Pursuing ‘Joe-Ness’
A cartoonist tests whether AI can expand his drawings without sacrificing his unique style
By Joe Dworetzky • Bay City News
August 2, 2026
IN MARCH OF 2024 I wrote a story for Bay City News Foundation that considered whether I could land a cartoon in The New Yorker if I used generative AI to enhance the limited drawing skills that I brought to my cartooning. The question was rhetorical, but I used it to explore whether I should outsource the work of illustrating my cartoons to AI.
I evaluated the pros and cons and did my best to make the case for using the technology. But despite my genuine desire to level up my cartooning, there were serious problems with the way the AI models had been trained on other people’s work. The images it produced had been copyright-laundered but using them still felt cheap and vaguely sleazy. I concluded that I wouldn’t be using AI generated images in my work, unless there was a clear reason for doing so, like where the technology was part of the story.
My conclusion had a substantial caveat; I said if there was a way to train the AI on my body of work so that all it did was draw things in my style — with all the warts and imperfections — that might change the balance.
In the two years since I wrote that story, several things happened that encouraged me to revisit my decision.
First, AI had gotten much better at generating illustrations. In 2024 I couldn’t work iteratively. The AI would get 95 percent correct, but when I’d try to fix the last 5 percent, the app would start anew; the 5 percent might be fixed, but it would have generated a new drawing that had fresh issues. Not only were the illustrations better, but it was now much easier to use a reference drawing or drawings to shape the AI-generated outcome.
Second, in April-May 2026, I covered the Musk v. Altman trial for Bay City News. My colleague Jay Harris and I provided daily stories from the courtroom as well as weekend wrap-ups and analysis. The courtroom was swarmed by national and international media and in looking for a way that our work would stand out visually from our competitors, we decided we would use cartooning. And, because Sam Altman and Elon Musk could be seen in a very loose way as parents of ChatGPT, we thought it would be appropriate — and appropriately ironic — to use ChatGPT, a kid with parents in a custody fight, to draw the cartoons.
We worked up a set of ChatGPT drawings of the principals and gave each of the main actors an animal personality. The judge a stern owl. Musk and Altman bantam cocks, etc.
We created more than two dozen characters before the trial started. Once it was underway, I could log into ChatGPT, reference the characters and the general scene I wanted, and it would generate an illustration for me to caption and insert dialogue as needed. The process was amazing; with it I could call up cartoons 10 times faster than I could have drawn something, and what it drew was way beyond my skill level.
Art, technology, and tools
After I recuperated from the grind of the trial, I revisited the musing in my 2024 piece about whether I would feel better using AI drawings in my work if the AI was taught to produce images in my style.
I came to cartooning late in life. I loved coming up with the idea and working out the caption and dialogue, but I was constrained by my limited drawing skills. I tried to teach myself, and I made some progress, but I did not have much range. I usually just drew faces and put them in conversation with each other. That was fine for what it was, but there were many cartoons that required more complicated settings and situations that I couldn’t draw.
In the broader creative and intellectual property community there is an ongoing discussion of a middle category between AI-generated and human-created work. The middle ground — call it “AI-assisted” — is where the central creative idea and development is human, but AI is used in its creation.
Artists and creatives are tool users; the use of tools by itself doesn’t render a creative work illegitimate, as long as the artist is the creator of the work. Use of tools has made many cartoonists and illustrators more productive without robbing their work of its creative essence.
We accept today that photography is an art form and a photographer an artist. But for decades, there was debate over whether a photographer deserved that title given that a photo was a “mechanical reproduction,” not a work of imagination.
For example, we accept today that photography is an art form and a photographer an artist, even though the photographer is not a painter, nor the photo a painting. But for decades, there was debate over whether a photographer deserved that title given that a photograph was merely a “mechanical reproduction,” not a work of imagination. Ultimately, the view emerged that control over the process of creating the image (not the mechanical rendering of it) was an artistic process involving decisions about scene, setting, lighting, framing, and composition.
I had already decided that I did not want to use fully AI-generated drawings, but I would love to be able to take the faces that I drew and with the help of AI expand them into different poses and expressions. If I could teach the AI to do that in my style — without improvement or beautification — perhaps the AI would be a legitimate tool to help me extend my range.
