News

ChatGPT will not replace good media plans, it will only replace the bad ones.

Published
London, 17. March 2023

This week we continue our review of ChatGPT and its impact on our day-to-day work. Benjamin Pearton, our Head of Activation and Analytics examines whether AI can really help improve media planning.

So, this is the brave new world we find ourselves in. Artificial intelligence tools are here to take our jobs. Media planners have been cruelly exposed by the almighty power of ChatGPT and its ability to produce research and insights they would have previously wowed clients with. What’s more, it can do it much, much faster too! It appears we are at the dawn of a new reality in marketing industry.

Or are we?

I would pose a simple question to those of you out there that are a little spooked by ChatGPT (which stands for generative pre-trained transformer); Is it the technology itself that worries you, or your reflections on previous work produced for clients? In my opinion, this is a question of quality over quantity. ChatGPT will not replace good media plans, it will only replace the bad ones.

Asking ChatGPT about an example industry (e.g. car rentals) we can see it can easily set the scene and produce a couple of talking points:

It can suggest multiple audience options for a campaign in this sector such as business travellers or road trip enthusiasts:

It can even put these audiences into a typical funnel (what is broader, what is more qualified) and assign some KPIs based on this:

These, however, really are generalities and things the client would already know most of the time. Splitting audiences into business or holiday makers is pretty obvious, no? It also suggested some things that I would not agree with, such as putting solo travellers in the upper funnel (and therefore less qualified) and putting business travellers in the lower funnel. It provides no concrete reason for doing so.

To understand why this is the case, we need to consider how GPT actually works. We do not need to go into detail here explaining how the models behind ChatGPT might work (some of it is still a mystery) but we can focus on key points:

It is powered by training data

Put simply, it ingests large datasets to form the base of its ‘intelligence’. Some examples of this include ‘books1’ (which scrapes text from books) and ‘WebText 2’ (scraping data from reddit URL links)

It is an unsupervised LLM (large language model)

Whilst it uses fine tuned reinforcement learning (weighting ‘correct’ actions more heavily in the model) it is still ultimately vulnerable to mistakes when this model is combined with the data in point one above. At the core, they are just auto-complete systems. For evidence of this, please see the embarrassing demo of Bard (the Google AI solution to rival ChatGPT) where the software essentially made up an answer to a question about the James Web Telescope.

The dangers of technology like this confidently stating things that are not true is a deeper question than a frothy 500 word article about marketing can handle. The point is that what underpins ChatGPT is missing key elements that we as marketing experts should be providing:

  1. Context & Nuance (we know the reality of the client’s position when it comes to business objectives etc. ChatGPT does not)
  2. Relationship (we know what is more likely to resonate with the clients brand specifically through working alongside them and understanding how they position themselves)
  3. Actual data (we know if things ChatGPT suggests did or didn’t work because we may have already tried them)

For me, this makes it impossible for the tool to get to that next layer and add anything of true quality. Good for ideation, not good for story.

To conclude, we should take the advent of technology like this as a wake-up call. If what ChatGPT produces looks eerily similar to what you produce for clients, use this to push your work to a level it could not possibly replicate. Let the machine do the quantity – you should focus on the quality.

On this page