There’s a witch hunt underway for anyone who uses AI to make things, which is overshadowing the more important question of whether the generated content is any good. My unpopular opinion is that using technology to unlock creativity is not the real issue. We’ve always done it. The printing press, the synthesizer, the camera, and Photoshop were each accused of killing craft in their day. The real problem is that everything generated by AI is becoming too similar. The content isn’t bad. It is just too predictable.
Why prediction leads to compression
In 1948, the mathematician Claude Shannon published a paper that established modern information theory. His concern was practical: how could messages be transmitted efficiently and reliably? A central insight was that information could be understood in terms of uncertainty. Predictable events require less information to describe; unexpected ones require more. When a sender and receiver share knowledge of a pattern, that pattern can be encoded economically.
The connection between prediction and compression also helps explain AI models. In a way, language modeling Is compression. During training, a model learns to assign probabilities to the next small piece of text. It is penalized when it assigns too little probability to what actually follows. The training objective, known as cross-entropy loss, has a direct relationship to compression: with an appropriate coding method, a better predictor can describe text using fewer bits on average.
Ted Chiang once compared ChatGPT to a blurry JPEG of the web. JPEG compression is lossy. It keeps the broad shapes and throws away fine detail, the rare and specific features that don’t appear often enough to be worth storing. Unfortunately in writing, that discarded detail is where the brillant metaphor lives, along with the contrarian insight, the joke only you would make, and the idea that sounds wrong until it suddenly sounds obvious. If something can be predicted, it usually isn’t new. This is creative compression.
After pre-training, language models are further tuned on human preferences, and people tend to reward answers that feel familiar, polished, and safe. This can reinforce the problem, rewarding familiar answers until they become the house style. That’s why AI writing has a recognizable accent: the dash-heavy rhythm, ideas grouped in threes, and the “it’s not this, it’s that” framing. None of these are bad on their own. In fact, prior to the LLM era - they were hallmarks of classic, incisive writing. Now, they are marks of shame.
So, how do you avoid repetition and similarity? In the early days of LLMs, people played with temperature, the setting that controls how much randomness goes into each word the model picks. It was, quite literally, an entropy dial. Temperature settings are disappearing or being deprecated in newer frontier large language models because modern platforms increasingly manage response variability dynamically based on the input prompt and system instructions. Even if you can find a way of adjusting temperature, turning it up rarely helps. A heavily tuned model has already narrowed its options, so raising the temperature mostly adds noise rather than real novelty.
For most of us, there is no simple technological solution. We just have to get smarter at how we use AI tools. The danger is that when we ask a model for a compelling opening, a persuasive argument, or a professional-looking image, we often accept the first plausible direction it provides. We may revise the wording without reconsidering the premise. The model’s initial suggestion becomes the boundary of our exploration.
In fact, in an experiment by Anil Doshi and Oliver Hauser, AI-assisted stories were judged more creative and enjoyable, yet became more similar to one another. Everyone improved. The collection became less varied.
The problem and the solution lie with the user. Rick Rubin has made the comparison between modern AI use and Andy Warhol. At the Factory, Warhol often chose the image and the treatment, then directed assistants to screen-print the work. The Marilyns are still Warhols. As Rubin puts it, the art lives in the ideation: the prompt, not who put the paint on the canvas. Using AI is not the problem. Accepting its default taste is. Warhol brought a unique point of view to his work that superseded who actually made it. Too many people outsource their point of view as well as the labor.
A new playbook for corporate rebels
Creative compression is not just an artistic dilemma. We also face the risk of algorithmic monoculture when we pose strategy questions to AI models without a plan for differentiation. Teams feed the same public models the same kinds of prompts, get back the same plausible answers, and ship work that sounds professional and is utterly interchangeable.
A strategy can be impeccably presented and still contain nothing worth doing. You cannot fix this by banning the tools. People will use whatever lets them move faster. If you take away official AI platforms, they will just use personal ones - exposing your data and secrets in the process. A better strategy is to consciously design and deploy smarter patterns that force people to think and bring differentiation back to their work.
