What Is AI? Part 2 – Demystifying AI Through Four Acts

© 2025 Keren Wang — Licensed under a Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International License (CC BY-NC-ND 4.0). Educational use permitted with attribution; all other rights reserved. For permission requests, please contact the author 

*Continuing from: “What Is AI?” Part 1: A Survey of Imaginaries, Mythologies, and Rhetorical Structures of Thinking Machines

In the previous segment of our genealogical survey, we have traced ancient imaginaries that seeded today’s visions of machine intelligence. From Yan Shi’s mechanical performers to Talos’ bronze vigilance and the ritual logic of golems and Homunculi, we explored how mythic prototypes shaped the rhetorical terrain on which “AI” would later emerge as a technological category.

Whereas earlier societies used myth and ritual to make sense of uncanny forms of artificial agency and “thinking machines,” in this second part we turn from historical mythologies to present-day technological possibilities. The term “artificial intelligence” circulates in our cultural lifeworld with astonishing fluidity. It appears in commercials, policy reports, sci-fi movies, legal disputes, and everyday conversations. It is used to describe everything from basic computer programs to sophisticated generative models capable of complex knowledge-performance.

To untangle this discursive terrain, it is helpful to trace four rhetorical “acts” that structure the modern evolution of AI discourse. These acts are not purely chronological; they are symbolic reactions, each representing a different way of responding to shifting technological scenes and imagining what AI is, what it could be, and what it ought to become.

---

Act One: AI as Metaphor & Speculative Frame for Procedural Automation (1980s–1990s)

Although the phrase artificial intelligence entered academic literature as early as the 1950s—hinted at in Alan Turing’s 1950 essay “Computing Machinery and Intelligence,” and explicitly invoked in John McCarthy’s 1956 proposal for an “artificial intelligence” research project¹—AI did not become a vernacular concept until the early 1980s.²

From Laboratories to Everyday Life

By the 1980s, computing technologies were no longer confined to military labs and elite research institutions. Personal computers, corporate networks, and industrial automation began to populate offices, factory floors, and even private homes. The rhetorical scene changed: computers became increasingly embedded in the architecture of everyday life.³

As the technological landscape evolved, the agency attributed to machines began to shift.⁴ Companies increasingly marketed procedural automation as “AI,” even when such systems were little more than CNC-controlled mechanical tools or robotic arms executing strictly preprogrammed sequences. Early “expert systems,” such as the Symbolics 3640 Lisp machine, mimicked expertise through rigid rule-based procedures rather than adaptive reasoning.⁵

A system like this might diagnose failures in a specific model of aircraft engine or assist in specialized medical image processing, but it operated entirely through thousands of human-authored IF–THEN rules. The machine could act, but only through procedural execution; it lacked the capacity to revise or reinterpret its own logic. If conditions changed, a human engineer had to manually rewrite the algorithm. Yet corporations promoted these systems as “AI” to harness the cultural momentum of personal computing and the speculative fervor surrounding the emerging digital economy.⁶

Cultural Imagination: Between Utopian & Dystopian Frames

Even at this procedural-automation stage, AI was more than a marketing gimmick; it became a fertile site of cultural imagination. The era’s public imaginary drew heavily from the “Frankenstein complex”—the deep reservoir of mythic projections we explored in Part 1. AI served as a symbolic placeholder for society’s tacit desires, anxieties, and futurist fantasies.⁷

The bipolar utopian/dystopian narratives of this imagination are visible across many iconic cultural texts of the era:

• 2001: A Space Odyssey (1968), whose iconic HAL 9000 dramatized the fear of machine autonomy turning inward, becoming opaque, neurotic, and unpredictable (just like human consciousness).

• WarGames (1983) and The Terminator (1984), which both crystallized anxieties about runaway military automation and the fragility of political oversight.

• William Gibson’s Neuromancer (1984), which imagined AI as simultaneously liberating and disorienting—entities that subvert corporate control yet also escape human interpretive horizons.

• Iain Banks’s Culture series (1987–2012), which portrayed highly advanced worlds overseen by benevolent superintelligent machines known as “Minds,” articulating a radical post-scarcity vision in which the lifeworld itself flourishes as a self-subsisting artificial superintelligence habitat.

• Star Trek: The Next Generation (1987–1994), where androids like Data become utopian mirrors for humanity, exploring optimistic possibilities for embodiment, sentience, sensibility, and the desire for self-realization.

This wave of speculative creativity in the 1980s and 90s functioned partly as a societal rhetorical response to the acceleration and proliferation of computational systems and automation. Like the ancient myths of the golem, the Homunculus, Talos, and Yan Shi’s automata, these narratives dramatized the tension between AI’s profoundly liberating potential and its equally potent dehumanizing dark undercurrents—an ambivalence comparable to our existential anxieties in response to the advent of the nuclear age.

This gap between speculative fiction and technological reality also proved to be a short-lived one, with the rise of machine learning.

