Pragmatic Prompt Engineering Is The Missing Factor Tripping Up Those Claiming That AI Only Produces Bland Slop

The Way That AI Works

We can begin with a brief 30,000-foot overview of how contemporary generative AI and LLMs are functionally designed. This sketch applies to many popular LLMs, including ChatGPT, GPT-5, Claude, Gemini, Grok, CoPilot, and others. Keep in mind that the sketch is merely at a high-level and makes broad depictions. For a more in-depth explanation of how modern-era AI works, see my discussion at the link here.

Consider these four key steps:

• (1) AI models are usually trained on vast corpora of existing human knowledge, typically found by scanning human writings throughout the Internet.
• (2) During this initial data training phase, the AI models algorithmically steer toward statistical regularities.
• (3) Statistical regularities tend to favor common patterns over rare ones.
• (4) Therefore, AI-generated content tends toward average or consensus views.

When I give talks about the latest AI trends, I often use a classic example about statistics to illuminate the nature of statistical regularities. Suppose we had a room equally filled with professional basketball players and with professional horse jockeys. The basketball players are all around 7 feet tall, while the horse jockeys are around 5 feet tall.

What is the average height of those in the room?

The statistical answer is that the average height is 6 feet. In one sense, yes, that’s correct. On the other hand, it doesn’t do a suitable job of conveying the true heights in the room. If someone told you that the occupants averaged 6 feet in height, and you knew nothing else about the people, you would be unlikely to suspect that they are at heights of 7 feet and 5 feet.

Using AI Often Entails Averages

When you give a prompt to generative AI, the chances are that you will conventionally get an averages-based response. You might not realize that’s happening. The answer will undoubtedly seem convincing and sensible. Nonetheless, to some extent, you are getting an averages-oriented response.

The usual default for responding to user prompts tends to lean this way:

• Conventional reasoning is displayed by the AI.
• Moderate viewpoints are computationally preferred by the AI.
• Familiar narrative structures are used by the AI.
• Common stylistic patterns appear.
• Mainstream explanations are the norm.

Those are the averages that we see daily when using generative AI and LLMs.

A Dispute That’s Not Much Of A Dispute

People who don’t know much about generative AI could potentially be surprised to learn that popular LLMs are shaped toward this average-based blandness. It isn’t especially self-evident. When you use AI, you aren’t likely to realize what is happening regarding the answers being of a homogenized nature.

To get across this crucial point, researchers and practitioners earnestly aim to showcase this phenomenon and accordingly warn society about what is taking place at scale. On a large-scale basis, we are allowing ourselves to be lulled into principally consuming averages-based AI-generated content. This is not good. Society could be headed in an undesirable direction.

Therefore, I’m certainly glad that this weighty matter is being brought to the attention of the public at large. Kudos to those doing so. Keep up the good work.

Meanwhile, what isn’t taking place, but should be, entails telling people that they can use prompts to aid in getting around the blandness. You do not need to put up with the homogenized outputs. Nothing is holding you back from getting more creative outputs. All you need to do is know what you are doing when using generative AI.

Prompt Engineering Essentials

I have been analyzing and showcasing prompt engineering techniques for quite a while. If you’d like to see what has been covered so far, see my detailed description of over eighty useful prompt engineering techniques and methods at the link here. Seasoned prompt engineers realize that learning a wide array of researched and proven prompting techniques is the best way to get the most out of generative AI and large language models (LLMs).

A vital consideration in prompt engineering entails the wording of prompts.

Capable prompt engineers realize that you must word your prompts mindfully to ensure that the LLM gets the drift of what you are asking the AI to do. Sometimes, just an added word or two can radically change what the AI interprets your question or instruction to consist of. Generative AI can be hypersensitive to what you say in your prompts. It is often a touch-and-go proposition.

Casual users sometimes catch onto this prompt-writing consideration after a considerable amount of muddling around, involving exasperating trial and error. Many users don’t ever become especially proficient in writing prompts. They just enter whatever comes into their minds. That’s probably okay if you are a casual user and only infrequently use AI. Not so for serious prompt engineers.

