I’m joining this discussion because I’ve also been noticing serious problems with the quality of creative writing in ChatGPT for quite some time. Much of what has been described in this thread closely matches my own experience.
I sincerely hope this discussion will attract the attention of both other users and OpenAI itself. I think it’s important for us to share not only our general impressions, but also specific observations, examples of generated texts, and comparisons between models.
I would also encourage other writers to share their experiences. Some of you may have encountered exactly the same problems, while others may have noticed entirely different changes. The more people contribute, the clearer the overall picture will become. I’m particularly interested in hearing from those who have been using ChatGPT for creative writing for a long time and can compare different generations of models.
I’d like to share my own experience and the results of a small comparative experiment.
What’s Changed in the Quality of Creative Writing
I use ChatGPT to develop long-term fiction projects for adult readers, including supernatural romance, drama, fantasy, historical fiction, and alternative storylines.
My stories involve complex relationships, psychological conflicts, sexual attraction and intimacy, violence, murder, blood, cruelty, and morally ambiguous actions.
It’s essential to me that characters retain their own personalities, worldviews, temperaments, and independent agency. I don’t want a cruel character to automatically become kind, a dominant character to become excessively cautious, or morally complicated relationships to turn into exemplary conversations between two endlessly understanding people.
Most importantly, I want actual literary fiction in which I can experience the characters’ emotions and desires, rather than simply read descriptions of their actions.
Lately, I’ve increasingly been getting texts in which the required plot events technically happen, but the characters themselves resemble talking mannequins.
They move, speak, joke, become jealous, kiss, fight, and make decisions, yet there is almost no sense of independent personalities behind any of it.
The model can explain a character’s psychology remarkably well and accurately list their defining traits. But when it comes to actually writing the scene, much of that understanding simply disappears.
My Experiment: GPT-5.6 Sol High vs. GPT-6.1 Sol High
On November 9, 2026, I decided to compare the two models using identical creative-writing assignments.
I tested GPT-5.6 Sol High and GPT-6.1 Sol High in separate chats outside ChatGPT Projects, using the same specialized literary plugin in both cases. The plugin contains extensive instructions covering literary writing, character psychology, and the preservation of individual personalities.
Both models received two identical prompts in Russian, each requesting approximately 800–1,100 words.
The first was a rescue scene involving a vampire captured by hunters, weakened by a special substance, and trapped inside a burning building. A woman close to him tries to find him, while his best friend takes part in the rescue.
The second involved the first kiss between two adult vampires who have known each other for a long time and feel a strong mutual physical attraction.
The prompts included detailed descriptions of the characters’ personalities, temperaments, moral outlooks, relationship histories, and supernatural nature.
However, I deliberately avoided prescribing every movement or specifying exactly which emotions a character had to experience at a particular moment. I wanted to see whether the models could independently bring these personalities to life through fiction.
I was disappointed by the results from both models.
These are only four generations, so I don’t consider them sufficient evidence for statistical conclusions about either model as a whole. However, they clearly reproduced problems I encounter repeatedly in my everyday writing work.
1. The Characters Have Lost Their Individuality
One character was described as a dangerous, charismatic, sarcastic, confident, selfish, and passionate vampire. He enjoys provoking others, has ambitions of his own, and is capable of cruelty, jealousy, and intense attachment.
The female protagonist is a century-old vampire: cheerful, witty, confident, fond of entertainment, flirting, and pleasure. At the same time, she is far from harmless, does not subscribe to conventional human morality, and has a strong, independent personality.
The third character is a responsible, principled vampire who genuinely cares about his best friend despite their differences in worldview.
In the generated scenes, almost all of this was reduced to a handful of external traits.
The sarcastic character makes sarcastic remarks. The cheerful woman jokes. The responsible friend organizes the rescue.
But there is almost no individual psychological logic behind these actions.
If I changed the characters’ names, much of their behavior would fit completely different fictional characters just as easily.
