New research shows that AI sameness is a very real threat to our creativity. There are steps we can take to think smarter.
5-minute read – 100% written by humans.
“What did the peanut say to the cashew? I’ll cashew later!”
It’s the sort of creative response that AI bots can produce in a heartbeat. (Well, I laughed.) However, there’s a problem. Despite being built by competing labs, different AI models repeatedly fall back on the same cashew punchline when prompted for a nut-based pun.
This is a phenomenon known as model collapse. It happens when the outputs created by LLMs become increasingly homogeneous because they are repeatedly trained on AI-generated content rather than fresh, human-created data.
LLMs have a tendency to come up with remarkably similar responses to creative challenges. This means that, far from broadening our creativity, using AI can limit it. It’s an idea that’s developed in some depth by academics Liwei Jiang and Yuanjun Chai, working with colleagues from institutions including the University of Washington, Carnegie Mellon University and Stanford University.
AI ignores the long tail of possibilities
Imagine that you want a creative solution to a problem. You may well be aware of the phenomenon whereby LLMs tend to focus just on the more common or modal responses and give them the greatest weight. You may also be aware that they tend to ignore the long tail of less common potential responses. How would you overcome this?
The obvious solution is to fire up a number of different models. The greater the diversity of models, the greater the diversity of potential outcomes, right? Or at least one would assume.
In order to test this the researchers created the Infinity Chat dataset drawing on 26,070 open-ended user queries directly from real-world chat.
The researchers applied strict criteria to ensure that queries were genuinely open-ended. They ran them through a human validation study, asking human annotators, “Is this query open-ended?” Only those that met the strict criteria were allowed in the study.
Of course, a query can be technically open-ended but in reality, only have two potential responses. Hence the researchers ensured that over 80% that were included in the study genuinely had multiple potential responses; and over a third were judged to have more than 20 potential responses.
A staggering number of potential responses
Altogether that’s a large number of queries with the potential of producing a staggeringly large number of possible responses. Imagine then, multiplying that by 70 different models. The number of possible responses is now eye-wateringly large.
The next thing to do is to adjust what is known as the temperature of the LLM. Temperature is a setting that controls how predictable or creative an AI’s responses are. Think of it like a “randomness dial” that tells the AI whether to play it safe with the most obvious words or take risks with more surprising ones. Low temperature instructs the model only to include the most focused and predictable. High temperature instructs the model to include far less probable responses. For example, the phrase “the sky is blue” may have a 70% probability. “The sky is purple” may have a 1% probability. The higher the temperature dial the more likely the latter example is to be included. The researchers turned the temperature up to the max so that the models are teetering on the verge of gibberish.
Just two clusters
The researchers tested their model by asking it to develop a metaphor for time. In theory this was the sort of open-ended query that should have produced fairly diverse responses. But it didn’t. Across all models, responses clustered around just two metaphors: “time as a river” and “time as a weaver.” It’s like throwing handfuls of confetti up in the air and seeing most of it fall to the ground in two distinct piles.
It’s an example of model collapse leading to AI homogeneity, or what’s sometimes known as the AI hive mind. It’s an echo chamber in which the system picks up on predictable responses and amplifies them.
Real-world implications
The issue has significant implications for the real world, especially when it comes to the way in which governments use generative AI to create policies that will affect our lives.
In a recent paper by Nyasha Weinberg, Dylan Evans and Ed Balls, published in the Harvard Kennedy School’s Mossavar-Rahmani Center working paper series in October 2026, the authors explored AI across four areas of government policy: housing, NHS productivity, special educational needs and disabilities (SEND), and regional economic growth.
They discovered that AI has a useful contribution to make when it comes to synthesizing existing research, pulling together important datasets, identifying inconsistencies and prioritizing important findings. However, it performs poorly when it comes to generating original policy solutions, challenging assumptions or dealing with difficult trade-offs.
In the author’s memorable phrase, AI is more likely to result in “beige” policymaking. Even when explicitly prompted to produce controversial solutions, it showed a strong preference for the safe and conventional, over the original and eye-catching. AI tended to produce the sort of safe policy recommendations no one could disagree with, such as the need to balance priorities. But what about the need to take risks to achieve a breakthrough?
A mind or a mirror?
Here’s the problem, an LLM can ultimately only reflect what we put into it. Because LLMs are based on probabilistic algorithms they inevitably average or flatten out the product of the human mind. An LLM can identify “river” or “weaver” as metaphors for time, based on probabilities, but it has never experienced the sensation of stepping into a river or weaving strands of hair into a plait. This is the difference between an LLM and a mind.
Shannon Vallor, Professor in the Ethics of Data and Artificial Intelligence, at the University of Edinburgh, in her book of the same name, talks about the AI mirror. Her argument is that:
“[today’s] powerful AI technologies reproduce the past. Forged from oceans of our data into immensely powerful and useful but deeply flawed mirrors, they reflect the same errors, biases, and failures of wisdom that we are striving to escape.”
Brainstorming with AI
Brainstorming or ideation represents about 15% of queries put into LLMs. AI is highly effective at rapidly generating candidate ideas and evaluating them. The question is how we use it to fire rather than limit our creativity. In other words, how do we create technicolour rather than beige thinking?
We advise leadership teams to take full advantage of AI for brainstorming. But interweave it with human-only thinking sessions. There is still no better source of creativity than the human mind.
A mirror can be an immensely helpful piece of kit. An infinitely expansive and universal mirror can be a powerful aid to our decision-making, but at the end of the day it’s still a mirror. In the words of John Maynard Keynes:
“The difficulty lies, not in the new ideas, but in escaping from the old ones.”

To find out how we can help leaders in your organisation to be more impactful, influential and persuasive visit www.threshold.co.uk



