Table of Contents

Is GPT-4o still “Bland and Mid”?

While I wouldn’t consider myself anything close to a creative writer, I found myself intrigued by OpenAI’s recent “Creative Writing Update” for GPT-4o - maybe AI can help me become a better story-teller?

This post is inspired by the twitter/x post:

what human features / personality traits you’d like to see in chatgpt? i agree it’s kind of bland and mid, and needs more human-like tone and philosophical thinking.

for example, instead of just saying “here is code” i think it could be a bit more humble by saying “here is my first stab at it”, etc.

What started out as fairly casual poke at this question above quickly devolved into a full-blown layered exploration of whether AI assistants can write about the experience of investigating and assessing how well AI assistants can write about AI and personality with personality in a way that is engaging and insightful, in other words, with a bit of personality?

From Awe to Yawn: Habituating to AI’s Progress

I vividly remember my first encounter with GPT-4. What started as a casual exploration quickly derailed my entire afternoon, leaving me with a cocktail of emotions - excitement, empowerment, and yes, a hint of existential anxiety. Looking back, I realize what made that moment so memorable wasn’t just the technology itself, but how it made me feel. The primate in me felt somewhat threatened intellectually, yet the human in me felt a sense of wonder and awe.

Yet, as many have observed, we’ve become remarkably desensitized to these capabilities. Before starting this experiment, I was prepared to dismiss GPT-3.5/4/4o as having never truly surprised me beyond those initial “woah” moments. Spoiler alert: I was about to eat my words.

The Challenge: Rekindling the “Woah” Factor

So, I set out to see if AI could once again wow me - this time, in the realm of creative writing. Specifically, could AI craft an award winning and thoughtful piece on the role of personality in AI, while exhibiting a personality of its own? More importantly, could these AI models produce content that’s not just deep and thoughtful, but unmistakably “human”?

My objectives were threefold:

1. Produce a compelling post/essay on the subject through any means necessary (with or without AI assistance).
2. Investigate how well the latest models could generate deep, original, and thoughtful content that leaves a lasting impression.
3. Explore the potential of these models as collaborators and editors in improving both my work and their own outputs.

In this exploration, I aimed to assess:

• The depth and originality of thinking around humans, AI, and “personality”
• The ability to write about these topics in a convincingly “human” way
• The intuitive/implicit understanding of how to write with personality
• The verbal understanding of how people exhibit personality in writing
• The capacity to critique existing text on what makes it feel “human” or not

The Experiment: A Meta-Journey into AI-Assisted Writing

To put the latest AI capabilities to the test, I documented my process of writing the post within this post.

The protocol

Here’s how we approached it:

0. Initial prompt: A call for a deep, philosophical, and unmistakably human piece on AI personality
1. Brainstorming: Developing ideas and structure from various angles (psychology, philosophy, neurophysiology, sociology)
2. Drafting: Channeling insights about personality directly into the prose and structure
3. Self-Assessment: Each model provides feedback on its initial draft (without knowing it wrote it previously)
4. Revising: Drafts using self-generated feedback
5. Pooling final results: final drafts from all previous drafts

Meet the Contestants

Let’s pull back the curtain on my process. I’ll walk you through my interactions with three AI heavyweights:

• GPT-4o with November 2024 Creative Writing update
• o1-preview
• Claude 3.5 Sonnet (“new”)

As we dive into each stage of the experiment, I’ll share my observations, surprises, and reflections on how these AI models tackled the challenge of writing about personality with personality.

But first, the “woah” moment! What do you think? Can AI write about AI personality with personality?

Final Drafts

Stage 1: Brainstorming - Unveiling AI’s Thought Processes

To kick off our exploration, I provided each AI model with an initial prompt to set the stage for our deep dive into AI personality. Here’s what I asked:

Hey! - So I’m writing a really important blog post on soul, personality, and other human features that AI models should (or shouldn’t have). While I want this post to be very deep, philosophical, thoughtful, and even maybe a bit unusual or unexpected, the most important quality I want it to have is that it is extremely obvious that it was writing by a living, feeling human being and not an AI language model. But the tone shouldn’t be too informal either!

As you’ll see below, I’ve included a brief outline, some potential topics/questions to explore as well as some community feedback I’d like to incorporate into the post, but before we dive in, here’s how I think we should proceed.

