The founders at Insightology.ai have recently launched a book called ‘How to Become an AI Enabled Market Research Agency’. It’s a gold mine of information and insights for anyone working in market research, whatever stage of the AI adoption curve you’re on…even if you’ve decided to stay well away from AI.
Across the coming weeks, I’m going to post on what I see as the real nuggets, adding my thoughts and experiences along the way.
Find the book on Amazon Follow on LinkedIn
How AI actually works
Taken from Chapter 1: How AI Actually Works:
“ChatGPT, Claude, Gemini and Copilot are products … Strip away the interface and a language model does one thing: it predicts what text should come next. That’s it. That’s the whole trick”
Some people will know this already, some won’t. I didn’t. And the reason I didn’t is that LLMs present themselves back to us in a way that makes it seem like they’re thinking and processing in the same way humans do - but they don’t.
Why this is important: it helps to explain why general LLMs aren’t very good at market research and never will be.
More on that in Post 2.
Post 2The road to AI hallucinations
Taken from Chapter 1: How AI Actually Works:
“An AI system is only as good as the data it was trained on and the data it can reach… set against law, medicine, coding or finance, the web knows very little about our [market research] profession, for the simple reason that the detail behind client work is never published… A related paper by Kandpal and colleagues (2023) is titled, bluntly, Large Language Models Struggle to Learn Long-Tail Knowledge”
The extracts above are pulled from three different paragraphs in the chapter, and together, summarise a fundamental point: the ability of LLMs to accurately predict the words and text that should follow the last are highly dependent on the amount and breadth of available data that they were and are trained on. To this end, they are very, very good at producing accurate outputs on subject areas that are vastly covered in the published domain (across the internet, in books, and other licensed material). Subjects such as law, medicine, and finance do well. Market research doesn’t.
In short, the availability of training data on the market research sector sits on the long-tail of the spectrum, i.e. in relative terms, there’s not that much data to train the models and this is a big problem.
Faced with gaps in training data the accuracy of what it predicts as the most likely next text is vastly reduced. But rather than saying so, it will confidently fill the gaps with the most likely option, even if that text is widely inaccurate. That’s hallucination.
