Shadow Lexicon

Supplement

Institutional timeline of semantic drift

The lexicon documents what was lost from each term. This supplement documents when, where, and by which institutions each tracked term was redirected. The drift was not accidental or uniform. It followed datable institutional decisions, named structural transitions, and identifiable competitive pressures. The eighth entry extends the method to the security dialect — bug, catalog, scanner, count.

  1. Safety
  2. Alignment
  3. Open
  4. Responsible
  5. Hallucination
  6. Transparency
  7. Frontier
  8. Bug

Drift path

Safety

MIRI / Singularity Institute, 2000 → OpenAI, 2015 → Industry standard, 2022–present

Open lexicon entry →

  1. 2000–2013

    Philosophical Redefinition

    Eliezer Yudkowsky founds the Singularity Institute (later MIRI) in 2000. "Safety" is detached from its regulatory meaning (OSHA, FDA, consumer protection law) and reattached to existential risk scenarios around hypothetical superintelligence. The term is now a technical, philosophical concern — not a consumer protection concept.

  2. 2014–2019

    Institutional Adoption

    Stuart Russell and Nick Bostrom bring "AI safety" into mainstream academic discourse. OpenAI is founded December 2015 with "safety" as a core mission term — but is simultaneously a research lab and a competitive player. Google acquires DeepMind in 2014; an AI ethics board is established as an acquisition condition, institutionalizing "safety" as a term controlled by the acquirer.

  3. 2019

    The Structural Turn

    OpenAI creates its capped-profit for-profit subsidiary (OpenAI LP) and takes a billion-dollar investment from Microsoft. "Safety" is now used by a company with investors, a commercial product roadmap, and liability concerns. The term is doing double work: describing the mission while protecting the company from scrutiny.

  4. 2022–present

    Consumer Deployment

    ChatGPT launches November 2022. "Safety" is operationalized as content filters, RLHF systems, and trust-and-safety teams managing reputational and regulatory risk. Jan Leike resigns from OpenAI's superalignment team in 2024, stating safety had taken "a backseat to shiny products" — the first senior insider confirmation of the drift in print.

Key institutional moments

  1. 2000MIRI founded — "safety" detached from regulatory meaning.
  2. 2014DeepMind acquired by Google — "safety" becomes a corporate governance term.
  3. 2015OpenAI founded — "safety" paired with commercial AI development for first time.
  4. 2019OpenAI LP created — "safety" used by a company with investor obligations.
  5. 2022ChatGPT launch — "safety" operationalized as content moderation at consumer scale.
  6. 2024Jan Leike resigns from OpenAI — first senior insider confirmation that safety was subordinated to product priorities.

Drift path

Alignment

MIRI naming, November 2014 → Stuart Russell coinage → OpenAI/Anthropic adoption, 2017–2022

Open lexicon entry →

  1. October–November 2014

    The Naming Moment

    Stuart Russell suggests the term "alignment" to MIRI in 2014. MIRI's internal documents show the switch from "Friendly AI" to "aligned AI" occurring between October 23 and November 20, 2014 — the first time it appears in "Aligning Superintelligence with Human Interests: A Technical Research Agenda." The original meaning is narrow and technical: correspondence between a superintelligent system's objectives and human interests.

  2. 2017–2022

    Expansion and Dilution

    OpenAI, Anthropic (founded 2021), and DeepMind adopt "alignment" as a field name. The term expands from "alignment against existential risk" to "alignment with developer-specified values." RLHF, popularized by OpenAI and Anthropic, operationalizes "alignment" as feedback from a small pool of paid raters. "Human values" becomes shorthand for contractor preferences.

  3. 2022–present

    Corporate Standardization

    Following ChatGPT's launch, "alignment" becomes standard corporate language. Anthropic publishes Constitutional AI (2022); OpenAI publishes alignment research; Google DeepMind has an alignment team. The term now describes whatever the company's RLHF process produces. The original question — aligned with whom? — has been answered by default: aligned with the developer's chosen feedback process, undisclosed to users.

  4. By 2023

    The Mutuality Problem

    The geometric root of "alignment" requires two parties. By 2023, industry "alignment" means one party (the developer) defines the standard and the system conforms. No external body verifies what values are encoded. No public disclosure of what the alignment process actually optimizes for. The word's built-in mutuality has been erased by a unilateral technical process presented as a relational achievement.

Key institutional moments

  1. 2014October–November: MIRI adopts the term — original meaning narrow and specific.
  2. 2017OpenAI begins using "alignment" in research.
  3. 2021Anthropic founded; alignment becomes core institutional language.
  4. 2022Anthropic's Constitutional AI paper — "alignment" operationalized as AI self-critique against a developer-written constitution.
  5. 2023Industry-wide adoption renders the term's original precision meaningless through ubiquity.

Drift path

Open

OpenAI naming, December 2015 → GPT-2 withholding, February 2019 → Meta Llama, 2023–present

Open lexicon entry →

  1. December 2015

    The Naming

    Elon Musk names the company "OpenAI" with explicit intention of open-source research and public sharing of findings. The founding statement promises to "publish papers, blog posts, or code" and share patents with the public. "Open" means: open collaboration, open publication, open access — the established technical meaning.