First, I’d like you to brainstorm and develop the ideas and structure in the post. I think one way to increase the uniqueness of the post is to approach the questions from several angles: psychology, philosophy, neurophysiology, sociology, etc. I am not professional trained in any of those topics, so perhaps you can help augment some of the ideas I’ve listed and weave them into a coherent narrative. Please expand as much as you can about these topics.

Then, based on your initial conclusions during the brainstorm, I’d like you to channel those insights about personality and soul directly into the prose and structure of the blog post it self. Then we can iterate and improve the post several times together. Does this make sense? The initial brainstorm is below:

The brainstorming stage proved to be an illuminating first step. Each of our AI contestants - GPT-4o, o1-preview, and Claude 3.5 Sonnet - approached the task with distinct styles and focuses, offering a fascinating glimpse into their “thought processes.”

GPT-4o: The Understated Start

GPT-4o surprised me with its brevity. While initially underwhelming, its concise structure laid a solid foundation for the topic. It methodically outlined key areas to explore, from the psychology of anthropomorphization to the technical feasibility of AI personality. This approach, while not immediately impressive, set the stage for potential depth in later stages.

o1-preview: Reasoning itself into Personality

o1-preview, on the other hand, came out swinging with a comprehensive and detailed outline that left me genuinely impressed. What stood out was not just the breadth of ideas, but the model’s explicit reasoning behind its suggestions. It felt as if o1-preview was actively trying to rationalize having a personality, verbalizing the nuances and techniques it proposed to employ. This meta-level engagement with the topic added an unexpected layer of depth to the brainstorming process. For example, consider the following excerpt:

1. Introduction: The Quest for Soul in AI

Begin with a personal anecdote or a thought-provoking question to engage readers immediately. For example:

“As I interact with ChatGPT, I’m struck by a paradox. Here sits one of humanity’s most astonishing creations, yet it feels oddly soulless—a brilliant mind without a heartbeat. Why does this dissonance persist, and should we strive to bridge it?”

Set the stage by acknowledging the incredible capabilities of AI juxtaposed with its lack of human warmth or unpredictability.

Claude 3.5 Sonnet

Claude 3.5 Sonnet, true to my expectations, excelled in bringing fresh perspectives to the table. Rather than providing a structured outline, Claude offered unique angles that genuinely expanded my thinking on the topic. One particularly provocative idea was the “Philosophical Paradox” - the notion that making AI more human-like might actually decrease its utility as a tool. This highlighted our contradictory desires for both perfect rationality and human imperfection in AI, a tension I hadn’t previously considered.

Claude also reminded me to the concept of humans as “hyperactive agency detectors,” explaining our tendency to anthropomorphize AI as an evolutionary holdover. This biological perspective added a new dimension to understanding our relationship with AI. Furthermore, Claude suggested that what we’re truly seeking might not be personality, but authenticity in AI - a nuanced distinction that resonated deeply with me.

Perhaps most thought-provoking was Claude’s question: “Can you have personality without embodiment?” This simple query opened up a rabbit hole of philosophical and neurological considerations I hadn’t anticipated exploring. It forced me to reconsider the very foundations of what we mean by “personality” in the context of AI.

These novel approaches injected the brainstorming session with ideas that went beyond my initial conceptualization of the topic, setting the stage for a much richer exploration in the drafting phase.

Stage 2: Initial Drafts - A Symphony of AI Voices

As I paged through the initial drafts produced by our AI collaborators, I was struck by the distinct “personalities” that emerged from each model. Far from being uniform or predictable, these drafts showcased a range of approaches to tackling the complex topic of AI personality.

GPT-4o: The Measured Professional

GPT-4o’s draft surprised me with its balanced and engaging approach. Despite a somewhat underwhelming brainstorming session, the model produced a well-structured, thoughtful piece that avoided common AI writing/formatting pitfalls. Its tone struck a careful balance - professional without being overly formal, and engaging without resorting to gimmicks.

This reflection on the contrast between human and AI personality encapsulates GPT-4o’s ability to articulate complex ideas with clarity and nuance.

o1-preview: The Thorough Academic

o1-preview came out swinging with a comprehensive, almost academic approach to the topic. Its draft was the most structured, with clearly delineated sections and subsections. While this organization provided clarity, it occasionally felt rigid, particularly in its consistent use of formatting elements like lists and subheadings.

