As I’ve reached the biblical three score and ten years of age, I’ve gained increasing appreciation for the old saying that history doesn’t repeat itself, but it does rhyme.
I was thinking about this phrase after I returned home from IIEX in Austin a few weeks ago, where, as others have discussed, AI featured in many talks. There was a lot of exciting work discussed, either work in progress, or more finished work being hawked as ready for prime time. As part of my consulting work over the past few years, I’ve worked with LLMs in social media data analysis, and have seen the power of what they can do. There is enormous potential for AI improving the operational side of market research practice, and even more importantly, the substance and actionability of what we do.
However, after mulling it over in the weeks since, an experience from my distant past came to me that should perhaps give us pause.
In the late 1960s and early 1970s, there was a line of research that caused quite a stir in psychology, anthropology and linguistics – can chimps (and other primates like gorillas) learn to “speak” using American Sign Language (ASL)? One of the distinguishing characteristics of human beings is our ability to speak, to use words and sentences to express ideas and concepts. Chimps don’t have the right vocal apparatus for human-like speech, but maybe they could express ideas through ASL. If so, then maybe we aren’t so special after all (a bit reminiscent of the discussions about whether AI is sentient, and what that means for us).
There were several teams that claimed to have great success, and it caused great excitement in the scientific and popular press. Film clips of Washoe the chimp and Koko the gorilla signing away were shown all over and many articles appeared. The lead researchers were even claiming that the primates could even combine signs to express new concepts a single sign couldn’t express.
However, in the mid-1970s, I attended a symposium of the leading researchers in the field, and one of the speakers was Herbert Terrace of Columbia University’s Psychology department (where I was studying for my Ph.D.). He was very critical of the other research teams, saying that they hadn’t published enough hard data to really evaluate the case they were making. Without comprehensive data, others couldn’t evaluate whether they were just cherry-picking the most attention-grabbing examples for public viewing, and leaving out “signing” that didn’t really signify anything.
Terrace went on to start his own research program, acquiring a chimp named “Nim Chimpsky” (a play on the name of the famous psycholinguist Noam Chomsky). While I didn’t work on the program, I did meet the young fellow several times in the hallways. The research program lasted a few years, but had to be shut down because of funding and other issues (as recounted in the book written some years later called “Nim Chimpsky: The Chimp Who Would be Human” by Elizabeth Hess).
While the program was functioning, there seemed at first to be good progress – Nim seemed to learn a number of signs. Terrace submitted a paper to Science magazine, but then made the highly unusual decision to withdraw the paper. Terrace realized that their analysis left out something critical. The entire focus was solely on Nim, and didn’t take into account the larger context of the human-Nim interaction. If you took a “wide-angle view,” you could see that Nim didn’t “know” what he was signing, he was making hand gestures to try to get food rewards from the human. No food, no signs. In 2019 in his last writing on his experiences, Terrace concluded that not only couldn’t chimps learn sentences, they couldn’t learn individual words either.
As I thought about all this, it raised some key issues for me about the development of AI-based services in market research:
- Be wary of the shiny object. Don’t let your enthusiasm over a new technology overcome your scientific skepticism.
- Are the case studies being presented in sales or at conferences really representative of what a given company’s AI offering can do?
- I recently was emailed something from a market research company about their new AI offer that summarizes and extracts insights from qualitative transcripts. They claimed that you can expect the system to deliver at least one insight that human analysts can’t find. Really? Is there a way for me to go back through the transcripts and see where the “insight” came from? Or is the system hallucinating? You don’t want to be like the lawyer who is in hot water because he used ChatGPT for his research for a brief, which then made up a bunch of cases out of thin air.
- Are there insights that an experienced human researcher would find that the AI system misses?
- Validate, validate, validate
- If someone comes to you to try to sell any kind of AI product, ask to see their validation research – not just a single case study (and if you’re trying to sell one, show your validation studies before you’re asked). If I was a buyer and someone came to me without real validation research, I’d just say thanks, come back when you have it.
- Without quantitative measures of accuracy, and a real understanding of how the AI models were created, you just don’t know what you’re dealing with:
- What data have been used for training? How much? Are there any biases in the training data that can cause problems or failure down the road? How often is the model updated?
- One of the most valuable of the AI presentations at IIEX was by Lorin Drake of Publix, who worked with his team to divide a research project into different phases to test how well ChatGPT could do each one. They graded each phase, and most got B’s, but some got A’s. This gives you a solid view into where you could let ChatGPT rip, and where you need to be sure to give it “human help.”
- Don’t anthropomorphize
- Just like the researchers who fell in love with the idea that their chimp or gorilla could communicate, don’t fool yourself into thinking that any given AI product is more than it is. It isn’t another person on your team (and if you’ve seen the movie “Her”, don’t fall in love with it). It is a statistical model that makes predictions about what it should output based on what you input. And, as the saying goes, all models are wrong, some are useful.
There is a tremendous amount of momentum building towards incorporating AI into market research practice. The enthusiasm is great to see, and if channeled well, will lead to services that can greatly benefit business. My concern is that the enthusiasm will cause people to let their guard down and ultimately lead to embarrassment. So, let’s be real researchers and do the right thing.