How to avoid clichéd responses on ChatGPT (or any AI model)

Like a lot of folks in my world of work, I'm a daily user of AI tools (both public ones like Gemini, Claude and ChatGPT, and excellent specialist tools we have access to at AKQA on WPP Open).

As a qual and quant researcher, trained in the Malcolm X School of Research (motto: "insights by any means necessary"), I know not only that an insight can come from anywhere, but that every information source is inherently biased.

Putting these two apparent halves of a conundrum together means I use AI tools, but with the same caution that I use for any single data source (and we can save arguing whether it is a "single" data source another day).

One danger I found is AI tools propensity for cliche (as they reboil the pre-boiled liquid from the internet, concentrating the minerals and impurities found in it).

So I asked ChatGPT to help.

Now, I'm sure the great number of you have already adopted this kind of thinking into your prompting (and have fed this through to other friends and colleagues who are newer to, or less frequent or accomplished users of AI tools), but here's some useful advice about the current self-declared edges of one LLM platform.

Here's the link: https://chatgpt.com/share/67a0d004-62e8-800f-8f73-13ec55442a41

Here's the 100 word ChatGPT summary:

The conversation explores how clichés form and whether ChatGPT is likely to produce them. ChatGPT follows a five-stage model where phrases become clichés through repetition. AI is prone to clichés due to training on existing human discourse but can be prompted toward originality. The discussion shifts to Rumsfeld’s “knowns and unknowns” and Žižek’s “unknown knowns.” ChatGPT struggles with true “unknown unknowns” but can surface implicit biases (“unknown knowns”) through critical prompts. To uncover new insights, interdisciplinary research, paradoxes, and questioning assumptions are suggested. The conversation ends with the idea of experimenting to push AI beyond clichés.

And here's a transcript of the full conversation (for when ChatGPT's developers move the goalposts, or go the way of my beloved Posterous blog...)

AM: How are cliches created?

ChatGPT said:

Clichés are created when a phrase, idea, or expression is used so frequently that it loses its originality and impact. The process usually follows these stages:

1. Fresh and Effective – A new phrase or idea emerges that captures something in a vivid or memorable way. It resonates with people because it feels insightful or unique.

2. Popularity and Repetition – The phrase spreads through literature, media, conversation, and culture. It becomes a go-to way of expressing a certain thought or emotion.

3. Overuse and Predictability – As more people adopt it, the phrase starts to feel less fresh. The emotional or intellectual punch weakens because everyone has heard it before.

4. Becomes a Cliché – Eventually, it turns into a lazy shortcut for communication. Instead of sparking new thought, it signals predictability and lack of originality.

5. Ironic or Revived Use – Some clichés get repurposed with irony or humor, or they evolve into something fresh again with a twist.

Clichés aren't inherently bad—they became overused because they were effective at some point. The trick is knowing when to use them for familiarity and when to avoid them to keep writing or speech engaging.

AM said:

Following your five stage model above, how likely is it that ChatGPT will produce outputs that are cliches?

ChatGPT said:

ChatGPT is quite likely to produce clichés, especially when generating common phrases, advice, or narratives, because it is trained on a vast dataset of human writing, much of which contains well-worn expressions. Let's analyze it using the five-stage model:

1. Fresh and Effective – When AI first started generating human-like text, it felt new and impressive. Early outputs might have included some fresh phrasing simply because AI was a novel tool.

2. Popularity and Repetition – As more people use AI models, patterns emerge in common responses. Frequently requested topics (e.g., motivational advice, storytelling, writing tips) often lead to repeated phrases.

3. Overuse and Predictability – Because ChatGPT aims for clarity and broad appeal, it tends to default to commonly accepted phrasing, which can lead to cliché-like responses, especially in areas like inspirational writing, greetings, or storytelling tropes.

4. Becomes a Cliché – Certain AI-generated phrases or structures might become so overused that they feel mechanical or unoriginal, leading to a sense that AI-generated content is formulaic.

5. Ironic or Revived Use – Users sometimes play with AI clichés by using them in a self-aware or ironic way. AI models can also be fine-tuned or prompted to subvert clichés by deliberately seeking more original phrasing.

How to Avoid AI Clichés

• Use specific, unique prompts (e.g., "Describe a sunset without using the word ‘golden’").

• Encourage fresh perspectives (e.g., "Explain love as if you were an alien encountering it for the first time").

• Request unconventional styles (e.g., "Write like a 1920s newspaper columnist covering a futuristic event").

Would you like me to rewrite something in a less cliché way?

AM said:

Kind of. I'm going to start with a quote: Donald Rumsfeld (in)famously said "There are known knowns. That's the knowns we know we know, then there are known unknowns. That's the knowns we know we don't know. But there are also unknown unknowns. That's the unknowns we don't know we don't know." This quote was critiqued and expanded by Slavoj Žižek: "What Rumsfeld forgot to add was the crucial fourth term: the "unknown knowns" - things we don't know that we know, all the unconscious beliefs and prejudices that determine how we perceive reality and intervene in it." Bearing in mind your five stages, and the likelihood for ChatGPT to produce cliches, is it possible for ChatGPT to output either a) unknown unknowns or b) unknown knowns? If so, how can I prompt ChatGPT to make these outputs more likely? If not, what research methods would you recommend instead?

ChatGPT said:

This is a fascinating question because it gets at the core of how AI processes information and whether it can go beyond the predictable into true novelty or deep introspection. Let’s break it down in terms of Rumsfeld and Žižek’s categories:

Can ChatGPT Generate "Unknown Unknowns"?

"Unknown unknowns" are things we don’t even realize we don’t know.

ChatGPT is unlikely to generate true unknown unknowns because:

• It doesn't discover knowledge in the way that scientific research does.

• It works within the boundaries of its training data and human-posed prompts.

• It doesn’t engage in random speculation without a structured input.

However, it can help surface unexpected insights by combining knowledge in novel ways. If you want to push AI toward the unknown, try:

• Asking for paradoxes or contradictions (e.g., “What’s an assumption in my field that could be completely wrong?”)

• Requesting cross-disciplinary insights (e.g., “How would an AI theorist and a Buddhist monk define intelligence differently?”)

• Using counterfactuals (e.g., “What if gravity worked in reverse—how would biology evolve?”)

Can ChatGPT Generate "Unknown Knowns"?

"Unknown knowns" are unconscious biases, assumptions, or things we intuitively know but don’t recognize.

ChatGPT is somewhat better at surfacing these because:

• It has been trained on vast human discourse, which includes implicit biases.

• It can analyze patterns of thought and expose assumptions.

• It can be prompted to critique its own responses.

To push for unknown knowns:

• Ask it to reveal biases (e.g., “What are hidden assumptions in how I framed this question?”)

• Make it self-reflect (e.g., “What are the blind spots in how AI describes love?”)

• Ask for a psychoanalytic or philosophical critique (e.g., “Analyze this from a Lacanian perspective—what’s the real desire behind this logic?”)

If AI Falls Short, What Research Methods Work Instead?

If AI can’t generate a true unknown unknown, here are human-driven research strategies:

1. Engage in Interdisciplinary Exploration – Read outside your field to expose blind spots.

2. Use Thought Experiments – Science fiction, paradoxes, and counterfactuals are great ways to stretch thinking.

3. Analyze Your Own Assumptions – Journaling or Socratic questioning can help.

4. Expose Yourself to Radical Ideas – Read thinkers who challenge your worldview (e.g., Žižek, Foucault, Deleuze).

5. Seek the Margins – Explore non-dominant perspectives (e.g., indigenous knowledge, underground philosophy, esoteric traditions).