Springboards.ai co-founder Pip Bingemann has said the growing reliance marketers have on large language models (LLMs) risks completely flattening creativity, with AI-generated ideas increasingly drawing from the same predictable patterns rather than pushing brands into new territory.
Sitting down with B&T at the humAIn conference in Sydney on Tuesday, Bingemann said “the big problem” is rooted in how language models work, with their reliance on existing data and probability making genuinely original thinking increasingly difficult.
While AI tools can of course produce impressive results, he argued that marketers risk mistaking a plausible, well-presented answer for an original idea – leaving brands with the same taglines, visual identities and creative approaches as their competitors.
“The big misconception with language models is that because we’re told that they’re super intelligent, and we use words like thinking and hallucinating, we attach all these human-like words to a language model, and it is not human,” Bingemann said.
“It is not creative, it regurgitates a past. It doesn’t think. It’s literally predicting the next token on a distribution of what those next tokens could be, and that’s all they are. Just the bell curve of likely responses.”
As models are trained on their own outputs, Bingemann argued, their responses can become increasingly similar, pushing unusual ideas out of the picture and leaving the industry with a narrower pool of creative possibilities.
“When a language model is trained on its own data, it flattens out all the responses,” he said. “And so, while the language models are getting more capable, we’re also getting more and more like each other.”
Bingemann said the process risks eliminating the most unusual ideas first, with the less predictable possibilities that can lead to breakthrough creative work disappearing from the models’ capabilities.
“The interesting stuff is in the tails of the distribution, like the 1 per cent or the half a percent idea. They get completely deleted from the language models’ capability of even coming up with those ideas,” he said.
“Everything just ends up becoming more alike,” he said, with models trained to reproduce the patterns of previous outputs.
The problem is already visible across marketing, according to Bingemann, from the language used in social posts to the design choices appearing across brand websites.
“LinkedIn posts [are] the obvious one. Like overuse of emojis, [turns of] phrases. You see it straight away,” he said.
“It’s everything from the images you see, the videos you see, the language you hear about. These models are making everything slowly” look and sound more alike.
Bingemann also pointed to colour palettes, typography and website layouts as signs of increasingly formulaic creative work.
“I see websites now, and I just go, ‘I 100 per cent know that’s AI’, because it’s got the weird spacing, like monotype typefaces that they always use. You can just spot it a mile away,” he said.
The risk, he argued, is that marketers can believe they have landed on something fresh without realising the same idea is being generated and used elsewhere.
“People don’t realise that they’re all getting the same stuff until it’s too late,” Bingemann said. “Until they go, ‘Oh wow, look our competitors got the same tagline as us. How did that happen?’”
How marketers can steer away from AI sameness
Bingemann’s advice is not to abandon AI, but to change the role it plays in the creative process.
Rather than treating a language model’s first answer as the finished product, he said marketers should use the technology to explore a wider range of possibilities, giving human teams more material to challenge, discuss and develop.
He argued that one of AI’s most valuable applications is allowing creative teams to move beyond the handful of ideas they might traditionally have time to develop.
“When I was at agency, we’d pitch three ideas because it takes a fucking shitload of time to come up with three good ideas, do all your matching luggage, and because it’s fun,” he said.
But AI could change that equation by giving teams the ability to explore many more concepts before settling on a direction.
“What I think is interesting is now you can explore more ideas, and it’s in that exploration that you get to good work,” Bingemann said. “Sometimes you don’t know something’s good until you see someone else respond to the idea.”
He said the excitement generated when a team starts building on a concept can be an important indicator that it has potential, something that cannot necessarily be captured by asking a model to produce a single polished response.
“If you use AI’s fodder to feed that conversation, that’s when you get to really interesting stuff,” he said.
Bingemann pointed to the experience of an industry speaker at the conference who described generating a range of ideas before waking up the following morning with a stronger concept.
For him, that illustrates the distinction between using AI to produce an answer and using it to stimulate human thinking.
“Use the models as a way to feed your brain interesting stuff,” he said.
That approach also has implications for how marketing leaders manage AI adoption. While brand guidelines, consistent outputs and responsible use remain important, Bingemann said organisations should not allow the pursuit of consistency to eliminate experimentation.
He argued that teams need both the ability to work within shared brand parameters and the freedom to explore variations beyond their usual creative territory.
“There’s a time when you want consistency and a time when you want exploration,” he said. “I think we want to have one thing that can do both of those things.”
What’s the danger of believing AI is always right?
Bingemann also cautioned marketers against treating AI outputs as evidence of their own creative brilliance.
Because language models are designed to predict plausible responses, their agreeable tone can make mediocre ideas sound far more impressive than they really are.
“I think they should stop pretending they’re human. They’re not. They’re machines,” he said. “They try to predict and try to guess what you want to hear.”
To illustrate the problem, Bingemann described a user who uploaded a three-minute recording of his own farts to ChatGPT, presenting it as a soundtrack he had been working on for a month.
According to Bingemann, the model responded with enthusiastic feedback about the supposed track’s mood and creative qualities, despite the audio being exactly what the user had uploaded.
He said this tendency to affirm users rather than offer meaningful criticism “risks creating a generation of people who become increasingly convinced of the quality of their own work”.
“We’re actually creating the deluded generation, where I think we just raise a whole generation of people who are f**king deluded because they do think they’re f**king amazing.”
