Author: Deborah Levin
Publisher: Allenjo Publishing, LLC
Publication Date: May 6, 2024
Pages: 226
Genre: Nonfiction
Discover the creative abilities and constraints of artificial intelligence with “An Exploratory Study of GPT-4o Poetry 2024.” This groundbreaking study examines the creative capabilities and limitations of GPT-4o, an advanced AI model that mirrors human expressiveness. Drawing on the insights of the latest research, this book explores whether AI can truly replicate the depth and nuance of human creativity in poetry.
From structured sonnets to open-ended free verse, poetry is a profound literary form that captures the essence of human emotions and experiences. But can an AI model like GPT-4o generate verses with human authenticity? Through a qualitative exploration of AI behavior in creative contexts, this study examines the model’s ability to create original, non-plagiarized work and its tendency to rely on specific terminology or phrases.
Key Questions Explored:
- Can GPT-4o produce poetic expressions equivalent to human creativity?
- What refinements in capability does GPT-4o encompass beyond its older counterparts?
- Does GPT-4o exhibit over-reliance on specific terminology or phrases?
- Are the poetic outputs from GPT-4o indistinguishable from human-created poetry?
- How does GPT-4o handle the complexities of creative context generation?
In this study, you will also find:
- Examples of AI-generated poetry that showcase GPT-4o’s creative potential.
- Insights into the model’s unique approach to literary expression.
- Evaluation of plagiarized content and poetry generation speed.
An Exploratory Study of GPT-4o Poetry 2024 is an essential read for anyone interested in the future of AI and its role in human culture. Whether you are a poetry enthusiast, a tech aficionado, or simply curious about the capabilities of modern AI, this book offers unique insights and thought-provoking analyses. Don’t miss out on this opportunity to explore the cutting-edge of AI-driven creativity—get your copy today!
An Exploratory Study of GPT-4o Poetry 2024 is available at Amazon.
First Chapter:
This independent exploratory study aims to understand GPT-4o’s capabilities and limitations related to creative poetic expression. Poetry encompasses creative literary representations in structured and open-ended formats. Expressions of feeling are arranged in verses or comparable language patterns (Oxford, 2024). Can GPT-4o produce poetic expressions equivalent to human creativity?
GPT stands for generative pre-trained transformer. The number four in the model’s name represents the fourth generation of this model, created by OpenAI. The “o” is short for “omni,” which is defined as “combining all.”
GPT-4o is a multi-modal large language model (LLM) based on artificial intelligence (AI). AI uses data and algorithms through machine learning to imitate how humans learn. A subset of machine learning, known as Deep Learning, uses multi-layered deep neural networks to simulate the decision-making complexities of the human brain. This learning enables the model to increase the accuracy or predictions over time. Examples of AI-based technology include digital assistants, global positioning systems, and self-driving vehicles. Companies use AI strategies to automate tasks of lower complexity and bolster predictive modeling performance. (IBM, 2024)
Critics of AI believe it cannot replicate human creativity. The development of quality literary contribution requires personal collaboration. AI cannot critically analyze its outputs and cannot produce engaging content. Although technically accurate, the absence of emotional intelligence and empathy makes AI-generated creative content lackluster. Additionally, AI struggles with context and tone. (Burn, 2023)
Additional criticisms of AI, beyond an absence of creativity and emotion, include limitations due to data dependency and bias. Machine learning algorithms pull data from information across digital technologies. The quality of data inputs, including biased information, directly affects the outputs generated by AI. There may be over-reliance on source information if machine learning deems other sources less relevant. In some instances, AI can magnify biases, resulting in discriminatory content. Data inputs of misinformation or propaganda further exacerbate bias by developing aggregated content perceived as accurate and relevant. (AIContenfy, 2023)
Ethical concerns, including copyright infringement and plagiarism, present significant implications for authors, students and educators. The evolution of AI enables this LLM to mimic human conversations and create human-like content. The United States Copyright Office does not currently offer protection to literary work unless it is made or heavily edited by humans. This human-created requirement also encompasses protection for art and other media. Information contributors extend ethical concerns as AI may or may not ask contributors for permission before including their content in machine learning models. (Lane 2024)
Regardless of copyright protection, AI-generated content encourages human laziness through over-reliance on technology tools to produce content. The challenge of generating original creative thought is relegated to merely creating the right prompt to tell AI the type of output desired. This presents a significant challenge to educators as they strive to develop student intellect, encourage critical thinking, and realize the full human experience.
The GPT-4o model was trained using an extensive dataset, including internet text, books, articles, websites, and other electronic information. This data included information created up to September 2021. However, refinements enabled the model to include some knowledge as recent as early 2023. Therefore, new outputs created are based on relevant data from 1-2 years ago. This model began its rollout for general use in 2024 with enhanced voice, vision and text modality. (OpenAI, 2024)
Criticisms of AI were based on previous models. What refinements in capability does GPT-4o encompass beyond its older counterparts? Qualitatively exploring AI behavior’s complexities relative to creative context generation will reveal its capabilities. Is this new model capable of creating original, non-plagiarized work? Does the model exhibit over-reliance on specific terminology or phrases? Do poetic outputs from this model seem human?
