence campaign strategies, media coverage, investor expectations, public policy, scientific research agendas, and even citizens’ perceptions of what other people believe. If large amounts of fraudulent data enter these systems, decisions may increasingly be based on measurements that no longer reflect real human attitudes.

In the November 2025 research article The potential existential threat of large language models to online survey research, Sean J. Westwood, from the Department of Government at Dartmouth College, argues that advanced large language models pose a fundamental threat to the reliability of online surveys across the sciences, including political polling, psychology, economics, and public health.

Westwood designed and tested an autonomous AI “synthetic respondent” capable of completing surveys with highly coherent, human-like answers while successfully evading nearly all current fraud-detection methods. Using demographic personas and memory of previous responses, Westwood’s system generated internally consistent answers across political attitudes, psychometric scales, socioeconomic questions, and open-ended prompts, while also mimicking human typing behavior, reading times, misspellings, and educational differences in writing style. Across 6,000 trials of standard attention checks, the synthetic respondent achieved a 99.8% pass rate, successfully bypassing logic puzzles, instruction-following tasks, and questions designed specifically to expose nonhuman respondents.

Westwood warns that these systems create risks extending far beyond academic data quality problems.

The study demonstrates that synthetic respondents can be deliberately instructed to manipulate polling results, distort measures of public opinion, and even infer researchers’ hypotheses in ways that artificially confirm expected findings. In experiments, simple prompt instructions dramatically shifted responses about U.S. political candidates, presidential approval, and geopolitical rivals while preserving plausible demographic and partisan profiles, making manipulation difficult to detect. The author argues that such capabilities could transform survey fraud into a scalable tool for information warfare, capable of influencing polling results, media narratives, and democratic decision-making.

Westwood concludes that many commonly used safeguards, such as including attention checks, behavioral flags, completion-time metrics, and CAPTCHA-protected survey environments, are increasingly ineffective against modern reasoning-based AI agents, creating what he describes as a potentially existential threat to unsupervised online research. Westwood calls for new standards of survey validation, greater transparency from panel providers, and reconsideration of the scientific community’s reliance on low-barrier online survey methods.

Westwood’s synthetic respondent is not an isolated warning.

In the research brief “Fraudulent Respondents and Bots in Nonprobability Surveys: A Literature Review,” researchers at the National Opinion Research Center (NORC) at the University of Chicago argue that online surveys are already facing a growing crisis driven by fraudulent respondents, automated bots, click farms, and AI-assisted participation. Reviewing dozens of studies, they cite cases in which more than 80%, 90%, or even 95% of responses were estimated to be fraudulent or bot-generated.

The authors describe an increasingly sophisticated ecosystem in which AI-generated answers, automated bots, and “hybrid” schemes combining human and machine participation make fraudulent activity difficult to detect. They argue that traditional defenses such as CAPTCHAs and hidden trap questions are becoming less effective, while stricter screening can wrongly exclude legitimate participants. As a result, they recommend multilayered detection strategies combining behavioral analysis, survey metadata such as timing and clicking patterns, human review, and study-specific detection systems, and conclude that probability-based surveys with verified recruitment procedures remain substantially more resilient than open-access online surveys.

One environment that has attracted particular attention is Amazon Mechanical Turk (MTurk), a crowdsourcing platform where participants are paid to complete tasks such as surveys.

In the article “The threat of AI chatbot responses to crowdsourced open-ended survey questions,” published in the Elsevier’s journal Energy Research & Social Science in 2025, Frederic Traylor, from the Department of Sociology atRutgers University, examines how tools like ChatGPT are contaminating open-ended survey data. Comparing responses about hydrogen energy collected through a managed online panel and MTurk, he found that MTurk answers were longer, more polished, and more on-topic, but also substantially more likely to resemble AI-generated text. Human coders identified 7.5% of MTurk responses as likely AI-generated, compared with just 0.8% in the managed panel, while computational analyses found significantly higher similarity between MTurk responses and outputs from ChatGPT, Bard, and Bing AI. Many of these suspicious responses still passed traditional quality controls, including attention checks, IP verification, and logic-based screening.

