Skip to main content

Transcript Metaphor Audit: An LLM Experiment - The Last Invention

An image from the static

About This Analysis​

This framework is an experiment in using large language models to analyze discourse patterns in podcast transcripts. It involves two layers of AI processing:

  1. Transcription: Audio converted to text using OpenAI's whisper-large-v3-turbo model
  2. Analysis: Discourse patterns identified using Google's gemini-3.0-pro model

I cannot attest to the complete accuracy of either layer. The transcription may contain some errors since this is an LLM doing the transcription. Also, nothing here should be read as a claim about authorial intent, after all, these are interpretive outputs generated by probabilistic text systems, not scholarly conclusions.

The Source Material:The Last Invention​

The Last Invention is a podcast series that traces the history and trajectory of artificial intelligence research. I enjoyed it and it's a well-produced piece of audio journalism that provides one version of a coherent narrative framing for "how we got here.” From early cybernetics through the current era of large language models and existential risk discourse of AI “doomers.”

It's also, as it turns out, a master class in anthropomorphism.

The series features prominent voices from the AI safety and effective altruism communities, many of whom employ some vivid metaphorical language to describe AI systems: machines that "want," models that "lie," optimization processes that "seek" goals. This makes it an ideal candidate for Discourse Depot's analytical frameworks. The question isn't so much whether the speakers are wrong, it's how their linguistic choices shape what listeners can think about these technologies.

Outputs:​

The Framework: Five Analytical Tasks​

The Transcript Metaphor Audit prompt instructs Gemini to perform five distinct analytical tasks on the text of the transcripts. Each task operationalizes a different theoretical lens for examining AI discourse.

Task 1: Anthropomorphism & Metaphor Audit (25-30 instances)​

Identifies moments where speakers project human qualities onto AI systems. Each instance captures:

  • The conceptual metaphor at work (e.g., "AI as intentional agent," "AI as living organism")
  • What human quality is being projected (desire, understanding, deception, judgment)
  • The acknowledgment type: whether the anthropomorphism is presented directly, hedged with uncertainty, or so naturalized it becomes invisible

This task draws on cognitive linguistics (Lakoff, Gentner) to reveal how metaphorical framing shapes reasoning about AI capabilities and risks.

Task 2: Explanation Slip Audit (20-25 instances)​

Tracks moments where speakers slip between different types of explanations, using Robert Brown's Explanation Typology:

  • Mechanistic: How it works (causal processes)
  • Functional: What it does (system purposes)
  • Intentional: Why it wants to (goals and desires)
  • Dispositional: What it tends to do (behavioral patterns)

The "slip direction" is critical: do speakers drift from technical accuracy toward anthropomorphic framing, or occasionally self-correct back to mechanistic language? This reveals how explanatory frameworks compete within discourse.

Task 3: Agency & Causality Audit (15-20 instances)​

Identifies linguistic constructions that obscure human actors or misattribute agency:

  • Agentless passives: "The model was trained" (by whom?)
  • Nominalizations: "The training process" (who trained it?)
  • Displaced agency: "The AI decided" (software doesn't decide)
  • Reification: Treating abstractions as concrete actors
  • False symmetry: "Humans and AIs both..." (category error)

This task exposes how language can often render human labor, corporate decisions, and engineering choices invisible.

Task 4: Pattern Synthesis​

Synthesizes findings across Tasks 1-3 and 5 to identify:

  • Dominant Frames: The 3-5 most prevalent conceptual metaphors
  • Speaker Comparison: How different voices employ different rhetorical strategies
  • Explanation Patterns: Overall direction of explanation slips
  • Fallacy Summary: Which logic traps appear most frequently
  • Pedagogical Highlights: 5 instances with high teaching value

This section also includes Overall Analysis (the "so what") and Corpus Notes (how this transcript relates to broader AI discourse).

Task 5: AI Discourse Fallacy Audit (8-12 instances)​

Applies a custom 10-fallacy taxonomy to identify logical errors specific to AI discourse:

CategoryFallacyThe Logic Trap
I. LinguisticSynecdochePart (narrow competence) mistaken for whole (general intelligence)
Wishful MnemonicPsychological term (hallucinate, think) applied to statistical process
BiologicalIterative optimization described as organic growth/learning
II. AgencyHomunculusImagining a conscious agent inside making choices
Black Box LaunderingUsing system complexity to evade accountability
III. PoliticalObjectivityAssuming math/data is neutral, free of ideology
Ex NihiloIgnoring material costs (labor, energy, copyright)
IV. EpistemologicalRearview MirrorPredictive engine presented as capable of genuine novelty
ScalarIntelligence treated as simple volume (bigger = smarter)
Library of BabelInformation retrieval conflated with knowledge/truth
Each instance includes a Logic Trap (what the fallacy conceals) and a Correction (mechanistic reframing).

The Technical Components​

The Prompt​

A ~3,000 token system instruction that:

  • Assigns the model the role of "critical discourse analyst specializing in AI rhetoric"
  • Provides the theoretical foundations for each task (Brown's Typology, the fallacy taxonomy)
  • Specifies instance counts and structural requirements
  • Enforces negative constraints (no opinions, no external information, mechanistic language only)
  • Requires JSON-only output conforming to a strict schema

The JSON Schema​

A Gemini-compatible schema that enforces structured output with:

  • Type constraints: Strings, integers, arrays, enums
  • Required fields: Ensures no instance is incomplete
  • Enumerated values: acknowledgmentType must be "direct," "hedged," or "naturalized"; fallacyCode must match one of 10 codes
  • Nested structures: Each task's instances are arrays of objects with specific properties

The Processing Script (TranscriptMetaphorAuditProcessor.js)​

A Node.js processor (~1,100 lines) that:

  1. Loads JSON output from Gemini and optional thought summaries
  2. Collects metadata via interactive prompts (episode info, model settings, token counts)
  3. Validates instance counts against expected ranges
  4. Generates MDX output for Docusaurus with:
    • Frontmatter for site indexing
    • Tabbed About/Metadata sections
    • Overview and synthesis sections (summary first)
    • Collapsible instance details (evidence second)
    • Fallacy reference tables
  5. Outputs three files:
    • .mdx for the documentation site
    • .json archive with full data
    • .jsonl line for database import

Reading the Output​

The MDX output is structured for progressive disclosure:

  • Overview - The "so what": overall analysis and corpus context
  • Pattern Synthesis - Summary findings: dominant frames, speaker styles, pedagogical highlights
  • Task Instances - The evidence: every identified metaphor, slip, error, and fallacy

Each instance includes timestamps for audio verification. If something seems wrong, check the source.

Limitations​

  • Transcription errors: Whisper is imperfect, especially with multiple speakers
  • Speaker attribution: "Speaker 1" and "Speaker 15" may be the same person
  • Pattern-matching, not understanding: The model identifies patterns I described; it doesn't comprehend discourse
  • My biases encoded: The prompt reflects my theoretical commitments; different frameworks would surface different patterns
  • No ground truth: There's no objective answer to "how many anthropomorphisms are in this transcript"

Finally, this is exploratory methodology and I'm not flexing it as established science. The value is in surfacing patterns worth investigating. No pretense here of definitive claims.

Discourse Depot © 2026 by TD is licensed under CC BY-NC-SA 4.0