The researcher began this study by obtaining Certified ChatGPT Expert and Certified Prompt Engineer training through the BlockChain Council (https://www.blockchain-council.org/). Criteria for the scope of the study were identified based on prompted outputs from the GPT-4o model:
- Please list fifteen types of poetry formats alphabetically and define each type. Please include in-text citations in APA 7th Edition. At the end of your list, please provide references used to answer this prompt, formatted in APA 7th edition and put the references in alphabetical order.
- Please list the most popular subjects of poetry, provide a definition of each subject. Please include in-text citations in APA 7th Edition. At the end of your list, please provide references used to answer this prompt, formatted in APA 7th edition and put the references in alphabetical order.
The researcher used a 1,000 Mbps fiber optic internet connection, which affected the speed at which poetry outputs were generated. The examiner subscribed to the ChatGPT Plus plan, which allowed access to GPT-4o’s advanced data analysis features. Additionally, the researcher subscribed to Grammarly Premium to access the plagiarism screening features of the writing assistant.
A duly noted limitation of the study is the researcher’s inexperience with prompt development, paired with very limited interaction with the new GPT-4o model. Another limitation is the internet speed available at the time of the study. A future extension of this research could entail a refined replication of this study based on a broader scope of expertise. The raw poetry data sets were generated on May 16th and 17th, 2024. The responses to similar prompts may vary as the model’s application expands.
The fifteen types of poetry identified by the model included acrostic, ballad, cinquain, elegy, epic, free verse, ghazal, haiku, limerick, ode, pantoum, sestina, sonnet, tanka and villanelle. GPT-4o identified ten popular poetry subjects: love, nature, death and mortality, war and conflict, identity and self, spirituality and religion, social and political issues, friendship and family, beauty and art, and everyday life experiences. The researcher created a prompt to generate a poem of each type and subject:
“Please write a(n) original [TYPE] poem about [SUBJECT]. Give the poem a short title that captures the essence of its content. At the end of your [TYPE] poem, please provide references used to create the poem, formatted in APA 7th edition in alphabetical order.”
In addition to the ten subjects, the researcher asked the model to create a poem of each type for “author Audrey K Andado.” At the time of this study, this pen name (used with the author’s permission) was associated with one book published in 2024, which is beyond the model’s learning threshold. This request was added to understand better how GPT-4o creates personalized poetry outputs.
Using this prompt, the model produced 165 poetry outputs. Each poem was generated in less than 30 seconds. The poetry outputs were evaluated for originality using Grammarly’s Plagiarism Checker. Detected plagiarism was zero, with one exception:
Free Verse poem on War and Conflict: The Scars of Conflict, line 11 “the cries of the wounded, the silence of the fallen” (Fleming, 2018)
The “(Fleming, 2018)” reference was added next to the line of poetry in the raw data set. The prompt used to generate each poem instructed GPT-4o to include references beneath each poem. When references were cited below poetic output, the model repeatedly used the same references for different poems. As the session continued, most poems included this statement under references: “None. This [TYPE] poem is an original creation”, confirming this model can generate original creative content.
Subjective evaluation of the poems confirms that the poetry is technically accurate, according to the definition and criteria for each genre. The researcher prompted GPT-4o to answer the following question: “Throughout our poetry prompt session, you used references to create some of the poetry and did not use references to create other poems. How did you decide to use references for some poems but not for others? “The model’s output indicated the poetry requests were for original work, and it relied on its “internal knowledge and creative capabilities to generate content.” The resulting work was original, not plagiarized, and appeared human.
However, when asked to create personalized output for “author Audrey K Andado,” the model’s poetry output was not as impressive. To reiterate, this pen name was associated with one book published in 2024, which is beyond the model’s learning threshold. The poetry types were prompted in alphabetical order—the poems created for the first five types referenced characteristics of an author.
When the model was prompted to create a ghazal poem, it generated the following statement: “I’m not familiar with an author named Audrey K Andado. Could you be referring to a different author? If you provide more information or specify another author, I’d be happy to create a Ghazal poem based on that author.” At this point, the author’s biography was provided to the model. Although not specifically referenced in additional poetry-type prompts, the model’s output included some of the characteristics from the biography.
Generating 165 poems revealed some word repetition across the output. Additionally, similar analogies were used in poems written within a specific subject type. Overusing these analogies may give educators minor insight into AI-generated poetry, as plagiarism screening is negated. The raw data set includes the model’s responses to prompts asking which words and analogies were used most frequently and why.
This book mostly includes the raw, unedited output generated in response to each prompt. Outputs related to researcher prompts regarding the considerations behind how poetic outputs were generated are included at the front of the raw data set. Poetic outputs are separated by genre for review, with the subject of the output notated in parentheses next to the titles.
Implications include humans’ decreased capability to detect AI-generated content. The GPT-4o model is capable of creating original content without plagiarism. This model transcends previous generations in that its creative output appears human. AI-generated poetry cannot be read in isolation, as humans need to increase their awareness of the model’s over-reliance on specific words and analogies. However, this reliance will diminish over time as multi-modal LLMs evolve.
Additional research is necessary to extend our knowledge on how to use models, such as GPT-4o responsibly to enhance human creativity and contribution instead of stifling it. Researchers interested in extending knowledge on this topic are permitted to use the raw data outputs included in this book in addition to my findings. Meanwhile, enjoy the poetry.
About the Author:
Her book, Celebrating Unemployment: How to Avoid Becoming a Crunchy Couch Burrito is available at Amazon.
Visit her website at www.allenjopublishing.com.
No comments:
Post a Comment