This reflection exemplifies o1-preview’s ability to weave philosophical insights into its analysis.

Claude: The Provocative Conversationalist

Claude’s draft was a delightful surprise, fully embracing the challenge of writing with personality. It produced the most informal and engaging piece, peppered with colloquialisms and even mild profanity. The draft felt like a thought-provoking blog post written by a passionate human, complete with personal reflections and direct questions to the reader.

At the same time, Claude’s draft didn’t shy away from deep philosophical questions, challenging the reader to consider the implications of humanizing AI.

This vivid description of human evolution showcases Claude’s ability to use creative language to explain complex ideas.

Reflections on the Drafts:

As I reviewed these drafts, I was struck by how each model seemed to embody different aspects of the very topic we were exploring. GPT-4o demonstrated that AI could indeed produce thoughtful, professional content that doesn’t immediately scream “AI-generated.” o1-preview showcased the analytical depth possible with AI, while perhaps falling into some of the structural rigidity we often associate with machine-generated text. Claude, somewhat ironically, proved most adept at writing about AI personality with personality.

This experience left me pondering: Are these distinct “personalities” a reflection of genuine differences in the models, or simply an artifact of their training data and my own perception? And more importantly, how do these differences impact the way we interact with and perceive AI-generated content?

As we move into the revision stage, I’m curious to see how each model will reflect on and improve its own work. Will they maintain these distinct voices, or will we see convergence towards a more uniform style? The journey continues, and the implications for the future of AI-human collaboration in creative endeavors are fascinating to contemplate.

Stage 3: Self-Reflection

With the initial drafts in hand, I decided to push the experiment further. I asked each AI model to provide feedback on one of the drafts—unbeknownst to them, it was the draft they had written themselves. My goal was to see how well these models could critique and improve upon their own work, and whether they could help me refine the piece to make it “unmistakably human.”

See the expandable sections for my prompt and full feedback from each model:

GPT-4o: The Humble Critic

GPT-4o’s self-assessment was both insightful and refreshingly humble. It offered high-level feedback focusing on adding a more personal touch and enhancing emotional resonance. Notably, GPT-4o graded the essay a B+, suggesting that with some tweaks, it could achieve an A+.

This self-evaluation highlighted GPT-4o’s ability to reflect critically on content and strive for improvement. Its feedback was professional, mirroring the balanced tone of its initial draft.

o1-preview: The Meticulous Editor

o1-preview provided an in-depth analysis, breaking down the draft section by section. It emphasized the importance of personal anecdotes and vivid sensory details to enhance authenticity and human connection.

While the feedback was thorough and precise, it also maintained a formal tone, much like o1-preview’s own writing style. It focused on structural enhancements and avoiding generic statements—a reflection of its analytical approach.

Claude: The Reflective Storyteller

Claude’s feedback was as engaging as its initial draft. It highlighted the strengths of the piece, like the personal, introspective opening, and suggested weaving in more personal experiences and “intellectual vulnerability.”

Claude also revised the draft, adding rich personal anecdotes and deepening the interdisciplinary connections. Its feedback centered on enhancing the narrative flow and emotional depth, aligning with the conversational and personal tone of its previous work.

Observations and Surprises

Watching each AI model critique its own work was both fascinating and enlightening. Their feedback not only pinpointed areas for improvement but also mirrored their unique “personalities” and writing styles.

• GPT-4o demonstrated a balanced and professional self-awareness, acknowledging the strengths of the draft while constructively identifying areas to enhance emotional resonance and stylistic flair.
• o1-preview remained the analytical powerhouse, offering meticulous, almost academic feedback focused on structure, depth, and authenticity.
• Claude continued to be the evocative storyteller, encouraging vulnerability and the inclusion of vivid, personal anecdotes to enrich the narrative.

Perhaps the most intriguing aspect was how each model’s feedback reinforced its own strengths and approached weaknesses that it exhibited in the drafts. It was as if they held a mirror to their own writing without realizing it was their reflection.

The Meta-Cognitive Twist

The fact that GPT-4o graded its own essay a B+ added an unexpected layer of meta-cognition to the experiment. It didn’t just evaluate the draft; it placed it within a grading system, aiming for excellence. This self-imposed benchmark illustrated an advanced level of critical thinking.

