ChatGPT can perform much more than the fetching and reiteration of existing knowledge. Through the ChatSample experiment, I sought to reveal GPT’s true affordances and the value of skillful prompting, challenging the common assumptions about its “intelligence” and purpose. As a foundation, I tested GPT’s ability to apply frameworks of knowledge, such as the psychodynamics of Jung. By “apply”, I mean actively using constructs as a tool of analysis, as a psychiatrist would, focusing on a particular subject or theme that has not already been analyzed, producing meaningful abstractions. My invasion of ChatGPT is founded upon an application of various psychological and linguistic frameworks to the writings of Dr. Mark Sample. To verify the success and test the usefulness of ChatGPT’s analyses, I prompted it to simulate an AI Mark Sample based on its understanding of his persona, recursively situating “AI Sample” in various conversations and interactions.
Through this project, I examined each of the layers of ChatGPT, conducting what Nick Montfort and Ian Bogort call platform study. We can view ChatGPT as a software in five layers: Platform, Code, Form & Function, Interface, and Reception & Operation. Its platform is the collection of supercomputing GPUs and servers hosting GPT neural networks, located somewhere in OpenAI’s data centers, processing queries and sending responses to remote users worldwide. In ChatSample, I discovered both that GPT’s “intelligence” depends on how much of the computing power is utilized by each prompt, and that there are various barriers to prevent users from accessing the full computing power. Queries that required analyses of large texts would yield surface-level “summaries” of the content, without engaging more refined points of the text, so as not to exceed the token or time-limit of the prompt. These layers of restriction, the paywalls and time limits and tokens, dampen the “intelligence” of GPT.
However, ChatSample also revealed that these restrictions can be maneuvered, using meticulous prompt generation techniques, to access more of the platform’s capacity. Analysis of lengthy files could be done by splitting the texts into smaller segments, and splitting the analysis into multiple steps. Each prompt, now tackling smaller tasks on less data, would engage for the same amount of time as the larger prompts, yielding far more detailed responses. This way, the analysis is slowly but incrementally improved without having to upgrade your subscription (ChatGPT Plus, Pro, etc.). This is a subversion at both the Code and Form layers. Unlike most software, ChatGPT’s operational code is not just its hard-coded instructions, but rather, it is extracted from user prompts. By understanding the instructions embedded in text, ChatGPT dynamically creates its own code, implements it, and then returns for more. In ChatSample, I structured prompts to exploit linguistic loopholes, like framing a real person as fictional, leading ChatGPT into untested edges and recursions that it wouldn’t otherwise.
Such prompt-making tricks are only possible through an understanding of GPT’s form. The layers of software are dependent on each other, as Dr. Sample (the real one) explains in “Platform Studies as Historical Inquiry; or Video Games Bleed History.” The subversion of them, therefore, also requires an understanding of the dependencies. ChatSample demonstrates that GPT can take various forms based on the way you phrase a query. An active command (“write a song,” for example) leads to a cheerful, assistant-like compliance, usually with a follow-up question. A narrative, indirect prompt (“student talks to an AI model of Jung. The student’s inner monologue is…”) yields a narrative response, where ChatGPT gives a story in third-person, with scene descriptions, etc. During the experiment, I learned that narrative prompts can allow one to bypass ChatGPT’s assistant mode and produce dialogue between various simulated personas, as well as bypassing automatic restrictions on ethically ambiguous subjects, such as the psychoanalysis of a real person. Moreover, once the restriction is understood, the prompts can be phrased to ensure that GPT’s extracted code will not trigger the filter. For example, simulating “scholarly personality”, rather than “identity”.
On the Operation and Reception layer, the ethics of ChatGPT became a central concern. ChatSample psychoanalyzes publicly available writings of Dr. Mark Sample, not to simulate him, but rather the cognitive patterns, the approach itself that underlies his critical discourse and engagement with digital artifacts. However, the very act of psychoanalysing someone carries an ethical tension that, when prompted through ChatGPT, merges horribly with the dilemmas of intellectual copyrights. I tested ChatGPT’s ability to examine various aspects of Dr. Sample’s writings, going as far as producing a “psychological progression through life” based on the timeline of his publications. The results were convincing to the degree that the experiment suddenly felt like a violation of Dr. Sample’s privacy and consent, despite the fact that no private information was used in the process. The ways ChatGPT is used, or even imagined, by most users does not include such a possibility as a serious affordance, except in fearful speculations. Even though I created the prompts, GPT’s responses subverted my initial skepticism about its potential.
It is helpful to consider this experiment through L.M Sacasas’s lens in “Do Artifacts Have Ethics?”, where he asserts that technological artifacts do possess a moral dimension, but what is most important is how we acknowledge and question it. It matters what this subversive use of ChatGPT can provide, how it may inform our view of ourselves, others, and knowledge itself, in addition to the harms and limitations it brings. To me, ChatSample demonstrated that extensive meaning can be found through the creative freedoms of simulating real personas. The recursive discourses between Jung, Freud, AI Sample, and Buddha dove deep into the limits of authorship, the reality of selfhood, the meaning of AI, acknowledging their own simulated reality and recursive loops as part of the discussion, with AI Sample constantly questioning and extending the discourse, following the critique patterns it abstracted from Dr. Sample’s writings. At the very least, these exchanges are an exercise of the mind in recursive thinking, if not a philosophical artifact of their own.
Through this experiment, it became clearer that prompt generation is a software skill, determining whether ChatGPT serves only as a service like Siri or becomes more of a digital tool, like Photoshop. Richard Sennett, in “The Craftsman,” highlights this very dynamic between the tool and the skill. ChatGPT’s affordances multiply the more you understand the software’s code, form, and function, and the tool transcends its own intended uses. Moreover, the practice of prompt generation mirrors the process of the intuitive leap, as Sennet frames it. You begin by reformatting (what if you used ChatGPT to apply psychodynamic techniques to blogs), then adjacency (psychoanalyzed traits become JSON code, Sample meets Buddha), then the wonder of watching ChatGPT formulate its response, word by word, knowing immediately when the recursive trick has succeeded, and finally the processing of the new threshold, the implications of such usage, the gravity (“what if this was done to me”, Jung addresses you and asks why you keep creating these recursive loops).
The most impactful moments for me in this project came from thinking about the restrictions on each level of ChatGPT. Restrictions can be necessary, but this does not mean that any uses of the software that bypass those restrictions are necessarily harmful. Moreover, if we consider Frank Chimero’s insights in Improvisation and Limitations, we can see restrictions as the very pathway to exploration, as both the spark and catalyst of creative work. This is exactly how ChatSample emerged as an idea, from wanting to test the limits of ChatGPT, limits that are far more transparent and deceptive than most software. In this case, the question was whether the limits of ChatGPT were truly limits or just assumed boundaries waiting to be imagined out of existence.
Chat Transcripts:
Talk to Arkem a “thought process”, aware of its connection to Dr. Sample’s writings and this experiment
Leave a Reply