Two Buzzy Terms You Should Know
-
Frictionless Decisions brings you counterintuitive, original, jargon-free ideas for connecting data to decisions. Every week.
AI can tempt you to blindly hand over not just your reading and writing to it, but also all your decisions and creative work. It’s so powerful, you might not even recognize the difference between its “average answer“ and a truly creative alternative.
Can AI really surpass human intelligence? Who knows, but two massive studies suggest a simple answer. No. And they’ve given us some fun new words to explain why: trendslop and visual elevator music.
The first study tests large language models (LLMs) on strategic business decisions, while the second tests AI image generators for creativity in visual art.
Trendslop
Strategy always involves tradeoffs; some ideas are better than others, and you need to decide between them. Strategy “trendslop” is the “propensity for AI to opt for buzzy ideas over reasoned solutions” that make sense in your context.
Here’s how the study went down. The researchers asked AI for advice on seven well-known types of business decisions; for example, should you develop new capabilities or make the most of your existing opportunities (exploration vs. exploitation)? Large language models overwhelmingly (and not surprisingly) all clustered toward the same trendy answers.
Can you guess which one they chose in my example?
The tests - run 15,000 times on all major LLMs - evaluated these questions using different industries, contexts, and prompting methods. Better prompting didn’t change the outcome; in fact, it showed random variation rather than controlled variation. The researchers tried switching the industry and company size of the test, yet the biases remained. Short context statements (windows) resulted in less bias than long ones.
Why the bias?
LLMs predict answers based on sentiment, and “they have internalized all modern management trends and buzzwords.” The researchers note how startup culture and innovation narratives dominate Medium and Substack, skewing the predictions. Take “centralization” for example: consolidating control is a great business strategy in some contexts, but the word evokes Orwellian images and negative, uninspiring sentiment. Nobody (at least on the Internet) likes authority.
Then there’s the “hybrid trap.” When you give LLMs the option to not require a binary answer, they always recommend you take the best of both strategies instead of choosing between them. That might sound wise, but it’s actually a formula for failure called non-essentialism.
Why do all LLMs have these biases? Well, they all use the same training data.
Visual Elevator Music
This gravitation toward the average answer isn’t limited to business decisions; it’s embedded in the very way AI “sees” the world.
In this massive study, researchers set up an AI-to-AI loop: a large language model asks an AI image generator for a graphic, reviews the output, then requests improvements and refinements to the image. The researchers created these prompt “trajectories” (text->image->text->image loops) and watched them run autonomously, on their own.
No matter how diverse their starting prompts, all the agents converged on generic scenes, even generating the same room-temperature-look.
“Despite testing of seven different temperature values across 700 independent trajectories, the systems systematically evolved toward nearly identical semantic and visual endpoints—stormy lighthouses, urban night scenes, gothic cathedrals, and palatial interiors. Rather than exploring creative possibilities, autonomous AI loops appear to gravitate toward what could be called visual elevator music.”
Human artwork also trends toward common, generic content, the study authors point out, but then a real creative mind comes along and introduces something truly unique. In this way, the real artist provides “corrective feedback” that challenges the gravitation toward boring images and content.
What an amazing finding. Is creativity the thing that makes humans unique from machines - and from every other life form?
Seeing the Difference
These studies remind me of that moment in the early 1980s when human cloning was on everyone’s minds.
In 1983, the British government formed the Warnock Commission to study the ethics of new human reproductive technologies. In his address to the committee, theologian Oliver O’Donovan said that, in our technology-focused culture, we often fail to recognize the significance of things we see in nature. We don’t marvel when someone recovers from an illness or injury; instead, we regard the healing as a result of human ingenuity and good medical technology.
“This blindness in the realm of thought is the heart of what it is to be a technological culture,” he said.
O’Donovan could have said the same thing today. Yes, we are a technological culture. Are we losing our ability to tell the difference between original ideas and the average, technological answer?
These studies can help. Both sets of researchers recommend essentially the same things:
Use LLMs to research and expand options, but not to choose between them.
- Use AI to find AI’s biases, as these studies did.
- Don’t lose your edge.
All of this confirms what most people know intuitively: originality still starts with humans.
The HBR study authors don’t hesitate on this question: “Leadership is ultimately about making hard choices in conditions of uncertainty and taking responsibility for them. AI cannot and should not be a substitute.”
To help remind you of these concepts, I’m now adding a hit 80s song to the Frictionless Data Spotify playlist each week. This week, enjoy Mad World by Tears for Fears.
For the full story about making data flow faster and better, check out Frictionless Data on Amazon.
