Blogs

By, Piers Lee, Managing Director, Novema Pte Ltd

November 2025

To examine the types of errors that AI can make specifically in the analysis of research findings, Novema used AI to analyse the results from the eleven depth interviews conducted for this research and assessed where the AI either missed or misinterpreted key points.

A human Director-level researcher conducted the eleven interviews for the qualitative phase and personally reviewed all interviews. The interview write-ups were then fed into ChatGPT (the leading AI tool currently used by researchers) to generate key findings. We then analysed ChatGPT’s output to identify weaknesses in its analysis.

These are summarised below:

LOSS OF FIDELITY & CONTEXT:

While the AI findings were generally correct, often the AI does not provide the examples or illustrations to clarify the finding, e.g. it “AI reduces the effort of research synthesis”, but this is a bit vague, and needs to be clarified with examples, e.g. to “synthesize findings across different research reports”.

Another example is not stating the examples provided by the respondent in their use of AI to undertake competitor intelligence and “why some competitors are doing well’ – in this case the respondent stated examples of ‘their investment decisions’ or ‘ability to attract staff’, but these were not given as examples when the AI reported the findings.

Important findings, such as some respondents stating that there is still some fence-sitting around AI, and preference to ‘wait-and-see’, was not stated at all in the AI analysis. Generally, AI is not providing many, if any, illustrative verbatims that help to clarify and bring the findings to life.

AI HALLUCINATIONS:

There were not many ‘AI hallucinations’ in this AI analysis, but it referenced ‘diverse data sets’ when the context was the ability to draw from different reports. AI might have the tendency to inject buzzwords like “diversity” and “inclusion” when the context is not quite accurate.

The AI analysis also determined ‘biases’ generated in advertising copy, but it was less about ‘bias’ but more to do with ‘standardisation’. The AI might have linked an earlier comment about bias and assumed it also applied to the follow-up comment, but this was not correct. In one example, its summary could potentially confuse, e.g. “billing clients disproportionately for minimal AI-assisted work.” This could be read as too limited use of AI, but the actual finding was that AI was doing most of the work with relatively little input from the human researcher.

INABILITY TO IDENTIFY ERRORS:

There could be times where the write ups from the interviews contained typos or errors that were picked up by the moderator who also did the analysis. It might take human judgement to identify these errors. While auto-transcription could get around this, we know that this approach is also prone to errors!

How much does AI know anyway?

Novema also ‘interviewed’ ChatGPT and Perplexity on the same topics, with the same questions put to the human stakeholders.

While they identified a similar range of issues, it was weaker on detail and did not identify some topics. While these could have been identified through these applications via more probing, it does illustrate how the first iteration of AI can be relatively weak.

Examples of questions asked of AI include:

Applications of AI in market research: ChatGPT and Perplexity made no specific mentions of the following:

• Application in creative development, e.g. “sparring partner” and creative development (including generating images for stimulus in surveys), although Perplexity was better here
• The specific application of synthetic data
• AI applications in quality control
• Perplexity did not mention the application of bot moderators

Advantages of AI in research: ChatGPT and Perplexity made no specific mention about:

• AI’s ability to obtain more insight from existing data, e.g. a “synthesis of findings across different reports.”
• The ability of AI to make surveys more engaging with respondents, the potential of AI to attract new talent to the industry and allow research agencies to use AI to extend their services into new areas, e.g. creative development

According to AI itself, the ‘judgements’ that AI cannot make include:

1. Ethical Decisions: AI lacks the ability to make ethical judgments that require a deep understanding of human values, morals, and societal norms. While it can analyse data and identify patterns, it cannot weigh the moral implications of decisions in the way humans can
2. Creative Problem-Solving: AI is limited in its capacity for creativity and original thought. It can generate content based on existing data but struggles to create truly innovative solutions or ideas that require imagination and emotional depth
3. Understanding Context: AI systems often fail to grasp the broader context of a situation, which is crucial for making informed decisions. They can analyse specific data points but may miss subtle interconnections that a human would easily recognize.
4. Nuanced Human Emotions: While AI can recognize certain emotional cues through data analysis, it cannot fully understand or interpret complex human emotions or social dynamics. This limitation affects its ability to make judgments in scenarios requiring empathy or emotional intelligence.
5. Intuitive Insights: Human intuition, shaped by personal experiences and insights, plays a critical role in decision-making. AI lacks this intuitive understanding, which can lead to poor decisions in ambiguous situations where data alone is insufficient.
6. Judgments in Uncertain Situations: AI struggles when faced with uncertainty or incomplete information. In rapidly changing environments where new information is constantly emerging, AI may not adapt quickly enough to make sound decisions.
7. Cultural Sensitivity: AI systems may not adequately account for cultural differences and nuances that influence decision-making processes. This lack of cultural awareness can lead to misunderstandings and inappropriate conclusions in diverse settings.
8. Long-Term Strategic Thinking: While AI can analyse trends and make predictions based on historical data, it often lacks the foresight required for long-term strategic planning that considers potential future scenarios and their implications

All these limitations highlight the importance of maintaining human oversight in decision-making processes where context, ethics, creativity, and emotional intelligence are vital, and AI agrees with this itself!

HOWEVER, many stakeholders in our research recognize that many of these limitations are just temporary, and AI can develop with ever more enhanced capability.