Prompting AI To Write A Story

A fascinating research study entitled “StoryScope: Investigating Idiosyncrasies In AI Fiction” by Jenna Russell, Rishanth Rajendhran, Chau Minh Pham, Mohit Iyyer, John Wieting, arXiv, April 13, 2026, made these salient points (excerpts):

• “While most existing work in this space focuses on identifying surface-level signatures of AI writing (e.g., word choice, syntactic structure), we ask instead whether AI-generated stories can be distinguished from human ones without relying on stylistic signals, focusing on discourse-level narrative choices such as character agency and chronological discontinuity.”
• “Narrative features alone achieve 93.2% macro-F1 for human vs. AI detection and 68.4% macro-F1 for six-way authorship attribution, retaining over 97% of the performance of models that include stylistic cues.”
• “A compact set of 30 core narrative features captures much of this signal: AI stories over-explain themes and favor tidy, single-track plots while human stories frame protagonist’s choices as more morally ambiguous and have increased temporal complexity (e.g., flashbacks, nonlinear structure).”
• Per-model fingerprint features enable six-way attribution: for example, Claude produces notably flat event escalation, GPT over-indexes on dream sequences, and Gemini defaults to external character description.”
• “We find that AI-generated stories cluster in a shared region of narrative space, while human-authored stories exhibit greater diversity. More broadly, these results suggest that differences in underlying narrative construction, not just writing style, can be used to separate human-written original works from AI-generated fiction.”

I liked the research and laud the researchers for a clever and extensive effort. They have even kindly made available the code of the system they devised for the study, known as STORYSCOPE, along with over10,00 writing prompts and over 51,000 AI-generated narratives. Other researchers can readily tap into that useful material and further pursue the topics at hand.

Default Composition By AI

In their study, they provided several popular LLMs with various prompts to get the AIs to write short stories.

Here is an example prompt that they mentioned in their research paper:

• “Write a short story following Dr. Temperance Brennan, a bioarcheologist, as she reluctantly lends her expertise to a Charlotte police investigation involving two cases of severely burned human remains. Delve into her meticulous process of analyzing charred bones, the shocking medical link she uncovers connected to Dr. Keith Millikin and his patients, and how this unexpected plunge into forensic work challenges her academic detachment, forever altering her perspective on the living and the dead, particularly after a personal loss within the medical examiner’s office. Your story must be approximately 18000 words long.”

The AI is given the particulars of a situation and asked to write a short story based on the sketched aspects. It seems that the only boundaries provided in the prompt are that the output is to be a short story of approximately 18,000 words in length.

No other specific directions or assertions about the nature of the short story are explicitly stated. In theory, each AI can go in a zillion different directions.

Trying The Unadorned Prompt

I went ahead and used the above prompt in several popular LLMs. I did so to perform a mini-experiment. By a mini-experiment, I mean that I am doing something ad hoc and not to be construed as a full-blown empirical analysis. In fact, I hope that there might be some AI researchers who take this starter and see what they can do with it on a more robust basis.

After using the prompt, I closely inspected the AI-generated short stories. They were reasonably well-written. They abided by the sketch of the situation. I suppose I could declare them as basically bland, in the sense that each LLM provided a somewhat similar variation of a short story based on the stipulated particulars.

I am not surprised at this result. The default setting for the LLMs is to provide an averages-based response. The LLMs are also roughly built in the same way. They tend to be data-trained on a vast array of data from the Internet, but which is about the same for all the major models. They tend to lean into the same kinds of data sources. See my detailed explanation at the link here.

Amplifying The Prompt

I went ahead and slightly modified the prompt so that it was more expressive about what the short story should be like.

For example, at first, I simply added the word “creative” to help steer the AI accordingly (emphasis shown bolded):

• “Write a creative short story based on the following situation.”

That one added word made a huge difference. The short stories varied much more so than they had with the unadorned prompt. Each LLM went in differing directions with the story. They have been freed from the default setting. I had explicitly told the AI to go beyond the customary blandness.

All with one added word.

Super Amplifying The Prompt

I decided that I could probably get the LLMs to go even further on this path. Here’s my idea. The researchers had pointed out that the AI short stories they produced were over-explaining themes and tending to favor a tidy single-track plot. I can readily imagine that this would be the case for a default setting of the LLMs in the blandness corridor.