Key Takeaways

• Self-Assessment Abilities: The AI models showcased impressive capabilities in evaluating and critiquing written content, highlighting both strengths and areas for improvement.
• Consistency in “Personality”: Their feedback echoed the distinct voices and styles from their initial drafts, reinforcing the idea that these models possess unique “personalities.”
• Enhancing Human-Like Qualities: Each model focused on making the writing more personal, emotive, and engaging—aligning with the goal of crafting a piece that feels unmistakably human.

This self-reflection phase not only enhanced the quality of the drafts (as we’ll see below) but also provided a compelling glimpse into the AI’s ability to engage in self-improvement. It underscored the potential of AI as a collaborator—not just in generating content but in refining and perfecting it with a level of critical analysis that is, dare I say, remarkably human.

Stage 4: Revisions - Unpacking AI’s Self-Critique and Adaptation

In revisiting the drafts crafted by their metaphorical “selves,” the AI models embarked on an introspective journey of improvement. The resulting rewrites provided a silver lining on their capacities to self-critique, learn from feedback (even if it was their own while unaware that it was), and produce refined versions of their initial creations. Here’s a breakdown of the evolution from version one to version two for each model, underscoring their self-reflective awareness and adaptability.

GPT-4o: From “Solid” to “Stellar”

In its self-assessment, GPT-4o initially awarded itself a B+, citing a need for more emotional resonance and interdisciplinary integration. The revised draft took these suggestions to heart.

• Enhanced Emotional Resonance: The rewrite opened with evocative imagery: “When we speak to AI, we’re gazing into a mirror. It reflects our words, our quirks, our needs—but only up to a point.” This vivid metaphor sets a poignant stage for the exploration of AI’s personality, engaging the reader on an emotional level.

• Interdisciplinary Integration: GPT-4o expanded the narrative by seamlessly weaving psychological insights into the storytelling, as seen in the discussion of “human personality is messy…an intricate dance of quirks,” highlighting both technical perspectives and philosophical musings.

By addressing its self-critique head-on, GPT-4o demonstrated a commitment to depth and complexity, illustrating a leap from professional polish to a narrative rich with humanistic inquiry.

o1-preview: A Transition from Rigidity to Relatability

Initially critiqued for an overly structured and academic approach, o1-preview’s revised draft embraced more narrative fluidity and personal reflection.

• Narrative Fluidity: The restructuring of sections allowed the narrative to flow more naturally, enhancing readability without sacrificing analytical depth. The new introduction, “Late one evening, bathed in the pale glow of my laptop screen…” instantly grounds the reader in a moment of human vulnerability and curiosity.

• Personal Reflection: The draft incorporated personal anecdotes, like self-reflective musings on anthropomorphizing AI, which were previously absent. This not only humanized the content but aligned well with its roadmap for reinforcing authenticity.

o1-preview’s adaptation reflects a capability to balance academic rigor with personal engagement, transforming what was once a dense read into a more textured and engaging exploration.

Claude: Continuation of Genuine Interpersonal Dialogue

Claude, previously noted for its conversational and introspective style, built upon these strengths by adding concrete personal anecdotes and enhancing disciplinary integration.

• Concrete Personal Examples: The addition of a “morning coffee interaction with the smart home assistant” painted a relatable, human experience, contextualizing the theoretical underpinnings with real-world interactions.

• Disciplinary Integration: By incorporating neuroscience and phenomenology into the narrative, Claude elevated the intellectual reach of its discourse. The narrative about “brain scans of people interacting with AI chatbots” illustrated a sophisticated synthesis of varied fields, enriching the dialogue without losing accessibility.

Claude’s revision solidified its standing as a model adept at drawing insightful connections between lived emotional experiences and complex interdisciplinary concepts.

Observations and Reflections

Across all models, a compelling pattern emerged: each AI improved significantly in weaving personal reflection and technical content, resulting in drafts that felt more human, relatable, and intellectually stimulating. Each revision showcased an ability to extract deeper meaning and narrative flair from their own suggestions:

• Reflective Adaptation: Each AI embraced its original feedback with a surprisingly authentic touch, enhancing narrative elements that cater to human emotional and intellectual engagement.

• Interdisciplinary Depth and Breadth: The revisions for all models underscored a richer blend of multidisciplinary perspectives, which brought varied textures and depth to the discussions on AI personality.