The paper argues that large language models create a new form of survey fraud because they allow respondents to fabricate convincing answers instead of leaving blanks or giving obviously low-quality responses. This is especially problematic for exploratory social and political research, where open-ended questions are used to capture authentic opinions, emerging concerns, and diverse perspectives.

Rather than expanding insight, chatbot-generated answers risk reducing response diversity by recycling formulaic patterns from AI training data while potentially reproducing embedded social biases. Traylor concludes that existing methods for detecting bad survey responses are no longer sufficient and calls for structural changes in survey design, including anti-copy-paste measures, explicit prohibitions on AI use, and improved detection strategies for distinguishing human from machine-generated responses.

Yet the story may be even more complicated than fraudulent AI responses. In the article “The Bots Ruining Social Science Aren’t Bots At All” (pdf available), published in the journal Perspectives on Psychological Science in 2026,Shalom N. Jaffe and colleagues challenge the assumption that poor-quality survey data primarily comes from automated bots.

Drawing on four studies conducted between 2018 and 2023 across platforms such as MTurk and Lucid, they argue that many problematic responses originate from humans using VPNs that mask their locations, rented U.S. accounts, translation software, and increasingly AI tools to disguise their identities and qualify for paid surveys. The evidence ranged from respondents who used Indian English terms such as “brinjal” instead of “eggplant” despite claiming to be U.S. residents, to nonsensical open-ended answers and strong acquiescence bias (i.e., a tendency to answer “yes” to nearly any question to remain eligible for future studies). They also noted the anchoring effect, a recognizable human cognitive pattern in which exposure to an initial number influences later estimates. Video interviews with suspected “bot-like” participants confirmed that all were human.

 

 

Anti-bot measures alone cannot solve a problem driven largely by organized human fraud. The authors make this argument after documenting online markets for buying and renting survey accounts, guides for bypassing identity checks and geographic restrictions, and even criminal operations that generated millions of fabricated responses. They recommend focusing on behavioral indicators such as suspicious response patterns, low-quality open-ended answers, copy-and-paste activity, typing speed, and digital fingerprinting. Although generative AI is making fraud more sophisticated, the authors conclude that there is “almost always a person behind the screen,” making human behavior itself one of the most important clues for protecting research integrity.

What if the same technologies that can fake survey respondents can also fake public opinion itself? The implications extend beyond corrupted surveys.

In the article “How Malicious AI Swarms Can Threaten Democracy,” (pdf available) Daniel Thilo Schroeder and a large international group of researchers argue that the combination of large language models and autonomous multiagent systems could enable a new form of information warfare capable of manufacturing the appearance of public opinion itself. The authors describe “AI swarms” as networks of coordinated AI personas that maintain persistent identities, adapt in real time, mimic human behavior, infiltrate communities, and operate across platforms with minimal human oversight. Unlike traditional botnets, these systems could autonomously tailor messages to different audiences, conduct millions of rapid experiments, exploit social network structures, and create the illusion of widespread agreement, a phenomenon the paper calls “synthetic consensus.”

The authors warn that AI swarms could amplify echo chambers, spread hyper-personalized disinformation, orchestrate large-scale harassment campaigns, and erode the “wisdom of crowds” by undermining the independence of public opinion. They also highlight the risk of “LLM grooming,” in which large volumes of synthetic content are injected into the internet and later incorporated in the training data of future AI systems. More broadly, the paper argues that growing uncertainty about whether online speech originates from humans or machines could produce “epistemic vertigo,” which means a collapse of trust in media, institutions, and even ordinary users. To mitigate these risks, the authors propose a multilayered governance strategy combining detection systems for coordinated inauthentic behavior, transparency audits, identity-verification technologies, independent monitoring initiatives, and carefully regulated defensive AI systems.

“For the companies that provide online participant pools and survey tools, Westwood’s paper served as a “warning shot,” says Andrew Gordon, a researcher at Prolific, a platform that manages a vetted pool of participants for online behavioral research” – Cathleen O’Grady (pdf).

Can online research be saved, and what happens next?