This was my creative work, and it was up to me to decide how I wanted to create it. If I could use the technology to expand my work so that it was more powerful, more connective, maybe I should embrace it. And if the technology couldn’t do that, maybe I shouldn’t. But in either case it was all an academic issue until I knew if I could teach ChatGPT to draw like me.
I decided to try.
The experiment
I chose ChatGPT for the test because of the excellence of its work on the Musk-Altman trial. I uploaded one of my cartoons and prompted the platform to build out a set of drawings of the cartoon character in different poses and with different expressions.
I made it clear to ChatGPT that its work should have all the imperfections and mistakes that my drawing had.
It took the app less than 30 seconds to create a “model sheet” with 15 different poses and expressions. It did a good job with coloration and the overall rendition, but when I looked closely, I could see that it had cleaned up some of the miscues and amateurness from my draft.
What it produced was definitely derivative of what I uploaded, but several of the poses were beyond my skill level — ChatGPT drew scenes much better than I did. It wasn’t perfect but all in all it was an encouraging, exciting, start.
Next, I uploaded a drawing I made of Kurt Vonnegut:
ChatGPT drew Vonnegut in six different poses, but they were terrible. They were far too good; well beyond my range. Vonnegut was a known commodity on the internet, and ChatGPT apparently did not limit its knowledge of the subject to my reference drawing. ChatGPT is a pleaser, and it wanted to please me by making my drawing better, more Vonnegut-like, even though that was the opposite of the mission.
I had to figure out how to teach ChatGPT to draw at my middling level, and there needed to be accountability if it failed to remain a bumbler.
The rubric
ChatGPT and I explored the issue, and we ended up creating a rubric that could be used to evaluate its work. The rubric had five tests, each worth 20 points. The first was called “Recognition Fidelity” and asked, “If I showed this drawing to someone who knows the original sketch would they immediately recognize it as the same person?”
To apply the rubric, the arbiter would look at facial structure, the shape of hair, nose, eyes, as well as the character’s posture before awarding 20 points if it were “instantly recognizable, 10 if “probably recognizable” and zero if unidentifiable.
The other categories included “Imperfection Preservation” (did the drawing preserve “the mistakes, awkwardness, asymmetries and eccentricities” of the reference sketch?); Anti-Beautification (did it “resist making the drawing prettier?”) and then a variety of lower points categories that bore into the AI’s fidelity to character, medium, and the naturalness of the drawings.
The rubric closed with a rule: ‘When the reference is weird, the expansion must stay weird.’”
While it was disheartening to see that ChatGPT described my cartooning as weird, unpolished, awkward, eccentric, etc., the rubric seemed solid. I instructed the app that when it prepared the next character sheet it would not only include my reference drawing and the new poses and expressions, but it would also evaluate them under the rubric and post the final score.
I chose a drawing of Eric Johnson, former podcast producer of “Recode Decode,” who spoke to my class in 2020 when I was a journalism student at Stanford.
ChatGPT quickly took my sketch and drew 11 expressions and poses. It rated its work a 94 under the rubric;, however, the evaluation did not identify the scoring by categories and was of little use.
In my next prompt, I added an instruction that the components of the rubric should be identified in the scoring.
I tried again.
I took a drawing of a colorful cartoon character and uploaded it to the platform.
Without my knowledge or approval, ChatGPT converted the 100-point scoring system to a 10-point scale and defined its goal as achieving “Overall Joe-Ness.”
While the 8.9 score suggested that it had come passably close to that achievement, close examination revealed that it had scored itself generously. Also, its unilateral decision to change the rubric on the fly would not allow continuity in any evaluation of the scoring. That would not do.
I tried dozens of the other characters I had drawn over the years and carefully focused on the drawings and the scoring. There were some high-fidelity scores, showing that — if the scores could be credited — the drawings were landing very close to the reference. That’s when I noticed that in some cases, the reference drawing wasn’t actually my drawing; it was ChatGPT’s redrawing of my cartoon. And there were clear signs of beautification of the redrawn reference drawing, which meant that fidelity to the reference was not a reliable measure.
