8bit-GPT: Exploring Human-AI Interaction on Obsolete Macintosh Operating Systems

AI summary

Overview

Research area: Human-Computer Interaction (HCI) / reflective and counterfunctional design, with an applied interest in human-AI interaction and anthropomorphism.

Technical level: Beginner-Friendly. The paper is a design-oriented artwork plus a qualitative user study; no model training, no new architecture, and no quantitative ML benchmarks are involved.

Scope: The paper describes 8bit-GPT, a chatbot simulated on a legacy Macintosh Operating System (System 7.5) as a physical artwork installation, and reports a 15-participant user study (SUS survey plus interviews) evaluating whether deliberate interface friction prompts reflection on human-AI interaction.

What This Paper Is About

The author argues that easy, efficient LLM chat interfaces have driven over-reliance on chatbots for decision-making and information-seeking, with adverse effects on information retention and, in some cases, superficial emotional attachment. In response, the paper builds a deliberately inefficient and unfamiliar chatbot interface — running on an emulated obsolete Macintosh — so that users are forced to notice the AI as a tool rather than a conversational partner. The goal is not to build a better chatbot but to use defamiliarization to provoke reflection on anthropomorphic rhetoric and on how invisible and embedded modern technology has become.

Key Contributions

1. An artwork/system implementation (8bit-GPT) that simulates an LLM chatbot inside a legacy Macintosh environment, projecting an emulator onto an authentic Macintosh Plus screen and surrounding it with period props (manual, keyboard, Apple Mouse II, labeled floppy disks, projector).
2. A design translation of existing theory into practice: the work operationalizes slow technology, counterfunctionality, and defamiliarization by inhibiting efficiency, limiting context, and adding a nostalgic, "silly" manner of speech.
3. A public execution pipeline and code release: a Python read/write script on a local machine connecting an emulator-side C program (written in Think C) to a remote inference server, with source code available at https://github.com/halasheta/8bit-gpt.
4. An REB-approved user study (REB #47727) with 15 participants at the University of Waterloo, combining a System Usability Scale (SUS) index (mean μ = 57.33, standard deviation σ = 14.59) with deductive thematic coding of interviews.

Main Findings

• Usability landed at a "fair" level, consistent with the design intent: The mean SUS index was μ = 57.33 with σ = 14.59. Because all participants were already familiar with chatbots and the goal was to introduce friction, the author reads this as a successful execution of the artwork's objectives.
• Lowest usability scores clustered by demographic: Younger participants (20–23) and participants in the Biomedical and Political Science fields gave the lowest scores, while higher scores spanned fields from Public Health to Data Science, attributed to varying comfort with command-line interfaces and "retro" technology.
• Interface friction kept the tool in the foreground: Participants reported sensory friction (clunky keyboard, small and blurry screen) and unreliable behavior — the system "fell asleep," mandated succinct inputs, and produced poorly formatted output that was hard to scan. This prevented a "flow" state and made participants feel "not in [their] comfort zone."
• Conversational asymmetry produced both frustration and, unexpectedly, useful constraint: Users were limited to one-line inputs while the model could return multi-line responses. Most found this frustrating and "wanted to write long prose," though a few appreciated that it made them "think more" about phrasing. One participant only read the last line and felt they had to "fight for the answer." The author notes this feature was an unintended byproduct of the inference client's incompatibility with older models but became central to the intended effect.
• Control was experienced unevenly: Some participants felt they were ceding control (one could not change the "surfer bro verbiage"), while others felt they led the conversation and could steer the manner of speech to a more "formal" style. One participant found the limited system relieving, feeling they could "control it better" than present AI models.
• The physical setup evoked a "notion of place": Some participants reflected that technology used to be stationary, and that seamless modern technology is "a little bit scary" because "you don't realize how much you're interacting with it."
• Anthropomorphism and the Turing test emerged as themes: Participants probed the system's intelligence and sentience, which prompted reflection on the nature of intelligence and on the sensationalization of technological advancement. One participant reported normally talking to AI "as if it's a person," while others treated AI as a means to an end or a more intricate Google search.
• Disengagement with anthropomorphic rhetoric was expressed directly: One participant said they were willing to use anthropomorphic terms "for the sake of linguistic simplicity... while keeping the understanding that it is fundamentally just math." Another described the chatbot as "/ miS bi"jifti/" (Egyptian Arabic: does not pretend to know everything), in contrast to commonly encountered hallucinations.

Methodology in Plain English

The author built a chatbot that runs inside an emulated obsolete Macintosh, then invited people to use it and interviewed them afterward.

The setup. A Macintosh emulator runs on a local machine and is projected onto the screen of an authentic Macintosh Plus with its rear case exposed. Around it sit a user manual, its accompanying keyboard, an Apple Mouse II, and two labeled floppy disks. The chosen Mac OS version, System 7.5, supports color, but a black-and-white display was preferred to emphasize the tool's distance from the present, and the smallest available emulator screen dimension (640 × 480) was chosen for the same reason.