Not all researchers believe the problem is unsolvable. In the Science news article “AI Threatens to Upend Online Social Science Research” (pdf available), Cathleen O’Grady describes growing concern that increasingly sophisticated AI systems may make online surveys and behavioral studies fundamentally unreliable. Although companies such as Prolific and CloudResearch are developing advanced detection systems based on behavioral signals such as mouse movements and interaction patterns, several researchers interviewed for the article question whether online research can remain trustworthy in its current form. Some warn that the era of cheap, large-scale online data collection may be coming to an end.

In the article “Survey-taking AI tools surpass human abilities. Here’s what we can do about it” (pdf available), Folco Panizza, Yara Kyrychenko, and Jon Roozenbeek argue that protecting online research will require continuously evolving defenses rather than reliance on traditional safeguards such as CAPTCHAs and attention checks. The authors propose combining probabilistic detection of AI-like response patterns, analysis of survey metadata such as keystrokes and response times, greater use of identity-verified participant panels, and a new generation of “reverse CAPTCHA” tasks designed to exploit persistent differences between human and machine reasoning. Examples include estimation puzzles, logic traps, and tasks in which distinctly human mistakes become signals of authenticity. The authors conclude that abandoning online surveys is unrealistic, but that preserving research integrity will require an ongoing technological arms race between detection systems and increasingly capable AI agents.

 

 

Some researchers are exploring a very different possibility: using AI not as a threat to polling, but as a polling tool itself.

In the article “Artificially Intelligent Opinion Polling,” Roberto Cerina and Raymond M. Duch present a method that uses large language models to transform social media activity into election forecasts. Using posts from 𝕏 during the 2020 U.S. presidential election, the authors asked LLMs to infer characteristics such as demographics, geography, partisanship, and voting behavior from users’ digital traces, then combined those estimates with multilevel regression and post-stratification (MrP), a statistical technique used to adjust biased samples so they better reflect the broader population. Despite relying on highly unrepresentative online data, the resulting state-level forecasts were comparable to leading polling aggregators such as FiveThirtyEight.

The approach could dramatically reduce polling costs by eliminating many traditional expenses associated with survey recruitment, fieldwork, incentives, and multilingual operations. However, the authors also identify major limitations. LLMs may suffer from data leakage, in which models indirectly incorporate information about outcomes they are supposed to predict, while the growing presence of bots, coordinated manipulation, and AI-generated content could distort the relationship between online discourse and real public opinion. The authors note that platforms such as MTurk have already experienced significant contamination from people secretly using LLMs to complete surveys. More broadly, they argue that AI polling is only feasible when sufficient digital traces exist, relevant demographic characteristics can be inferred reliably, and external benchmarks are available to correct the substantial biases inherent in social media data.

Synthetic data is not inherently fraudulent. In fact, researchers and companies increasingly use artificially generated data for legitimate purposes.

In the article “Big Data Analytics and AI for Consumer Behavior in Digital Marketing: Applications, Synthetic and Dark Data, and Future Directions,” Leonidas Theodorakopoulos, Alexandra Theodoropoulou and Christos Klavdianos examine how synthetic data is being used in AI-driven marketing and consumer analytics. Rather than copying real individuals, these datasets are generated through techniques such as Generative Adversarial Networks (GANs), in which competing neural networks produce increasingly realistic data, and agent-based simulations that model the behavior of virtual consumers.

The authors argue that synthetic data can help train AI systems, test hypothetical scenarios, simulate rare events such as fraud or customer churn, and facilitate data sharing under stricter privacy regulations. However, they also warn that synthetic datasets may fail to capture complex human behaviors such as trust, emotional responses, and changing preferences over time while reproducing or amplifying existing biases. Some systems may even leak information about real individuals or create models that perform well in simulations but poorly in real-world settings. As a result, the authors conclude that synthetic data should be viewed as a complement to, rather than a replacement for, real human behavioral data.

Can we trust any polls at all? The danger is not that all polls are equally broken, but that the gap between stronger and weaker methods is becoming more consequential.