When I called ChatGPT out on the scoring, it agreed there was a problem, though its proposed fix did not include the obvious one — always using my actual drawing as the reference. Instead, it proposed more categories and guidelines for the evaluation, among them “Make ugliness measurable,” “Make the reference the boss not an inspiration,” and “Draw objects as Joe would.”
I had to reign ChatGPT in.
I made a number of changes in the workflow. I changed the name of the output sheet to an “expansion sheet” to embed the point that the goal was to expand or extend the existing character into new poses and expressions, not improve the initial drawing.
I mandated that each expansion sheet would include my actual drawing as the reference, and right next to the reference drawing, ChatGPT would also include its redrawn version of the reference using its cumulative knowledge of my style. That would let me see how it was doing in actually learning how I drew.
Finally, and perhaps most importantly, the evaluation would be split into two forks — the fidelity of its redrawn image to my reference drawing, and then the fidelity of the poses and expressions to the redrawn version. The two separate fidelity scores would be multiplied by each other and divided by 100 to get the final evaluation. This approach would penalize the AI harshly if it didn’t nail the likeness of the redraw to the reference image which I hoped would force the AI to learn my style.
To help ChatGPT draw a faithful version of the reference, I created a zip folder with about 80 drawings I had made over the years, all of faces. The faces varied a lot — men and women, many ages and ethnicities — but I thought that by seeing a lot of my drawings ChatGPT might find a coherent style that it could employ.
The rule changes and different training made an immediate difference:
This sheet was markedly better than its predecessors — it used my actual reference drawing and its redrawing of it was not greatly enhanced, though the shirt and neck were drawn at a higher level than my work. The poses and sketches had reasonable fidelity to the redrawing. One problem, however. ChatGPT had not converted the individual scores into an overall score. If it had, the score would have been an 84, significantly less than either of the individual scores and a misleading evaluation of the work’s overall fidelity.
Claude joins the team
All users of ChatGPT are familiar with its cloying sycophancy and tendency to over-compliment its user. It also has a strong self-congratulatory bias. Neither of those things could be tolerated in an evaluator charged with faithfully applying the rubric. I needed to find a neutral to give me a second opinion.
Of course I turned to Anthropic’s Claude.
Claude was delighted to be consulted and immediately told me that self-scoring was a serious problem. “Self-evaluation,” Claude said, “has three compounding biases: the model grades its intentions rather than its output, it’s trained toward agreeable conclusions, and an unanchored 1-10 rubric compresses toward 8-9.”
To address that issue Claude suggested that I separate the scoring from the generation of the images.
Claude modestly observed that it would be well positioned for the role of scorer: “The honest version: I can’t generate the images, and that limitation is precisely what qualifies me for the other seat. Because I have no authorship stake in any candidate, I’m structurally positioned to be the independent judge your pipeline is missing…”
I decided to give Claude a shot.
First, I asked him if the rubric was “sound and valuable.” It said “Yes, and more than most evaluation instruments I see.” Claude particularly applauded use of the anti-beautification rule (“when the reference is weird, the expansion must stay weird”) and approved multiplying the separate scores rather than averaging.
I uploaded another sheet that had been scored by ChatGPT and asked Claude for a second opinion on the scoring.
Claude found little to complain about in ChatGPT’s application of the fidelity test, saying “the category scoring is honest, and I have very little quarrel with it.” However, it found a serious computational error: ChatGPT had inexplicably and unilaterally converted the two individual 100-point scales to 50 and 40 points respectively, and then incorrectly computed the resulting score. Claude was merciless to ChatGPT for what it perceived as a rookie mistake on the math.
I worked with ChatGPT to fix its arithmetic and moved to another test.
Over the years I have drawn many writers whose work I admire. I gave Claude an expansion sheet made from my drawing of David Foster Wallace.
Claude gave ChatGPT high marks. Claude found the category scoring honest and the arithmetic correct. It found the CES sheet for DFW to be “honest even conservative” at an overall score of 81. High praise from the rival bot.