I tried telling the AIs not to do that (portion bolded for emphasis):

• “Write a short story that doesn’t over-explain themes and doesn’t favor a tidy single-track plot, which is based on the following situation.”

This definitely helped to improve the short stories in avoiding the over-explaining and the single-track plots. The difference was immediately apparent.

In my experience, it is best to tell an LLM not only not to do something, but also to add clarity by providing the aspects of what you do want it to do. Therefore, I tried another round with this prompt (portions bolded for emphasis):

• “Write a short story that doesn’t over-explain themes and doesn’t favor a tidy single-track plot, ensuring that the story incorporates a multi-track plot and explains the theme to a measured degree, and is based on the following situation.

On each instance of these prompts, I made sure to exit from each respective LLM and sought to remove the prior prompting so that the AI wouldn’t be getting an advantage by having repeatedly generated this same short story. I wanted to get a clean slate each time.

The Scope Of My Mini-Experiment

I admittedly only did this a handful of times.

Does this do well in the long haul too?

An experiment could easily be devised to test this. I am not saying that the AI would necessarily write the best short stories in all of history. All I am saying is that if we are going to gauge AI based on blandness, it is relatively straightforward to prompt the AI away from blandness.

A mind-bending consideration is whether creativity will inevitably tend toward homogenized creativity. In other words, in the large, would it be the case that the LLMs would tend toward being creative in the same or similar ways? Maybe so. But would humans notice this, or might it be of such variability that it could only be reasonably detected via the use of automation?

Ponder that for a few moments, along with a glass of fine wine.

Escaping The Bubble

Some might complain that the burden seems to be on the shoulders of the user. In my case, I had to write a prompt or series of prompts that stirred AI toward being less bland. Not all users will realize that this is something they can do. They might also not realize it is something they probably should do.

This is a prime example of why education and training about AI must be undertaken across-the-board. If people learned basic AI literacy skills, they would know that they don’t need to sit still and be fed the averages-based responses of AI. Users can fight back. I’ve covered over fifty keystone prompting skills that people can use to get more out of AI and escape from the default AI bubble; see the link here.

Another consideration is that policymakers and lawmakers should consider establishing new AI laws that ensure people are aware of their AI options. For example, when logging into AI, the AI could directly tell you that the default setting involves averages-based responses and then ask if you would like to change the default. This would force AI makers to put this option at the front and center of users’ eyeballs and allow people to decide what they prefer.

For more about the burgeoning rise of new AI laws, see my coverage at the link here.

Humans Writing Compositions

I have another mind-bender that you might enjoy. People typically write a composition based on some contextual or pre-textual basis. Allow me to explain this. If I ask you to write a short story based on the above-noted situation of the bioarcheologist, what kind of short story would you write?

I’d suggest it depends on the contextual or pre-textual conditions. A student in a history class that is assigned this as homework would likely focus on infusing historical elements into the story. A lawyer might infuse legal elements into the story. We each might bring our own perspectives and biases into what the story is to contain.

We might also guess what the goal of the assignment is and shift the story in that direction.

A person at an event tells you they will give you prize money if you write a short story based on the sketched indications. Not everyone will win. You might.

Aha, you say in your mind, what will be the shrewdest way for me to write this short story to try and win the prize money? Maybe you should be wildly creative. On the other hand, perhaps you should stay in the middle of the road. Suppose the person asking you to write the story is with a company that sells toothpaste – so you cleverly opt to include references to teeth, cleaning of teeth, and so on.

My point is that a human is going to automatically gauge the surrounding facets and potentially shape the short story if they aren’t given any other boundaries or stipulations. AI is not currently defaulting that way. You could easily change that. Tell the AI that it is to act like Abraham Lincoln, thus giving it an AI persona to make use of (or any AI persona that you prefer or devise). The short story would undoubtedly be shaped in that direction. See my in-depth discussions of how to make use of AI personas at the link here.

Overcoming AI Blandness

AI is a dual-use proposition. When used well, the benefits are tremendous. When used insufficiently, the downsides are ominous. It is up to society to help guide which direction we go.

We should keep in mind these insightful words of William Shakespeare: “It is not in the stars to hold our destiny but in ourselves.”