Discussion of the concept of ‘synthetic data’ has been gaining prominence in marketing circles recently – but does it really have any validity for marketers?
Synthetic data is data that is generated artificially based on a smaller set of real-world data. It is already being used practically in training machine learning models. As there are limited pilots of actual self-driving cars active in the world with which to train its self-driving car algorithm, Google simulates scenarios and feeds them into the model. Amazon uses synthetic data to train its voice assistant Alexa, generating endless variations on real voice interactions to enhance its understanding.
How does this relate to financial services marketing? A recent panel at the Market Research Society annual conference focusing on synthetic data featured Rose Tomlins, Head of Brand and Customer Insight at Virgin Money.
Rose noted that the company is already having internal discussions of how synthetic data can help her team, saying that it might be able to provide a customer voice earlier in the planning stages of a campaign.
“We’ve started to look at opportunities where the customer isn’t present but we’d like them to be involved earlier,” she said, noting that it could provide a “sense check” on what questions should later be asked to a real audience.
She went on to add that she is interested in creating a “living persona”, which could serve as a “person in the room challenging you to think differently even before you get to the planning stage”.
In this example, ‘synthetic data’ in a marketing context seems to be primarily a new way of organising existing data. When Rose talks about creating living personas which she can ask questions to while creating a campaign, what she’s really talking about is essentially an intuitive ChatGPT-like interface to summarise existing information from proprietary surveys. This isn’t really making the data itself synthetic – it’s just making real data easier to understand.
This is similar to the application that Mark Ritson discusses in his Marketing Week column of a few months ago – essentially using something like ChatGPT to scrape existing information from across the internet and serve it up in a form resembling survey data. This could be useful, but hardly seems to justify such a futuristic descriptor as “synthetic data”.
But what about the approach Google and Amazon are taking, of actually generating new inputs? The trouble is a self-driving car or voice assistant is quite different to the world of market research. It might be reasonable to assume slight differences in how people drive on a road or how they speak, and it makes sense to train a self-driving car or Alexa to expect these minor differences.
But in a survey context, who has the insight into human behaviour to predict how the views of a random individual human might diverge from the average? How would an algorithm predict these random divergences? And who programmes this algorithm in the first place, potentially infusing it with their own biases?
If you are going to create new data inputs and plan marketing around them, the underlying real data would have to be statistically representative and robust enough that it should serve perfectly well as a piece of research anyway. At worst, synthetic data may simply end up potentially amplifying the biases within a dataset while exaggerating its credibility.
None of this seems to get around the fact that if marketers want to actually know what their customers think, someone will have to ask them at some point in the process.
We may get to a point where advanced algorithms can so accurately understand and predict human behaviour that actual surveys become obsolete. But for now, synthetic data in marketing appears to be more useful as a way of organising data rather than a new way of creating it.