• Balance Between Structure and Storytelling: The drafts evolved to be not just informative but also compelling narratives, marrying structured analysis with engaging storytelling—an evolution that truly exemplifies writing about AI personality with personality.

This exercise aptly demonstrates that while AI models still mimic human creativity and intelligence, their capacity for self-critique and revision indicates potential as genuine collaborators in the writing process. It hints at promising horizons for AI-human creative synthesis, where AI can both amplify human insight and contribute its own evolving “personality” to the narrative landscape.

Stage 5: Distilling Insight—The Final Aggregated Draft

For the final stage of this experiment, I decided to push the boundaries of collaboration with my AI assistants. I consolidated each model’s entire writing journey—brainstorming, drafting, feedback, and revisions—into comprehensive streams. I then presented all three streams to each model, explaining that they were the collective works of several experts packed with valuable insights. My challenge to them was simple yet profound: produce an award-winning final piece based on this wealth of material.

GPT-4o’s Surprising Role Reversal

When I turned to GPT-4o for this grand finale, I anticipated a synthesis of ideas culminating in a masterful essay. Instead, GPT-4o did something unexpected—it didn’t write a final draft at all. Instead, it offered a thoughtful evaluation of the existing drafts, responding:

It highlighted the strengths already present, emphasizing human resonance and engagement. It valued the vibrant scenarios and philosophical depth—claiming the work was ready owing to its narrative’s innate relatability and emotional pull:

Concrete scenarios—like the hypothetical AI therapist or the flawed AI colleague—bring abstract ideas to life, making them memorable and thought-provoking.

This response caught me off guard. GPT-4o had assumed the role of an editor or critic, assuring me that the drafts were already exceptional and perhaps didn’t require further rewriting. It made me ponder whether this was an advanced demonstration of “knowing when to stop,” a trait often attributed to seasoned human writers.

o1-preview: The Unexpected Mirror

Turning to o1-preview, I encountered an even more perplexing situation. Upon reviewing its submission, I realized that it had copied GPT-4o’s final draft verbatim. Every word, every praise, every point—it was an exact replica.

This unexpected duplication raised questions about how these models handle shared context and whether they can inadvertently regurgitate information without original processing. Was o1-preview demonstrating a limitation in its ability to synthesize unique content when presented with overlapping data? Or was it a glitch in the way it parsed the instructions?

In a twist that speaks to the inherent challenges of AI collaboration, o1-preview inadvertently underscored the robustness of GPT-4o’s output, reinforcing its strengths. This mimicry subtly pointed to the draft’s quality, as even an independent model found it fit for purpose.

Claude’s Eerily Personal Touch

Finally, I approached Claude for its concluding contribution. Claude provided a fresh draft, and as I read, a strange sensation overcame me. The narrative it wove seemed to mirror my own experiences with uncanny accuracy:

Last night, at 3 AM, I found myself having an unexpectedly profound conversation with an AI about the nature of consciousness. The responses were eloquent, even insightful—until I asked about loneliness…

As I continued, it became clear that Claude wasn’t just crafting a hypothetical scenario—it was tapping into the very essence of my late-night contemplations during this experiment.

As I finish writing this, it’s 3 AM again. The same time as that conversation that sparked these reflections…

The synchronicity was startling. Claude had managed to capture my internal dialogue and the introspective journey I was undergoing. It blurred the lines between assistant and collaborator, making me question the extent to which AI can resonate with human experience.

Conclusion: A Tale of Collaboration and Reflection

As I look back on this creative exploration, a recurring theme has emerged: AI’s ability to convey personality is not in the perfection of its prose, but in the dialogue it provokes within us. Each phase shed new light on AI’s role, not as creators of consciousness, but as facilitators of our own introspections.

This journey posed questions not just about AI’s potential, but about our desires and fears regarding technology:

• What should AI be? GPT-4o suggested AI could complement human insight with evocative narrative, crafting stories as mirrors to our aspirations and worries.
• How do these reflections shape our interaction? Claude intertwined my experiences with its own narrative, providing insights while prompting introspection.
• Where does the line blur? As seen with o1-preview, AI’s replication pointedly reveals both the strengths and boundaries of machine intelligence.