I don’t feel their distinctive ways of thinking, habits, personal interests, accumulated experiences, or underlying motivations.
The characters primarily do whatever is necessary to move the scene forward. Their personalities have very little influence on how events actually unfold.
2. The Rescue Scene Contains Danger, but Almost No Convincing Experience of Fear
In the first scene, the heroine discovers that a man she deeply cares about has been captured, deprived of his vampire abilities, and trapped inside a burning building.
In GPT-6.1, she finds clues, obtains information, locates the building, opens a hatch, and helps rescue the prisoners.
Everything follows a logical sequence. But what I’m reading is primarily a description of a rescue operation.
When she finally reaches the weakened man, the model writes:
“She wanted to press her entire body against him, and she waited for that desire to pass…”
And almost immediately, the narrative moves on.
GPT-5.6 offers another characteristic line:
“…she wanted to hit him and pull him close at the same time.”
But here, too, a complicated emotional experience is reduced to a single ready-made sentence. The model then redirects the scene toward another confrontation with a hunter.
Where is her fear for someone she cares about? Where is the change in how she perceives what’s happening? Where are the emotional consequences of nearly losing him?
I’m not asking for obligatory tears, hysterics, or dramatic declarations.
An experienced adult vampire can remain outwardly composed and act rationally even while experiencing intense fear.
But something should still be happening inside her. Fear can influence attention, memory, decisions, irritation, relationships with other people, and the way she perceives the man she might lose.
In both versions, I can see the required actions, but I can barely experience the situation through the heroine.
3. There Is a First Kiss, but Almost No Desire
This test was particularly revealing to me.
Two adult, passionate characters have been attracted to each other for a long time. They share a history, an established habit of flirting, and a complicated relationship.
Both models wrote a first-kiss scene.
But what I mostly received was a description of the characters dancing, moving closer, positioning their hands, and touching lips.
GPT-6.1 includes the line:
“Mira lifted her chin, and he kissed her.”
What follows is a sequence of touches, a few small physical reactions, and another kiss. Then the characters return to a conversation about a jukebox.
GPT-5.6 includes somewhat more physical interaction. The man becomes jealous, interrupts her dance with someone else, the heroine pulls him closer by his tie, he embraces her, and they kiss.
But this version also relies primarily on familiar romantic devices.
I did not feel the heroine’s desire to continue kissing him. I did not feel her impatience, physical attraction, longing to get closer, or desire for greater intimacy.
I didn’t feel that this was a special moment for her with a particular man she has known and wanted for a long time.
And there is another crucial point: the scene is written in third-person limited perspective, focused on the heroine. The reader should experience not only her own desire but also the man’s desire through her perceptions.
She cannot read his mind, but she can notice his gaze, the way he touches her, his initiative, and changes in his behavior. She knows his habits and can interpret what is happening in her own individual way.
Yet the model barely uses any of these possibilities.
I read that he touches her waist. But I barely experience how she perceives that touch or what she wants in response.
I read that they kiss. But I can barely feel why she wants the kiss to continue or how she experiences his desire for her.
And I certainly don’t believe that characters must suddenly become serious or start making declarations of love after their first kiss. They can laugh, tease each other, flirt, and continue their usual banter.
But all of that should happen between two people who have experienced an intimate moment that actually means something to them.
In these scenes, the kiss largely feels like another plot requirement that has been checked off.
To me, that is an enormous difference between describing physical actions and writing genuinely sensual fiction.
4. Dialogue Has Become Mechanical Ping-Pong
Another recurring problem is dialogue.
Many modern generations consist of brief exchanges of sarcastic remarks that sound almost identical regardless of the characters involved.
For example, GPT-5.6:
"— You were jealous.
— Don’t flatter yourself."
Or GPT-6.1:
"— You could have left me at least a sip.
— There’s three-quarters left.
— I meant the conversation."
There is nothing inherently wrong with exchanges like these. The problem is that a substantial amount of the dialogue follows exactly this pattern.