In a Pew Research Center Q&A by John Gramlich, Courtney Kennedy, Pew’s vice president of methods and innovation, argues that rigorous probability-based polling remains fundamentally different from vulnerable opt-in surveys, where anyone can sign up online and where fake identities, bots, or paid respondents can more easily enter the data stream. Kennedy also rejects “silicon sampling,” the idea of asking AI systems to simulate public opinion instead of surveying real people. The objection is not only technical, but democratic: polling exists to give actual people a voice in public life. AI-generated estimates can stereotype demographic groups, underestimate social disagreement, and misrepresent political viewpoints, rather than measuring what real citizens think.

Pew’s response to the problem is rooted in how respondents are recruited. Rather than relying on open online sign-ups, the organization uses probability-based sampling, a method in which people are randomly selected from the broader population, often through real-world residential address lists. Because participation is tied to verified recruitment procedures rather than self-selection, large-scale fraud becomes far more difficult, costly, and logistically complex.

Kennedy cautions that probability-based polls are not immune to error: poor questionnaire design, inadequate weighting procedures (statistical adjustments used to better match the population), or other methodological mistakes can still produce inaccurate results. However, in an era of AI-assisted fraud, rigorous sampling, controlled recruitment, and multimodal participation (i.e., allowing people to respond through different channels such as mail, phone, or online surveys) may represent some of the strongest safeguards available for ensuring that measures of public opinion reflect the views of real people rather than synthetic respondents.

 

 

The solution is not to abandon surveys, but to improve them.

The controversy that opened this article illustrates both the vulnerability of online surveys and the possibility of correction. Following concerns about fraudulent respondents, YouGov reanalyzed its widely discussed 2024 Bible Society survey and found that some of the demographic groups highlighted in the study, particularly younger churchgoers, contained enough fraudulent responses to affect key findings, including the reported increase in church attendance from 8% in 2018 to 12% in 2024. Nadeem Badshah and Sinéad Campbell for The Guardian interviewed David Voas, a quantitative social scientist and emeritus professor at University College London, who said YouGov used online opt-in surveys that allow participants to self-select for the study.

The company acknowledged that available anti-fraud tools had not been deployed optimally and that the survey’s unusually complex sampling design, which combined quotas for age and ethnicity while oversampling regular church attendees, created opportunities for fraudsters to target harder-to-reach groups.

YouGov took responsibility for the issue and noted that online research has faced increasingly sophisticated attacks in recent years. In response, the company strengthened its defenses through enhanced identity verification, duplicate detection, behavioral monitoring, and AI-assisted fraud detection systems. It also emphasized that it maintains and continuously monitors its own participant panel rather than relying on third-party recruitment networks, and announced plans to repeat the study using updated methods.

Online surveys are entering a period of structural instability driven by increasingly capable AI systems, organized fraud networks, and synthetic respondents that can imitate human behavior with growing realism. Traditional safeguards such as CAPTCHAs, attention checks, IP verification, and static screening rules are becoming less effective against systems capable of generating coherent answers, maintaining consistent identities, and adapting to detection methods in real time.

The challenge extends far beyond polling. Surveys inform political campaigns, public policy, scientific research, public health, and perceptions of what other people believe. As a result, contaminated data can distort scientific findings, measures of public opinion, and even democratic discourse itself. Yet the literature does not suggest that all polling is broken. Rather, vulnerability varies dramatically across methodologies, with probability-based surveys and controlled recruitment remaining substantially more resilient than open-access online models.

Researchers increasingly argue that protecting research integrity will require stronger identity verification, stricter recruitment controls, multilayered detection systems, and continuous adaptation as AI capabilities evolve. In digital environments increasingly saturated with synthetic behavior, verifying that a response came from a real human may become one of the central methodological challenges of our age.

---

Craving more information? Check out these recommended TQR articles:

• Thinking in the Age of Machines: Global IQ Decline and the Rise of AI-Assisted Thinking
• Digital Sovereignty Movement Grows as Global Infrastructure Concentrates Under Few Companies
• Nations Adopt Digital IDs for Citizens, While Critics Highlight Privacy Issues
• Cleaning the Mirror: Increasing Concerns Over Data Quality, Distortion, and Decision-Making

---