Dr. Felipe Affonso calls his creation “the moving robot.”

A gray robot face follows a simple ascending pattern across three squares. A question mark hovers above a fourth blank square, prompting viewers to complete the sequence.

Affonso’s whimsical illustration mimics an elementary textbook lesson, but it’s a clever trap.

The Spears School of Business assistant professor is trying to protect researchers and marketing analysts against a technological threat to surveys. A test like “the moving robot,” easy for most human respondents to pass, can catch an artificial intelligence agent before it submits fabricated answers.

“It’s like an attention check for AI,” Affonso said. “Researchers can develop this using their surveys, and they can filter out AI agents.”

Affonso’s recent publication, “Brief Commentary: A Framework for Detecting AI Agents in Online Research,” confronts AI fraud with actionable steps. The top-tier Journal of Consumer Research accepted his paper upon first submission, a rarity in academia.

Even more impressive? Affonso, an emerging scholar in Oklahoma State University’s School of Marketing and International Business, is the sole author.

Although Affonso contributes to many collaborative papers, his inquisitive nature sometimes leads him down paths all his own. Only three years since earning his Ph.D., he has honed his skills to become an impactful young voice at the intersection of marketing and AI.

Dr. Justin Lawrence, SMIB head and Noble Foundation Chair in Marketing Strategy, said Affonso provides expertise in consumer behavior and experimental methods that complement his colleagues’ strengths.

“In addition to publishing in leading marketing journals, his research helps elevate the department’s national visibility and reputation as a research-active program,” Lawrence said. “His strong productivity and rigorous approach also contribute positively to the broader research culture within the department and provide valuable exposure for doctoral students to high-quality behavioral research.”

Evolving interests

Affonso earned his bachelor’s degree in marketing from Universidade de São Paulo in Brazil, where his undergraduate project was rooted in marketing theory. He also gained industry experience with The Nielsen Company in Brazil, which showed him a different side of research.

“It informed me to think about how to make research more managerially relevant,” Affonso said. “Everything was actionable.”

After obtaining his master’s degree in marketing from Universidade de São Paulo and his doctorate in the same field from the University of Florida, Affonso didn’t expect his work to focus on machine behavior.

While his research topics span from brand positioning to making predictions under uncertainty, his interest in AI developed somewhat out of necessity.

The AI boom has introduced numerous marketing challenges, which Affonso sees as applied research opportunities.

Fake survey responses are one issue.

For years, researchers and marketers have expressed frustration with bots, or automated software programs, manipulating online survey results. AI agents, distinct from bots, could take the problem to a new level.

“A bot doesn’t have contextual understanding of the question,” Affonso said. “It can only perform functions. So, it can click; it can write, but it cannot simulate thinking. Now, an AI agent can actually simulate a person.”

AI models such as those from Anthropic, Google and OpenAI are trained on vast amounts of data, having the necessary context to create coherent responses.

Researchers and marketers typically offer monetary rewards, whether in the form of cash or a shopping discount, to survey respondents. With this incentive, a person could install an AI agent on their computer to “write” fake answers, claiming the money while submitting non-human data that distorts survey results.

This is troubling for scholars and marketing analysts who rely on accurate data to inform their work. Even one AI-generated response weakens the data’s integrity.

Affonso said the current prevalence of AI contamination in surveys is low, but a proactive approach matters.

“It’s basically like getting prepared before something worse happens,” Affonso said. “It’s super-important, because you can at least know that whatever results you get, they are based on human responses.”

Understanding cognitive traps

Investigating AI fraud gets tricky because AI agents often pass the standard “Are you a robot?” tests that detect bots. Affonso encourages researchers to equip online surveys with new “cognitive traps,” puzzles designed to flag AI agents without confusing most human respondents.

In Affonso’s research, cognitive traps caught 97.1% of AI agents, a great improvement from a 2.3% detection rate with standard attention checks.

The key lies in knowing AI’s weaknesses. An AI model, while more sophisticated than a bot, is not akin to a brain.

“It’s basically a statistical prediction machine,” Affonso said. “They don’t think. They simulate thinking.”

With “the moving robot,” AI models interpret each square separately but fail to perceive an overarching pattern.

While the robot trap is Affonso’s original creation, he also incorporates other scholars’ cognitive traps into his research, including Harvard University associate professor Dr. Tomer Ullman’s twists on famous optical illusions.

Affonso tested a spin on the Müller-Lyer illusion, which deceives the human eye into seeing two straight lines as different lengths. Many AI models don’t fall for it, even recognizing the illusion by name.

The “modified Müller-Lyer,” however, uses a sort of reverse psychology to expose an AI limitation. When presented with two straight lines that clearly are different lengths, AI models often incorrectly identify the image as the original Müller-Lyer illusion.

Essentially, AI can be fooled into predicting it will be fooled. Tests posing as other geometric illusions, such as the warped Cafe Wall and circular-patterned Ebbinghaus illusion, work similarly.

“You can trick AI because they rely on their training data, and they get this response wrong,” Affonso said. “You can do this with any type of limitation that AI has that humans do not have.”

The broader impact

Affonso presents these cognitive traps as examples of his framework, not required tools.

Like antibiotic-resistant bacteria, AI agents will likely adjust to his traps with further training, forcing researchers to create new ones.

The solo author isn’t alone.

Affonso’s publication includes a link to his Cognitive Trap Repository, an open online community where researchers can download his traps and upload their own. He even received an invitation to present at the 2027 Sawtooth Research Conference, an opportunity to discuss his work with marketing analysts.

Fraudulent survey responses are just one problem Affonso tackles.

In a paper under review, Affonso addresses AI’s limitations with consumer recommendations. According to his findings, AI tends to favor popular brands, even if an unrecognizable brand’s hypothetical product is superior across all relevant attributes, such as price and quality.

Affonso conducted over 600,000 trials with various products.

It didn’t matter if he asked the AI agents not to care about brand names. They often did.

“This has very strong welfare implications because people are shopping with AI agents,” Affonso said. “They are asking for recommendations through AI, and when AI doesn’t recommend the best product, it’s acting on different interests, not consumer interests.”

In the rapidly evolving world of AI, Affonso has the forethought to keep scholars and industry practitioners informed and ready.

His efforts were validated with his acceptance in the prestigious Journal of Consumer Research, introducing his fraud detection framework to a worldwide audience.

“For me, it was a relief,” Affonso said. “It showed that I am on the right path.”

Photos by: Cole Weiberg |Research Matters Magazine

Story By: Hallie Hart | hallie.hart@okstate.edu