The unexpected art of AI and language

This is the first in a series discussing some of the takeouts from the Soc•AI• Skylines podcast I’ve been co-hosting with Social Hills founder Christina Vetta.

The idea behind Soc•AI•l Skylines was to speak to people already using AI practically in their day-to-day work, whether that’s creative, marketing or other tangential disciplines, and to learn from their work on the tools.

AI creates a lot of uncertainty. There’s a lot of talk about what it will replace. But new models are only as good as people using them. My experience so far is there’s still a gap between theory and practice, and a lot of people haven’t thought beyond “Create me something.”

Our first episode is with Tomas Haffenden, somebody I’ve known for a long time and is excellent at bridging the gap between theory and execution when it comes to creativity. Tomas’ central point was around language, and how what we say and how we prompt are very different.

You can subscribe to the podcast as there’s a lot of good guests coming up.

Language is a curious thing. It’s continually evolving. Partly out of necessity - new inventions and concepts don’t have language to describe them - and partly through merging of cultures. Sick can mean something very different today depending on the context. Two people can take a different inference from a different sentence.

It matters in communications sectors like marketing and PR because interpretation matters. It matters because a play on language can be the difference between an ok advert and a really good advert. Good copywriters know how to play with language.

It also, unsurprisingly, matters in Large Language Models. But if you’re somebody who takes time to unpick why a LLM returns one response over the other, the way AI treats language quickly becomes very apparent.

Firstly, when we speak, we know the meaning behind what we say, and often use words interchangeably. That’s not the case for LLMs. Tomas makes the point that these are models that have been trained on every dictionary and millions of texts - and have mathematical logic built into them - so have a somewhat absolutist view.

In his example, we may use the words stunning and beautiful and mean the same thing, but using these two words in prompts may significantly shift the output (and this doesn’t take into account the slight differences in how different LLMs are built and trained). Knowing the exact language that gets the quickest and best result is a skill in itself.

Tomas’s example of knolling illustrates this. Knolling is the technical term for what most of us would call flat-lays. Some artists and professional photographers may know the term. LLMs definitely know what knolling is.

Most of us, including Tomas, had never heard of the term. He describes how he spent over and hour trying to find the right prompt to get an image of multiple pairs of sunglasses perfectly spaced, so he gave up. Soon after, he stumbled across the concept of knolling. Two seconds later, he had the image he wanted.

Knolling is a good example of “super tokens” - words that are highly specific and technical, rarely used in everyday language, but are easily understood by an LLM as the “correct” definition. This is where things get interesting. Maths tends to be absolute and LLMs are very good at interpreting data. But language is not maths, which is the bedrock of LLM programming.

Local dialects vary wildly. My wife, from a Sicilian-Calabrian family, has trouble understanding people from the north of Italy; their language is different from the Sicilian-southern Italian dialect her parents use at home. Assuming most AI is developed in Silicon Valley, how do the programmers handle regional dialect?

Say a native Welsh speaker gives a voice prompt in English, but with the back-to-front Welsh-language grammar. Would AI struggle compared to somebody from the home counties? How about Bristol? Or a strong Geordie accent?

There are no easy answers here. A general, rather than specific prompt, often reveals more about the bias of those who built and trained the model combined with those who are prompting.

Two implications around language jump out. One is that no matter how much language evolves, using language LLMs understand produces better results.

The second shows the value of creativity. AI has already started to disrupt creative industries. Nano Banana, Adobe and Canva make it easier for non-designers to design. People who wouldn’t know to begin where to build a website can vibe code. Anybody can create copy in seconds.

There’s a big difference between average AI output and high-quality results. Somebody who knows what knolling is should produce a better and quicker result than a non-designer or photographer. That has an intrinsic value over somebody who just asks for a flat lay or a description and who doesn’t know how to work backwards to ask what prompt they should use.

It’s why experienced creatives who embrace AI should hold an advantage over non-creatives. They know the language to use. They’ve built a career on taking ideas from their head to the page. They can assess the quality of AI output and adjust accordingly, either through AI or by editing and adjusting themselves.

An experienced creative who knows how to use AI can produce high quality work quicker and at greater scale than a marketing manager trying to produce the same end result.

Models evolve. People will get better at learning outputs. Juniors will, as is already happening, be disproportionately affected. But if you don’t know how to structure the language of your prompt then the output is only as good as the input, no matter how advanced the model.

Creativity isn’t always valued. But speed, quality and effectiveness often is. There’s a competitive advantage right now for creatives to learn the language LLMs expect.

Influencer measurement has been woolly for as long as I can remember. This is the first proper attempt at marketing science and impact classification I’ve seen.

A new report from System 1 using Effie data on how good creative multiplies commercial outcomes. Well worth a read.

Denstu recently tried to find a buyer for its international business. Nobody bit. Penri Jones’ analysis of the collapse of the holding company model and erosion of famous brands (mostly by WPP’s hand) is shaping a very different future for agencies.

Mutinex founder Henry Innis with a very thorough dismantling of last-click attribution and the cost to businesses who make it their main measurement. It’s a metric that helps with operational functionality at the end of the journey but cannot tell you anything about incremental sales. As you can tell, I’m not a fan.

Playing us out this week, the strangely named Brigitte Calls Me Baby with I Danced With Another Love In My Dream.

This is a band who clearly loves the 80s. You can get shades of Icehouse and The Cure in this track. As somebody write in the YouTube comments, if this was released 40 years ago, Molly Ringwald would have been dancing to it in a John Hughes movie.