---

Act Two: The Machine Learning Turn (1990s–2000s)

“Over hundreds or thousands of iterations, the network becomes increasingly adept at algorithmic rhetorical invention, producing continuously refined messages tailored to a specific user. Each output responds to the need or problem that occasioned the algorithmic communication in the first place.“

Machine learning (ML) marked a decisive break from the procedural automation of earlier systems. Instead of requiring humans to continually rewrite programs to accommodate new tasks, ML enabled computational systems to adapt by identifying patterns within data. This capacity for self-modification, albeit often constrained by human supervision, reconfigured the very grammar of artificial intelligence.

Agentic Shift from Static Programs to Adaptive Algorithms

As computational systems evolved from static programming to adaptive architectures, we witnessed a corresponding agentic shift from the human programmer to the learning algorithm itself. Under unsupervised, supervised, and semi-supervised paradigms, algorithms were granted expanding autonomy.

ML algorithms are programmed to act beyond fixed, preprogrammed rules, recognizing statistical regularities and adjusting internal parameters in response. Their agency, or operative power, derives from statistical inference, optimization functions, and nonlinear transformations distributed across increasingly deep and complex hidden layers. These processes unfold within a scene saturated by digital trails—user inputs, weather logs, location data, sensor readings, biomedical scans, and social media metadata—all of which provide the raw training data for pattern recognition and inference.

Taken together, the anchoring purpose of this semi automated optimization becomes clearer. It aims at producing generalizable prediction, personalization, and a steadily expanding capacity for autonomy. These systems increasingly pursue complex goals and carry out intricate tasks with very little human supervision or prompting.

Illustrative Example: A Machine-Learning Wardrobe Recommender

To understand how machine learning works beneath the surface, let us imagine designing an app that recommends what someone should wear each morning and suggests suitable activities based on weather, lifestyle, and health considerations.

At first glance, our “daily wear and activity recommendation app” seems like a trivial convenience tool, but beneath the surface it requires a full machine-learning pipeline: large datasets, pattern recognition, hidden layers of knowledge-production, and an ongoing feedback loop that allows the algorithm to respond to changing real-world conditions, and adapt to its target audience over time.

What looks like a minor convenience is, in fact, powered by an artificial neural network (ANN)—the central architecture of modern machine learning.

ANNs are inspired (loosely and metaphorically) by the structure of the human brain: layers of interconnected computational units known as “artificial neurons” that pass signals forward, adjusting themselves as they learn:

1.Input layer – nodes collect and internalize raw data from the external world:

Every ANN begins an input layer of nodes that collect and internalize raw signals from the external world into its network of neurons.

In our hypothetical “Wardrobe and Daily Activity Recommender App” example,  the process of machine learning begins  by gathering inputs from two domains:

• Environmental data: temperature, humidity, air quality, UV index, local traffic (collected from third-party providers and public databases).

• User data: location, age, hobbies, daily schedule, health information (collected from device sensors, location services, login credentials, app permissions, and related sources).

2.Hidden layers – recombine raw inputs into tacit knowledge or meaningful patterns:

The “hidden layers” of an ANN perform the “learning” task by identifying, differentiating, and organizing meaningful patterns from the input data that are not explicitly written in its human-authored codes:

Hidden Layer 1 - nodes trained to  identify basic patterns such as  “high UV level + fair skin,” then pass to nodes in Layer 2

Hidden Layer 2 - nodes recombines Layer 1 patterns to broader lifestyle categories such as “hiker" and "commuter," then pass to other hidden layers

Rather than following human-authored rules, neurons in these hidden layers would adjust internal weights — calculated based on a set of ratios between competing patterns (i.e., “how much should certain word embeddings matter relative to other representation of a word or sentence in this specific context?“) — through repeated internalization and recombination of raw data, and gradually synthetizing actionable knowledge on their own.

This adaptive knowledge-production structure is precisely what distinguished ANNs from the rigid procedural systems of Act One.

3. Output layer – externalizes tacit knowledge into tailored, actionable recommendations:

After  compressing and reconfiguring patterns from the hidden layers, the network externalizes tacit knowledge into explicit performance with a set of tailored outputs, such as: “sun protection level SPF 50+” and “outdoor run between 5-7pm”

4. Training loop 

Through its interconnected layers artificial neurons, the ANN improves its user response via a continuous knowledge-production feedback process:

• Initialize weights - Each connection between neurons starts with a small random value or ratio. At this point, the network is like an infant and knows nothing.

• Forward pass  - Input data move through the network. Each neuron multiplies its inputs by their weights or relevant ratios, sums the result, applies an activation function, and produces an output.

• Backpropagation - adjusts the network’s internal weights/ratios to reduce communicative error in the next round.

• Explicit (e.g., giving a negative rating) and implicit (e.g., ignoring app recommendations) user feedback — both helps refine the network’s internal logic over time.

Over hundreds or thousands of iterations, the network becomes increasingly adept at algorithmic rhetorical invention, producing continuously refined messages tailored to a specific user. Each output responds to the need or problem that occasioned the algorithmic communication in the first place.