I decided to extend the expansion sheet concept to buildings. I have not drawn as many buildings over the years as people’s faces, and I had a lot less to put in the training set. But even with that limitation, the “Building expansion sheet” for the White House came out fairly well:
Boom! What seemed to be working for characters could also work for other things! Over the next several sessions, I created expansion sheets for more buildings and for simple scenes and settings: a park bench, a table on a patio, a city sidewalk.
Real-world test
With a scoring process, rubric, and Claude’s watchful second opinion, I was making progress, but so far, the AI had only done the low-stakes work of making and scoring expansion sheets. The real question was whether ChatGPT could do something with those expansion sheets that would be useful in a real-world context.
I decided to upload a story to ChatGPT along with expansion sheets for the characters in the story, then prompt ChatGPT to create new drawings using them that would illustrate the scenes I identified.
Years ago, I had written a fabulist short story published in the literary magazine Fiction Vortex. called “The Girl Who Didn’t Know What To Be.” The story involved a young girl — Metrissa — who is in class when Mrs. Charity — a prim substitute teacher with a chip on her shoulder — arrives to take over. Despite Metrissa’s excitement at the change of pace that a new teacher could provide to her boring class, she quickly got into trouble when she wasn’t able to tell Mrs. Charity what she wanted to be when she grew up. She was forced to sit on a stool facing the corner as the other students continued class.
I wanted to use that story for the experiment because before I started working to train ChatGPT on my style, I asked the AI to illustrate the story on its own (without expansion sheets). It produced these drawings of Metrissa’s first encounter with Mrs. Charity and then Metrissa’s time in the corner:
The drawings might have been technically accomplished — hats off to the AI — but to borrow ChatGPT’s language, they had an utter and unredeemable lack of Joe-Ness.
I hated them.
I would never use them. They were too well drawn and glitzy, and it didn’t help that ChatGPT misread my story and made Mrs. Charity a perky confident teacher not a prim substitute.
For the new try, I uploaded two character expansion sheets, the first one for Mrs. Charity, the second for Metrissa, a girl I had drawn a few years before.
I prompted Chat GPT to draw new illustrations for the story using those sheets. I encouraged it to use the absolute minimum background needed to ground the scene.
When ChatGPT followed the instructions, Metrissa’s first encounter with Mrs. Charity looked like this:
And Metrissa sitting on the stool in the corner, like this:
The illustrations weren’t perfect — ChatGPT was still fighting the anti-beautification rule — but I was pleased. They weren’t glitzy and professional, and they were indisputably my characters. They were in poses I hadn’t drawn, but that had not caused them to shed their essential Joe-Ness.
Cartooning
Satisfied that it was not crazy to think I could illustrate one of my stories with characters drawn by ChatGPT from my reference, I decided to try another real-world example: a cartoon.
Most of my cartooning is done in a single panel with one or two characters. The U.S. Supreme Court had just handed down its decision in the Birthright Citizenship case. A perfect test case. I decided I would have a reporter doing a man-on-the-street interview and receiving a worm’s eye opinion on the high court’s latest decision.
I started with an expansion sheet for a character that I drew a few years ago. The sheet was created early in my experiment, and its scoring (99) was laughably over-inflated. Nevertheless, I liked the character. He would be my reporter for the cartoon.
The second sheet was for a character that ChatGPT called “Odd Lot” for his pained, skeptical expression. Like the reporter, Odd Lot was also scored indulgently.
I had my two characters, but I still needed a setting.
Years ago, I drew the U.S. Supreme Court building. It was not one of my best drawings. I drew it in a hurry on my iPad. I asked ChatGPT to make an expansion sheet from it.
The sheet was not ChatGPT’s best work either; it did way too much cleaning up and it awarded itself an unearned 86, although the score — in my opinion — ought to have been in the 60’s. However, it had the advantage of being quickly available for my test.
When I told ChatGPT to take the characters and building in the expansion sheets and create a scene I could turn into a cartoon by adding dialogue and a caption, it gave me this:
I took that drawing and converted it into a proper cartoon:
Ok, it wasn’t a great cartoon, but it was good for proof of concept. Moreover, in putting the cartoon together, I had been forced to focus on the disclosure I would make to the reader.
