“AI is very clever until it’s unexpectedly stupid”
Having had the last few weeks to reflect on the recent bout of conferences – QRCA, IIeX, Quirks – it’s clear AI models have arrived in our industry. In my 20 years in the industry, I have not yet seen the enthusiasm and child-like joy of anticipation this technology triggered. This is amazing. Research has been a super conservative, backward industry – protective of its scientific method – and here we are embracing infant technology. I think everyone predicting the demise of MR has just been proved wrong – the industry is ready to change, adopt and succeed.
The energy is astounding on the client side as well. Clients are eager, or are already seizing on the opportunity to have an AI model sift through the zillions of consumer data points across the organization to recognize and match patterns; GPT is already used to summarize FGD&IDI transcripts or open ended survey responses, summarize social media chatter or draft a discussion guide.
All the hype and attention is captured by generative AI – and righteously so. The technology has amazing potential for all. But for qual research in general and for qual social media intelligence the future can only be talked about in the superlatives. This is because AI will need qual insights to work. In fact, AI will need qual more than the other way around.
Let me explain, GPT is no different from the auto-complete function that Amazon or Google has. It predicts the next word based on a myriad of rules. (Those rules can be trained – and that’s where qual comes in.) While it appears to understand instructions – it only approximates them, and will predict the best answer. To put it another way, it can detect emotions, it cannot understand them. It cannot do what Bakamo analysts do naturally, read, feel, interpret, anticipate, observe themselves to shed light on the unsaid, underlying or hidden. This level of AI is referred to General Intelligence, and like the fata morgana of self-driving cars might elude us for a long time, if ever. Dave Carruthers line of thinking that divides research into doing and thinking tasks makes a lot of sense. All the tasks that do not require thinking are ideal for ChatGPT.
Sidenote: The low hanging fruit for Bakamo is the isolation of
authentic consumer voices in social media. One of the reasons
our social media analysis is different, is that it is based on
what real people say; typically only a fraction of data captured
by the listening tools is real; the rest is marketing, spam, ad,
promotional blog posts, and all of that needs to be cleansed for
the content analysis to begin. Our first tests have not
been great - but I’ll keep you in the loop when we make progress;
But back to the story.
Allow me to set the scene with a historic comparison. About 20 years ago when the Internet was young, WebAnalytics and AI bid management technology started appearing on the market. I was working at a startup offering this tech to early adopters. The companies who were quick to adapt the new technology were richly rewarded. It took competitors long years to weapon-up and get the same technology working for them. Once that happened the technology lost its competitive advantage – and the onus was returned to the brand to make a difference. This is the BRAND RENAISSANCE that Marcelo Ferrarini Carneiro spoke about recently on Joaquim Bretcha podcast. The half-life of GPT tech will be a lot shorter: every client can be expected to make use of the technology in the next few quarters. But with everyone having access to the same technology – it will be down to who leverages this technology best, and this puts the brand into the driving seat. I’m using brand and qual almost interchangeably, as a lot of quant data may well tell some parameters, but it’s going to be qual that can talk to a brand’s identity, purpose and charisma.
In technology terms, any AI algorithm is only as good as the data and instructions it is given. Newer AI models will have greater and greater data points to choose from. But for a general AI language model, like ChatGPT, to generate beyond banal outputs, let’s say identify communication hooks for a cereal brand’s target audience, it would be better if it’d based its recommendation on the needs, narratives and language that the target group actually uses rather than generic data it has picked up online. (It is not by accident that companies keep insight reports safely filed on internal systems)
This requires training, priming or prompting as it is now being called. Qualitative facts, narrative economies, as Susan Fader💥 likes to call them, personas with clearly defined needs, behaviors, values… all of that needs to be coded into the model for it to have a chance to make meaningful contributions downstream. The sharper and innovative qual thinking goes in the machine, the better, more specific its outputs will be.
To achieve this AI needs qual – a lot more than Qual needs AI. Quals need to figure out how to make our work – the findings & data & insights – accessible to AI models – to unleashed codified human expertise. I believe helping AI use qual understanding will create a lot more value than using AI in the making of insights. This is the journey Bakamo is on; and I will make an effort to keep you up to date about our progress. But don’t wait: let me know what you think? Can partner to make it happen? Please share, join and help! See you soon!