Here are five new patterns to think about:
1. Free thinking, enforce structure
Amazon had an infamous ban on PowerPoint presentations, requiring employees to structure their presentations into a six page memo that clearly outlined the decision to be made, what had been done in the past, and the supporting data. Excessive length was a red flag. In a similar way, your team members should be encouraged to use AI to explore ideas wildly, but impose structure on their outputs. Hundred pages of beautifully formatted, AI-generated slides with no real argument is just a waste of everyone’s time. Before generating the polished version, require the author to state the recommendation, identify the evidence they trust, and explain what would change their mind.
2. Create agents to disagree with you, and each other
AI wants to be our friend, and will by default, agree with you. So build a team of disagreeable agents whose job is to contest the working theory. Give them different data, different incentives, and permission to be contrarian. For example, before a strategy review, run the same brief through three agents: a bullish growth case, a bear case built only on disconfirming data, and a “what would our sharpest competitor do?” agent. Present all three side by side.
3. Source data that only you can
Originality comes from inputs nobody else can prompt. Interviews, field notes, win/loss calls, support tickets, and proprietary archives move the model off the public internet’s center of gravity. Instead of asking for “ideas for our next product,” dump anonymized customer interview transcripts and last quarter’s churn reasons into the context, then ask what patterns a competitor would miss.
4. Punish laziness, not tool use
Ban bad work, not AI use. Sloppy, generic, or unowned output gets sent back whether a human or a model wrote the first draft. If your marketing team returns campaign copy sounding like every other LinkedIn post, ask them to reconsider their purpose.
5. Let the chaos monkey loose
Don’t standardize on one chatbot. Build a controlled ecosystem of different models and agents so people can explore divergent ways of working without putting production at risk. Chaos Monkey is an open-source software tool created by Netflix that randomly disabled computer servers in a live production environment to test how well the system survives unexpected failures. We need something similar in our creative workflows to ensure original thinking doesn’t get lost. Occasionally remove your team’s favorite AI model, withhold the standard template, or ask them to work without their most convenient assumption. Find out whether the idea survives without the machinery that produced it.
If you are looking for inspiration on preserving your unique point of view, corporate leaders could learn a few things from artists. In creative communities, while the old guard argues about provenance and ‘what art really is’, there is always a new generation with nothing to lose, that gets on with finding ways to tell its stories.
How to start a creative revolution
Very few artists will be content to use AI systems according to their default settings. Creators need ways to interfere: introduce their own material, alter one element without regenerating everything, preserve an accident, or push a recurring flaw until it becomes an aesthetic. Some will build systems around their archives, obsessions, and personal mythologies. Others will subvert commercial tools.
Simply accepting an AI system’s standard output strikes me as the creative equivalent of pressing the accompaniment button on an old Casio keyboard. Choose a rhythm, adjust the tempo, and off it goes. We should demand more of a technology supposed to expand human imagination. Otherwise, we risk building an entire creative economy around increasingly sophisticated versions of the demo button.
In the early 1980s, Roland released the TB-303, a small silver box meant to give guitarists a bass accompaniment to practice with. It sounded nothing like a real bass. It sold poorly and was discontinued. Used units ended up in second-hand shops at bargain prices. The machine had six rotary controls, including cutoff and resonance. They offered more control than its modest purpose seemed to require. A few years later, musicians would discover what that freedom was worth.
Nathaniel Jones (DJ Pierre) and his group Phuture were looking for a unique sound and bought a cheap unit to experiment with. Unable to read the Japanese instructions, they hooked it up to a drum machine and began playing with the sequencer. One particular pattern caught their attention, prompting Pierre to tweak the control knobs, radically alter the sound, and record the historic jam session.
When DJ Ron Hardy played it at Chicago’s Music Box club, the crowd initially struggled to make sense of it. He played it four times that night. By the fourth, something had clicked. Released in 1987 as Acid Tracks, the recording helped define a genre known as acid house. The musicians pushed the Roland machine until they found something worth making their own. A piece of corporate engineering became the raw material of a counterculture.
AI needs that same creative insubordination: a DIY movement of people building their own instruments, subverting commercial tools, and making things that would never survive a product approval meeting. The tools should be cheap, accessible, and open to being pulled apart. You shouldn’t need a research lab, an engineering degree, or anyone’s permission to make something strange and entirely your own.
AI needs its acid house moment.