There wasn’t a lot of space on the side of the drawing, but I fit: “Original drawings by Dworetzky expanded with AI-assisted tools and reviewed for fidelity.”
That wasn’t perfect, but I also imagined that it would be published with my other cartooning in Bay City Sketchbook and there would also be an Editor’s Note that explained: “The cartoons and illustrations presented here are based on original drawings by Joe Dworetzky expanded with AI-assistance into new poses, expressions, and scenes. The resulting images were selected for use after evaluation under a fidelity rubric designed to assess consistency with the characters and visual style of the original drawings.”
I could improve the wording, but I thought the disclosures, read together, fairly credited the separate roles that ChatGPT and I had played in the creative process.
Takeaways
I had been at it for more than a month, and it was time to assess the takeaways.
Through the experiment I decided that there was a middle place between AI-generated and artist-drawn illustrations. That place — call it AI-assisted illustration — does not involve a complete surrender of the human illustrator’s artistry.
While I didn’t have it completely right yet, I felt some confidence that if I chose to continue, I would be able to make cartoons based on my own drawings and illustrations — expanded by ChatGPT into poses, scenes and settings that I had not drawn — that still had something of that ineffable quality known as Joe-Ness.
In that respect my experiment had been a success, but there still remained a nagging, bothering, question: Was using AI-assistance fundamentally sleazy, employing a shortcut that belied the human nature of my work and undercutting my credibility as a creative?
I posed the question to ChatGPT and Claude: “Is AI-assisted drawing as contemplated by this rubric and workflow (assuming fully disclosed to the reader) 1) unethical; 2) unartistic; 3) uncreative?”
They agreed that it couldn’t be unethical as long as the AI’s role was fully disclosed. ChatGPT thought the rubric was a strong ethical positive because it showed that I wasn’t using AI to create “random replacement art” but to preserve fidelity to my “original characters and visual style.” Claude agreed and suggested that the rubric was explicitly “anti-appropriative.”
Claude added, “’Artistic’ isn’t a property of which muscle made the mark; it’s a property of judgment, intention, and control over the result.” It said, “Look at what the rubric actually encodes: a refusal of beautification, a defense of asymmetry and wobble and patchy wash, a ‘when the reference is weird, the expansion must stay weird’ law. That is an aesthetic position — a fairly sophisticated one, closer to outsider-art and cartooning values than to commercial illustration.”
I heavily discounted Claude and ChatGPT’s views — after all what else would they say — but they were not without weight.
The experiment had definitely moved me closer to embracing AI-assisted cartooning. The excitement of seeing how I could both use and control the technology was intoxicating, and the speed and variety it could add to my drawing was seductive.
During the experiment I talked to dozens of people about the question I had posed to Claude and Chat GPT — friends, colleagues, creatives, lawyers — and found there was no consensus on the issue. Some thought that disclosure of AI’s role was all that was needed; others had deep distrust — and occasional disgust — with the technology and the companies behind it. My son Adam, trained in graphic design and enthusiastic about the technology, warned me that for many of his cohort, anything creative that uses AI is written off in the worst possible way: it is deemed uninteresting.
It was a close question, but after all the time I had spent on the experiment, I decided that I wasn’t ready to use AI assistance in my work. I would still experiment, and I would definitely use AI if another situation came along — as it did with the Musk Altman trial — where the technology was part of the story. But at least for now I would stick with hand-drawn images.
The ultimate deciding factor for me turned out to be a surprise.
I have always been embarrassed by my drawing. I have body dysmorphia for my body of work, at least as far as the illustrations go.
But over the course of the experiment, I looked carefully at hundreds of my drawings and cartoons. While I wish I was a better artist — and having to survey the amateurness of the work made me groan at times — I came away with something I never had before: an appreciation of Joe-Ness.
Yes, I am limited as an artist, but I can take comfort in one uncontestable fact: I am the world’s best at Joe-Ness, and at least for now, that’ll have to do.
The cartoons and illustrations presented here are based on original drawings by Joe Dworetzky expanded with AI-assistance into new poses, expressions, and scenes. The resulting images were selected for use after evaluation under a fidelity rubric designed to assess consistency with the characters and visual style of the original drawings.