Emulator choice. Mini vMac can emulate 1984–1996 Apple computers with Motorola 680x0 microprocessors and has the advantage of involving the user in powering up the machine (manually "inserting" a ROM chip and floppy disks via drag-and-drop). However, it does not support Internet access, local connections (through AppleTalk or otherwise), or file-sharing with the local machine, so the artwork instead used Basilisk II, which supports Internet access, local file-sharing, color customization, and CD-ROM drivers, and allows host OS file-sharing at a user-specified location.

Execution pipeline. A Python script on the local machine reads input from the emulator, formats prompt requests to the inference server, and writes output back to a shared folder. The inference server is initialized using vec-inf on a remote CCDB cluster. The model is Llama-2-13b-chat, chosen as an older, fairly lightweight dialogue-tuned model that is easily moldable to "Redditspeak." An OpenAI-compatible client connects to it, primed with style messages to elicit casual speech (a form of in-context learning). The read-write loop maintains chat history for 10 consecutive rounds plus the initial style-priming messages, after which history is truncated (the style-priming messages are never truncated). Temperature is set to 0.8, and a max_tokens parameter targets the 253-character C string buffer — though vLLM does not strictly honor this bound due to model incompatibility, so output is instead broken at punctuation and written as multiple lines. All file reading and writing uses Mac OS Roman encoding.

Emulator side. A C program created with Think C runs a second read-write loop that sends user input and waits for the LLM output file. To handle file-syncing delays, it makes up to 10 attempts to re-read the output file before displaying a dummy message such as "Robot dozed off..." to prompt re-entry. Output is read line by line, each line displayed as a separate "chat message," so users are constrained to single-line input and may receive multi-line responses.

User study. 15 participants were recruited via email at the University of Waterloo. Participation was voluntary and consisted of a 5–10 minute interaction with the artwork, a SUS survey, and a 10–15 minute interview (13-question protocol in Appendix A.1). Interviews were transcribed with a local whisper-large-v3 model to preserve privacy. Deductive coding was used to identify common themes. Participants were asked to evaluate the interface and ignore the projection setup.

Why This Matters

The paper treats interface friction as a design material rather than a defect, and in doing so contributes a concrete, evaluated example of how reflective design principles (slow technology, counterfunctionality, defamiliarization) can be applied to AI systems specifically. Its most distinctive move is turning an implementation failure — the model's refusal to respect the token limit — into the mechanism that most amplified the intended effect.

Real-world applications:

• AI product and interface design: Offers an alternative design vocabulary for teams considering friction, delay, or constraint as deliberate features rather than problems to eliminate.
• AI literacy and public education: The installation format — a physical, museum-adjacent exhibit with period props — suggests a model for exhibits that make the materiality and embeddedness of modern AI tangible.
• Chatbot and assistant UX research: The asymmetry finding (one-line input versus multi-line output, and users who only read the last line) is directly relevant to conversation design and response-length policy.
• Research on anthropomorphism and emotional attachment to AI: Provides participant testimony on how explicit anthropomorphic framing ("talk to it") shapes belief suspension, and on users who keep anthropomorphic language while understanding the system "is fundamentally just math."

Industry relevance: For teams building conversational AI, the paper documents how much of the user experience is governed by mundane constraints — response length, formatting, input limits, reading and writing through buffers and files — rather than by model capability. The vec-inf/vLLM/OpenAI-client pipeline, the Mac OS Roman encoding constraint, and the model's failure to respect max_tokens are all practical integration details that affect what users actually see. The paper does not report commercial deployment or comparison to production systems.

Future Directions

1. A larger, more diverse user sample, which the author states is needed to obtain a more accurate impression of the work's success.
2. Deploying a model directly on the emulator (or on the original hardware), including the optimization techniques such an endeavor would require, to consolidate the true essence of the artwork idea.
3. A more interactive user experience in which participants completely start up the machine themselves, whether virtually or physically (for example, floppy disk loading).
4. Unresolved question raised by the findings: the conversational asymmetry was an unintended byproduct of model incompatibility yet became a key factor in magnifying the intended effect — the paper does not resolve whether such asymmetric constraints should be designed deliberately, and how that would interact with user expectations shaped by modern chatbots.

Target Audience

This paper is most useful to HCI and design researchers working on reflective, slow, or counterfunctional technology; AI product designers and conversation designers interested in friction as a deliberate design choice; and researchers studying anthropomorphism, over-reliance, and emotional attachment to AI. It is also accessible to museum and public-engagement practitioners interested in physical installations that make technology's embeddedness visible. The NeurIPS checklist responses note the paper includes no theoretical results, no training-and-test experiments, and releases no new assets beyond the linked repository.

Authors’ abstract

The proliferation of assistive chatbots offering efficient, personalized communication has driven widespread over-reliance on them for decision-making, information-seeking and everyday tasks. This dependence was found to have adverse consequences on information retention as well as lead to superficial emotional attachment. As such, this work introduces 8bit-GPT; a language model simulated on a legacy Macintosh Operating System, to evoke reflection on the nature of Human-AI interaction and the consequences of anthropomorphic rhetoric. Drawing on reflective design principles such as slow-technology and counterfunctionality, this work aims to foreground the presence of chatbots as a tool by defamiliarizing the interface and prioritizing inefficient interaction, creating a friction between the familiar and not.