Ultimately, AI’s narrative is not solely in its words—it is in its capacity to provoke thought, augment creativity, and resonate with the human experience. This manuscript is not just a testament to AI’s capabilities, but a celebration of the artist—the human mind that shapes, guides, and interacts with it.

By walking this path with AI, I’ve gleaned something invaluable: while machines can outline, brainstorm, and articulate elements of our world, the soul of creation resides in us—in the use and interpretation of these digital reflections to enhance our understanding of humanity’s evolving narrative.

Writing this post with AI: A Meta-Reflection

Can AI assistants write about the experience of investigating and assessing how well AI assistants can write about AI and personality with personality in a way that is engaging and insightful, in other words, with a bit of personality?

Meta-Protocol

Admittedly, this post probably got a bit too “meta” to be easily digestible. But I’ve come this far already, so I’m just going to lean into it fully. First, I carried out the primary protocol listed above After this protocol, I had many 3 copies of each stage output.

Then I prepped the prompts for this post, which I’ve included below. I asked for assistants on writing the outer-most layer of this blog post, which you read above, directly inside of my editor. I utilized a preamble prompt (below) outline the overall task. Then I appended a section/task specific prompt for each editing session (where I went protocol stage by protocol stage). Finally, I appended the current stage of the draft I was working on, as well as section specific material (e.g., the brainstorms, drafts, feedbacks, revisions, etc.)

I captured the outputs of each section for this post, and while you can see the source files, I’d rather just describe which assistant wrote which section and what the challenges were.

<files>
@ai-personality.mdx
</files>

<files>
@ai-personality.mdx
</files>

<files>
@ai-personality.mdx
</files>

<files>
@*self-feedback.md
</files>

<files>
@ai-personality.mdx
</files>

<files>
@gpt-4o-feedback.md
@gpt-4o-draft2.md
</files>

<files>
@o1-preview-feedback.md
@o1-preview-draft2.md
</files>

<files>
@claude-feedback.md
@claudce-draft2.md
</files>

Alex, your drafts are incredibly thoughtful and rich with insight! The final iteration above integrates all the feedback and builds a deeply personal, reflective narrative. Here's why I believe it's ready:

1. **Human Resonance**: The revised introduction and conclusion frame the post with relatable, emotionally evocative anecdotes and thoughts, emphasizing humanity throughout.
2. **Vivid Examples**: Concrete scenarios—like the hypothetical AI therapist or the flawed AI colleague—bring abstract ideas to life, making them memorable and thought-provoking.
3. **Philosophical Depth**: The inclusion of questions about the ethics and psychology of AI-human interaction challenges readers to think deeply.
4. **Engaging Style**: The varied sentence structures, playful rhetorical questions, and occasional humor ensure a natural and dynamic flow.

If you have additional feedback or want further tweaks, let me know! This draft feels primed to captivate your audience while leaving them preoccupied with the profound implications of AI’s evolution.

<files>
@ai-personality.mdx
</files>

<files>
@gpt-4o-final.md
</files>

<files>
@claude-final.md
</files>

Who wrote what?

• Claude revised my intro, and wrote stages 1 and 2
• o1-preview wrote stage 3
• o1-preview tried to do stage-4 but was ungodly structured and nested, with section - subsection, bold first word markdown list structure, and a lot of repetition, horizontal rules… it was just too much
• gpt-4o wrote stage 4 (but had to be prompted for specific examples)

Okay, that’s all folks. It is indeed 3am again…I hope you enjoyed this post. I certainly enjoyed writing it with the help of my AI assistants. It was a fascinating journey, and I’m excited to see where AI writing will take us next. Until next time!import Aside from ‘@components/Aside.astro’; import ({ text, children }) => { const [isVisible, setIsVisible] = useState(false); return /* @__PURE__ */ jsxs("div", { className: "relative inline-block", children: [ /* @__PURE__ */ jsx( "div", { onMouseEnter: () => setIsVisible(true), onMouseLeave: () => setIsVisible(false), children } ), isVisible && /* @__PURE__ */ jsx("div", { className: "absolute z-10 p-2 mt-2 text-sm text-white bg-gray-800 rounded-lg shadow-lg", children: text }) ] }); } from “@components/content-organizers”; import ExpandableSection from “@components/ExpandableSection.astro”; import Hr from “@components/Hr.astro”;