Paradoxes of Machine Learning

Machine learning often depends on harvesting immense amounts of personal data—social media footprints, search histories, behavioral logs, medical records, financial metadata, and more. This “data-rich” knowledge-production process powers the personalized tools and media experiences we now take for granted. Yet the same process that enables algorithmic convenience carries profound risks. Sensitive information can fall into the hands of malicious actors, or be quietly repurposed for profit-driven AI training, academic research, or targeted advertising. In each case, users lose both informed consent and a fair share of the benefits.

With the rise of ML, we confront several interwoven paradoxes that shape our contemporary digital lifeworld:

Autonomy Paradox 

We want machines to learn, adapt, and make judgments—but not in ways that drift beyond human expectations, institutional norms, or moral boundaries. This is a difficult balance to strike. 

Example: Self-driving cars must learn from millions of road scenarios. Yet when confronted with edge cases (e.g., unusual lighting, ambiguous road markings, or odd pedestrian behavior), the car may act unpredictably. The very autonomy that makes them powerful also makes them difficult to constrain.

Privacy Paradox 
Most people claim to value privacy, yet their everyday actions tell another story. We routinely grant apps access to microphones, location data, browsing history, photo libraries, and contact lists—often in exchange for trivial conveniences. 

Example: Millions of users share sensitive biometric data with fitness and period-tracking apps that may later sell or share this information with advertisers or data brokers. Likewise, those “Login with Facebook/Google” buttons function as gateways streamline access, but they also funnel enormous amounts of personal social media data back into AI training pipelines.

Transparency Paradox  
Machine-learning systems thrive on massive and diverse datasets. Greater transparency—open data, detailed logs, public benchmarks—helps researchers improve models. But transparency also exposes users to risk: data leaks, deanonymization, identity theft, and algorithmic exploitation. Meanwhile, the ML models themselves grow increasingly opaque. Hidden layers and self-evolving weight matrices create black-box algorithms that even their original human designers struggle to interpret. 

Examples: Some medical diagnostic AIs have shown outperform clinicians on benchmark tasks, but their reasoning is often inscrutable. When they fail, they may do so without clear explanation. Similarly, credit-scoring and job-screening algorithms also produce decisions that are extremely consequential, yet their internal logic is protected as proprietary, leaving individuals with no recourse to challenge or understand the outcome.

This second act of modern AI discourse marks a pivotal recognition: once intelligence is encoded into statistical systems, human supervision becomes partial, uncertain, and retrospective. We intervene after errors emerge, not before. The shift from procedural logic to learning-based inference signals a fundamental transformation in how agency, accountability, and meaning unfold in our sociotechnical world.

With the rise of machine learning, artificial intelligence shifted from executing rules to interpreting patterns, and from following instructions to generating its own provisional judgments. This transition introduced new forms of agency, new vulnerabilities, and new ethical terrains that require careful attention. In the next portion of What Is AI (Part 3), we will delineate how this adaptive foundation gave rise to generative systems capable not only of classifying the world but also of producing new artifacts—images, texts, sounds, and synthetic personas. From there, we turn to the ongoing debates surrounding the prospect of Artificial Superintelligence (ASI), where questions of autonomy, responsibility, and human futures become all the more urgent.

To be continued…

---

Footnotes

• ¹
  Alan M. Turing, “Computing Machinery and Intelligence,” Mind 59, no. 236 (1950): 433–460; see also Lawrence Livermore National Laboratory, “The Birth of Artificial Intelligence (AI) Research,” in Science & Technology – A Look Back, 2021.
• ²
  Gerard A. Hauser, “The Moral Vernacular of Human Rights Discourse,” Philosophy & Rhetoric 41, no. 4 (2008): 440–466.
• ³
  John A. Lynch, “The Rhetoric of Technology as a Rhetorical Technology,” Poroi 9, no. 1 (2013).
• ⁴
  Ronald Walter Greene, “Rhetoric and Capitalism: Rhetorical Agency as Communicative Labor,” Philosophy & Rhetoric 37, no. 3 (2004): 188–206.
• ⁵
  Bruce G. Buchanan and Reid G. Smith, “Fundamentals of Expert Systems,” Annual Review of Computer Science 3, no. 1 (1988): 23–58.
• ⁶
  古川康一 [Koichi Furukawa], “第五世代コンピュータ・プロジェクトの概観 (<特集>「第五世代コンピュータ」)” [Overview of the Fifth-Generation Computer Project], 人工知能 4, no. 3 (1989): 254–257.
• ⁷
  Lee McCauley, “The Frankenstein Complex and Asimov’s Three Laws,” in Association for the Advancement of Artificial Intelligence (AAAI Workshop Papers, 2007).
• ⁸
  Tony Myers, “The Postmodern Imaginary in William Gibson’s Neuromancer,” MFS Modern Fiction Studies 47, no. 4 (2001): 887–909.
• ⁹
  Simone Caroti, The Culture Series of Iain M. Banks: A Critical Introduction (Jefferson, NC: McFarland, 2015).