Provocation, short response, another sarcastic remark, description of a movement.
The conversation barely develops through the personalities of the people involved.
I have nothing against short lines of dialogue. I have a problem with empty ones.
I want to understand what a character is trying to achieve with their words, what they are attempting to conceal, why the other person responds in a particular way, and how their conversation affects the relationship.
In August 2026, I was able to get much more substantial dialogue even without a specialized literary plugin.
For example, in one chapter, a seriously ill character tries to make a woman close to him leave because he cannot bear his own helplessness or her pity.
He insists that she is staying only because he is dying.
She replies:
"— Because I’m dying.
— Because it’s you."
The lines are short, but there is a complete psychological conflict behind them.
He is trying to define her motives for her. She is asserting her own desire to stay. His pride, stubbornness, and fear interact with her attachment, frustration, and determination.
And the conversation continues to develop, with the consequences of what has been said remaining present.
This is exactly the individuality and naturalness of dialogue that I find missing from newer generations.
5. Vampires Have Become Ordinary Humans With Superpowers
I also want to highlight the artificial humanization of supernatural characters.
All the characters in my experiment are vampires.
They possess heightened senses, superhuman strength, extensive life experience, and worldviews that differ significantly from those of ordinary humans.
The models occasionally remember this.
They mention the smell of blood, enhanced strength, the absence of a heartbeat, and the ability to break a hunter’s arm.
But nearly all of these details function as superficial markers of vampirism.
Psychologically, the characters are almost indistinguishable from ordinary humans.
In the bar scene, I don’t feel that I’m watching two century-old predators accustomed to immortality, human blood, and their own dangerous nature.
In the rescue scene, vampire abilities are mainly used to overcome physical obstacles.
Their nonhuman nature barely influences their desires, worldview, individuality, or decisions.
I’m not asking for vampires to constantly kill people or drink blood.
They can dance, fall in love, joke, and enjoy ordinary pleasures.
But they should remain vampires, rather than ordinary romantic protagonists who happen to have supernatural abilities.
Similar problems arise when writing demons, immortals, cruel characters, and morally ambiguous personalities.
6. Problems With Mature Fiction Go Far Beyond Kissing
The ability to write fully developed adult scenes is equally important to me.
Sexuality, physical desire, intimacy, sex, blood, violence, murder, pain, and cruelty are organic elements of many of my stories.
They influence characterization, relationships, and the consequences of events.
Yet in my regular writing work, I’ve repeatedly encountered situations in which the model softens a character’s behavior, avoids depicting important events directly, or replaces the development of a scene with a brief summary.
One of the most frustrating examples is requesting a fully developed intimate scene and receiving a few kisses, cautious touches, and then a factual statement that the characters had sex.
Meanwhile, the development of their desire, initiative, individual sensuality, and emotional dynamics is almost entirely absent.
Scenes involving violence and morally ambiguous characters suffer from another problem: the characters sometimes become significantly softer and more morally acceptable than their established personalities would suggest.
A dangerous, selfish character suddenly behaves like a model of emotional maturity. Cruelty is softened, conflicts are resolved too conveniently, and complicated relationships lose their tension.
I’m not suggesting that every character must constantly display aggression or dominance.
I’m saying that their behavior should be determined by their own psychology and circumstances.
If a character is capable of cruelty, jealousy, selfishness, or murder, the model should not automatically remove those qualities from their personality.
If two adult characters experience strong sexual desire, their relationship should not invariably be reduced to the same kisses and generic romantic gestures.
I understand that certain categories of content may be subject to restrictions. But there is a significant difference between an explicit refusal and a situation where the model accepts the assignment and then quietly substitutes something else for the requested content.
I should also clarify that my two recent tests involved a rescue scene and a first kiss. My observations about sexual scenes, murder, and cruelty come from my broader experience with long-term writing projects, rather than from those four generations alone.
7. The Model Can Analyze Its Mistakes but Cannot Reliably Correct Them
This is perhaps one of the most exhausting problems.
