Use Case #3,468 for GPT: Surveys
January 31, 2024
Kevin Schulman, Founder, DonorVoice and DVCanvass
You’ll find what I imagine to be at least 3,467 use-cases splattered across the internet. That makes this example, use case 3,468 and it’s uber novel, using GPT to generate survey research responses.
That’s right, a machine answering survey questions and using a wash/repeat method to produce a dataset of say, 1,000 responses. That may seem risky, useless or both. But, set that aside for a moment and think about practical benefits:
- Free (or nearly so) data collection.
- Massive time savings. We’re talking seconds versus weeks.
- Ability to do analysis on the data with the same analytical engine – more time and money saving goodness.
The obvious elephant in the room, is the “artificial” data garbage? The gold standard is human survey response and by comparing the two we address the large pachyderm.
This is a brand perception study (Determining the Validity of Large Language Models for Automated Perceptual Analysis) with humans and GPT rating the similarity of car brands. The visual plots these brands in a virtual space based on their perceived similarity by humans (right) and GPT (left). The plots are strikingly similar, over 75% the same in fact.
The plot on the left was free and done in minutes. The plot on the right, not free and weeks to months in the making.
I’m old enough to remember when all survey data was collected in person or over the phone and the advent of internet surveys derided as garbage.
We’ve replicated this approach for big brands with lots of data on internet and small brands with comparably little info on internet and it passes the eye test. The caveats are this,
- This seems to work well (enough) for brand rating and perception studies to see where and how your brand stacks up against competitive set.
- Relatedly, you can query GPT on how it arrived at it’s ratings and get a sense of validation but also qualitative to go with the quantitative
- It doesn’t work to collect direct experience ratings or preference data
- If all you need is good enough, this may suffice.
- If you’re making strategic decisions, you’ll want to pay for the “old-fashioned” way of real humans to pick up the 25% that isn’t correlated
For the interested reader, here is how we do the prompts:
- Assume the role of a survey taker using all your generalized knowledge.
- Please rate the degree of similarity between [insert your brands here]
- Use a 1-5 scale with 1 being not at all similar and 5 being extremely similar.
- Return integers only, 1 for each paired comparison. R
- epeat this exercise 1,000 times and create a dataset.
- Use the dataset to generate a perceptual brand map