We face a conflict between the economic pressure to produce fast and cheap research results and the scientific need for rigor. Hundreds, if not thousands, of people’s lives can be made in a few minutes by sellers who promise strong results. But these often act as methodological black boxes, producing results that cannot be verified, may contain hidden biases and can quietly mislead decision-making.
The synthetic data market is growing rapidly, estimates are expected to rise from approximately $267 billion in 2023 to over $4.6 billion in 2032. Driven by the demand for instant data in an always-on economy, 95% of insight leaders plan to use artificial data within the next year and the appeal is clear. Speed, scale, cost effectiveness and the ability to generate insights from niche audiences are key factors.
In order to transition from passive testing to a reliable, measurable practice, organizations need to address these risks head on. Several methods can help overcome doubts and create a more sustainable model. It is important to identify key problem areas and deal with them head on.
While cost savings and speed of data acquisition are compelling reasons for adoption, several challenges remain. The most successful organizations understand the strengths and weaknesses of different automation tools and when to use them.
Your customers are searching everywhere. Make sure it’s your product he appears.
The SEO toolkit you know, and the AI visibility data you need.
Start a Free Trial
Start with
General challenges and research methods for implementation
Why mainstream LLMs fail to live up to expectations
Why not just ask ChatGPT your research questions? A common misconception in applied research is that providing an LLM with a detailed backstory guarantees a representative exit. Recent large-scale experiments suggest the opposite.
Preliminary research shows that informing an LLM such as ChatGPT, Claude or Gemini to generate more content per person increases bias/homogeneity instead of creating a diverse set of results. For example, the people who used to predict the results of the 2024 US presidential election (with detailed background information provided by LLM) swept all the states for Democrats and failed to reflect the political diversity of the people.
This situation highlights a problem known as watering down, a pervasive issue in AI that affects everything from facial recognition to artificial intelligence, as LLMs are trained on Internet data that unfairly reflects a Western, educated, industrialized, rich, democratic (WEIRD) worldview. Asking the models to be different people produces a statistical explanation that is filtered by this bias, illegally excluded as AI neutral.
In addition, artificial responders can suffer from the Pollyanna Principle, or the tendency for LLMs to be overly agreeable and biased in their responses to user commands. Most users of productive AI chat rooms have probably experienced this: ideas are met with encouragement as ‘good idea’ or ‘good choice’ rather than objective evaluation.
For example, in a usability test comparing artificial versus human respondents, artificial users reported completing all online courses. Where human users may report dropping out of many online courses, virtual users report completion.
High dropout rates among real users ensured that artificial respondents were trying to say what they thought the testers wanted to hear. This sycophancy can lead to dirty product ideas being validated by helpful AI agents.
Fine tuning provides context that synthetic methods lack
Aren’t LLMs trained in a broad enough knowledge set to generate realistic use cases in almost any situation? The most effective way to align artificial and true respondents is to fine-tune using proprietary data. Although standard LLMs provide basic measurements for existing products, they struggle with new problems and underrepresented segments.
In another experiment, the team queried the basic GPT model about a fictional pancake-flavored toothpaste and directly encountered the Pollyanna Principle. Without training data, the model expected people to like it – in other words, it showed a preference for new things. When the researchers fitted the model to past survey data about toothpaste preferences, the output shifted to the negative.
In one study on the desirability of a built-in projector for laptops, the baseline model overestimated willingness to pay by a factor of three. After fine-tuning with research data on conventional laptops, the error was corrected, and the results were aligned with human benchmarks.
Getting the best results with synthetic
The competitive advantage in synthetic research is not the model itself – which is the property – but the content of the identity it puts. For example, Dollar Shave Club used synthetic panels based on cross-sectional data to verify segments of new customers in days rather than months, obtaining results that mirrored human behavior in a specific segment of the effort.
A few methods can help you get the best results from an artificial intelligence study.
Artificial train, real test
To address some of these challenges, the market research industry has proposed an industry-wide validation method known as train-synthetic, test-real (TSTR). In this way, models are trained on synthetic data and tested for predictive validity against a sample of real-world data. Early results have been positive.
In a study led by Stanford University and Google DeepMind, digital agents trained on interview data replicated people’s survey responses with 85% accuracy and 98% social interaction power.
This approach acknowledges the shortcomings of relying solely on off-the-shelf LLMs as a starting point, and the dangers of taking synthetic results at face value without validation. By using proactive methods and validating with real data, teams can realize time and cost savings while building confidence in results.
Governance and transparency are used
Success with artificial intelligence means that researchers and students cannot accept the fallacy of artificiality – the belief that LLMs have the equivalent of human intelligence and personality traits.
Instead, a robust verification process is needed, supported by governance guidelines, well-documented procedures and transparency in the methods used.
A transparent persona checklist can guide researchers as they engage with artificial personas:
- Application domain: A specific job that a person is meant to do.
- Target population: The target group is a persona that is intended to be developed, instead of relying on general definitions.
- Data availability: Whether existing data sets are reused or modified to create populations.
- Environmental suitability: That test interactions reflect the content of real-world applications.
Transparency solves two challenges. It addresses ethical concerns about disclosure and builds trust by showing how practices work and where they fail. As artificial influence increases, distinguishing between real and artificial content will become important.
Trust but verify
A practical approach to applied research means abandoning the belief that LLMs inherently reflect the science of human psychology and focusing instead on empirical measurement, fine-tuning and transparency.
Synthetic research is effective if you respect its limitations
Synthetic research shows great potential but is a catchy promise. The promise is unprecedented speed and scale and the catch is the risk of bias and illusion.
Acknowledging these challenges and building governance and mitigation measures will help you succeed. This also transforms internal doubt into a systematic management approach that measures efficiency and results, creating success.