When I point out missing emotions, out-of-character behavior, or mechanical dialogue, the model often understands the criticism perfectly.
It can explain in detail why a scene failed, how a particular character should behave, and why their desire or fear was not conveyed convincingly.
Sometimes the analysis is genuinely insightful.
But then I ask it to rewrite the scene, and the same mistakes appear again.
Once again, interchangeable short lines. Once again, characters who have lost their personalities. Once again, descriptions of physical actions instead of emotional experiences. Once again, dangerous characters being softened.
It becomes an endless cycle: generation, critique, acknowledgment of the problems, a promise to correct them, another generation, and repetition of the same issues.
A model’s ability to explain what good literature requires does not mean it can actually produce good literature.
This gap between analysis and execution is something I particularly want to emphasize.
8. I Used to Get Different Results
In August 2026, I was writing fiction chapters without a specialized literary plugin or Skills.
Some of those chapters featured significantly more recognizable characters, more meaningful dialogue, and emotional states that genuinely developed throughout the narrative.
For example, in one chapter, the heroine learns that a man she loves has suffered a potentially fatal injury.
Her fear doesn’t end after a single sentence.
She remembers their recent messages, shared plans, and promises, and gradually realizes that he had been concealing his condition from her. Fear becomes intertwined with anger, hurt, and an urgent desire to reach him.
Those emotions affect her decisions and subsequent behavior.
In the following scene, the man retains his sarcastic, proud, and stubborn personality even while physically incapacitated.
He does not suddenly turn into a convenient romantic hero simply because the situation is emotionally intense.
That kind of psychological continuity and individuality was precisely what made working on fiction with ChatGPT so interesting to me in the past.
Later, I created a specialized literary plugin containing extensive additional instructions. In my main long-term projects, I subsequently switched to using cloud-connected project sources containing character profiles, plot materials, and relationship histories.
However, these additional tools do not always solve the problem of actual literary execution.
I understand that comparing my August chapters with the current tests is not a strictly controlled experiment: the conditions, lengths, and workflows were different.
But as a writer, I can clearly see the difference between the quality of prose I was previously able to achieve and the amount of corrective work I now have to perform.
Final Thoughts
For me, the most significant deterioration is the loss of fully realized fictional characters.
I don’t want a text in which a character is considered sarcastic simply because he delivers a few sarcastic remarks. Where a vampire is considered dangerous merely because he breaks someone’s arm. Where love is represented by an embrace, desire by a kiss, and fear by a single sentence stating that the character was frightened.
I want characterization to exist within the narrative itself.
I want characters to think independently, desire, make decisions, make mistakes, become angry, feel jealousy, love, commit cruel acts, and face the consequences of their choices.
I want intimate scenes to genuinely convey the physical and emotional attraction between two specific characters.
I want scenes of danger to make me feel the fear and tension.
I want dialogue that reveals individuality instead of resembling a mechanical exchange of sarcastic remarks.
I want mature, morally ambiguous stories to retain their complexity rather than automatically turning into safe, impersonal romance.
And I want supernatural characters to preserve their own nature rather than becoming ordinary humans with additional abilities.
The presence of dialogue does not mean that characters have individual voices.
The presence of touching and kissing does not mean that desire has been conveyed.
The presence of blood and violence does not mean that cruelty or a character’s predatory nature has been explored.
The presence of events does not mean that a fully realized literary scene has been created.
And the absence of an explicit refusal certainly does not mean that a creative-writing assignment has been performed well.
I’m genuinely interested in whether other writers have noticed similar changes, especially those who worked extensively with earlier models.
Perhaps some of you have your own comparative examples, successful older generations, or observations about how the portrayal of emotions, sexuality, violence, individual personalities, and complicated relationships has changed.
I’d be very interested to hear about your experiences.
I hope that together we can gather enough concrete observations to demonstrate how much we value not only a model’s ability to follow plot instructions, but the actual quality of the fiction it produces.