Does it take one to know one? 

“Are you a hacker?” 

A few years ago, I was chatting with a data expert about survey fraud prevention.  As he described with intricate detail how click farms and bots operate to pass as “human”—and best practices for detecting and stopping them—I was sure that anyone this knowledgeable must be intimately familiar with both sides of the house.  

Today, I am amused to be undertaking a hacker-like task myself: prompting our own bots (also known as large language models, or LLMs) to provide human-like responses to our survey questions. If it really does take one to know one, then all our training in battling would-be fraudsters has prepared us to sidestep common pitfalls and cut right to building strong, synthetic respondents that don’t just mimic survey-taking behavior but reflect real, human insights.  They will be optimized to tell the truth of the market, drawing from our extensive consumer behavioral and transactional databases to drive accuracy and efficiency. 

Can synthetic respondents deliver accurate and “human-like” survey feedback?  

It’s an exciting time to be in the product innovation space, witnessing the capabilities of Generative AI (GenAI) models improve and demonstrate their utility for empowering human creativity. We’re already seeing how refined prompts paired with high quality market research can help innovators do their jobs more effectively. But that’s just the beginning of their ability to accelerate your product innovation cycles. The next phase? Quickly validating and optimizing the new ideas you’ve generated with predictive, simulated consumer feedback.  

It’s not as far-fetched (or ominous) as it sounds, but most GenAI users don’t have access to the level of quality, detailed data that NIQ can exclusively provide. Our vision is to do more than what we’re seeing in the industry today, which tends to be simple re-packaging of GPT with shiny, new interfaces.

When Artificial Intelligence models are carefully prompted with the types of  comprehensive, granular, and up-to-date data that only NIQ has, they can offer remarkably human responses.

This has big implications for manufacturers and retailers’ product innovation cycles: Simulated consumer research can provide quicker temperature checks for ideas and concepts, saving the time and expense of engaging actual human respondents for research that requires more complex feedback.   

Confronting the duality of synthetic respondents 

But as our efforts to integrate GenAI into our product innovation toolkit begin to bear fruit, we haven’t forgotten the underlying duality of this technology that we’re also contending with: its promise and its misuse. In a business where preserving data integrity is essential, how do we ensure we’re not falling into our own fraud-busting traps or diminishing the quality of real consumer data through AI’s limitations or ineffective use cases?  

Let’s make some important distinctions.  

First, our use of AI is fundamentally different from that of our more nefarious counterparts. Survey farms optimize their outputs to merely seem human, using a blend of technology and minimal human intervention to craft responses that mimic authenticity. In stark contrast, at NIQ we are harnessing real, predictive consumer data and detailed product attributes at a scale that is unmatched, to recreate accurate synthetic responses. This approach is powered by our sophisticated GenAI, which leverages our extensive databases to ensure responses are not only realistic but also genuinely reflective of true consumer sentiment. 

However, our experiments have revealed that today’s GenAI models alone are insufficient for creating truly accurate consumer responses. These models are highly susceptible to biases present in the training data. Without careful control, these biases can lead to responses that sound convincing yet are fundamentally incorrect—essentially, the worst type of output for decision-making purposes. This critical insight sets NIQ apart, as we continuously refine our methodologies to mitigate these biases, ensuring our synthetic responses are both authentic and accurate.

This capability distinguishes us from both unscrupulous actors and other providers who lack access to the breadth and depth of data necessary to generate precise consumer insights. 

 Second, we must also keep the use case in mind, taking care to apply GenAI only when the quality versus risk balance is equal. There are times when synthetic feedback can be extraordinarily helpful in the product innovation process—for example, in early-stage development, when the stakes are lower and there is still time to iterate or decide that the robots might not have gotten it quite right. GenAI is also great for more tactical implementations that are easy to modify — like development of social media posts or other marketing copy.  

But when it’s time to make critical decisions, the go versus no-go, putting dollars behind those big ideas, you can’t chance the uncertainty of whether an AI model was prompted with a relevant and recent dataset that will elicit the right feedback. It’s possible that, one day, the robots will know us better than we know ourselves…but for now we must trust that the complexity of human minds and their choices are beyond the reach of a model trained on publicly available text.  

So, when our clients ask how we’re managing the duality of this technology, the simple answer is … by leveraging our expertise to play both sides. We’re trying to push it to its fullest potential while accepting its limitations and directing it toward the best use cases.  We’re trying to trap its misuse in surveys while building our own army of robot respondents prompted for survey-response excellence.  And above all, we’re working to combine its efficiency with the integrity and reliability of our data, propelling our clients toward better innovations. If it truly takes one to know one, then our years of bot-busting and unparalleled data integrity have us poised to build the best synthetic respondents in the industry.