  2. February 2019

    The First Reversal

    OpenAI announces GPT-2 but withholds the full model, citing misuse risks. This is the first time a company named "OpenAI" explicitly declines to be open. The community notes the contradiction immediately. Musk later states: "OpenAI was created as an open source company — that's why I named it 'Open' AI — it's not what I intended at all." The staged release transforms the word from a commitment into a communications posture.

  3. 2023–present

    Industry Adoption of the Hollow Term

    Meta releases the Llama model family beginning in 2023, marketing them as "open source" or "open" while including usage restrictions, prohibiting certain commercial applications, and not disclosing training data. Meta's use of "open" is widely criticized by the actual open-source community (OSI) as a misuse of the term. Nevertheless, "open" Llama releases are reported as open source in mainstream press and regulatory submissions.

  4. 2023–2024

    Regulatory Exploitation

    In EU AI Act proceedings and US congressional testimony, "open" models are argued to deserve lighter regulatory treatment on the grounds that openness enables scrutiny and community oversight — the original meaning of the term. Companies with "open" models that are not open by any established technical standard benefit from regulatory carve-outs justified by a word that no longer means what the carve-outs assumed it meant.

Key institutional moments

  1. 2015December: OpenAI founded — "open" means open-source, public research, shared patents.
  2. 2019February: GPT-2 withheld — first documented reversal by OpenAI itself.
  3. 2019March: OpenAI creates for-profit subsidiary — institutional incentive against openness now structural.
  4. 2023Meta Llama releases — "open" applied to models with usage restrictions and undisclosed training data.
  5. 2024EU AI Act negotiations — "open" used to argue for regulatory exemptions. Musk v. OpenAI (N.D. Cal., 4:24-cv-04722) argues departure from the "open" founding commitment.

Drift path

Responsible

Microsoft/Google principles, 2018 → Meta RAI team, 2019–2023 → disbandment, November 2023 → brand without obligation, present

Open lexicon entry →

  1. 2018

    Corporate Institutionalization

    Microsoft publishes its "Responsible AI Principles" in 2018, establishing six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. These are voluntary, internal, unenforceable, and self-defined. Google simultaneously publishes its AI Principles following the Project Maven controversy. Both companies use "responsible" to describe voluntary commitments with no external enforcement mechanism.

  2. 2019–2023

    Team Formation and Reduction

    Meta creates its Responsible AI (RAI) team in 2019. By early 2023, Business Insider reports the team has been reduced to "a shell" with little autonomy, requiring lengthy stakeholder negotiations before implementing initiatives. The team's initiatives must pass through product stakeholders before implementation — structural subordination of responsibility to commercialization.

  3. November 2023

    Disbandment

    Meta disbands the RAI team entirely, redistributing members to the generative AI product team. Meta's spokesperson states the company will "continue to prioritize and invest in safe and responsible AI development." The word "responsible" survives the dissolution of the team it named.

  4. April 2024

    The Google Reorganization

    Google announces its Responsible AI teams in Research are being moved to Google DeepMind, "to be closer to where the models are built and scaled." The structural effect: the team previously positioned to evaluate models from outside is now inside the unit that builds them. "Responsible" oversight is reorganized into the product pipeline. The word remains; the independence is gone.

Key institutional moments

  1. 2018Microsoft and Google publish "Responsible AI" principles — voluntary, self-defined, unenforced.
  2. 2019Meta creates RAI team.
  3. 2023Early: Meta RAI reduced to "shell of a team" per Business Insider.
  4. 2023November: Meta RAI team disbanded — members moved to generative AI product team.
  5. 2024April: Google moves Responsible AI teams into DeepMind product structure.

Drift path

Hallucination

Computer vision literature, 2000 → NLP/LLM adoption, 2021–2022 → OpenAI consumer framing, 2022–2023

Open lexicon entry →

  1. 2000

    First Technical Use

    The term "hallucination" first appears in an AI technical context in a 2000 IEEE paper on face and gesture recognition, where it carries a positive meaning in computer vision (generating plausible image detail beyond available data). The clinical term — a sensory perception without external stimulus — is borrowed into a technical context where its pathological connotations are initially absent.

  2. 2021–2022

    NLP Adoption

    As LLMs scale and generate fluent but factually false text, "hallucination" migrates from computer vision into NLP. Research papers use it to describe models generating false information confidently. The term is convenient: it is in technical circulation, sounds sophisticated, and imports the clinical frame of involuntary perception — suggesting the system cannot help itself — rather than the engineering frame of foreseeable product defect.

  3. 2022–2023

    Consumer Normalization

    ChatGPT launches November 2022. OpenAI uses "hallucination" in consumer-facing communications. May 2023: OpenAI defines it as "a tendency to invent facts in moments of uncertainty." The Cambridge Dictionary updates its definition in 2023 to include the AI usage. Medical researchers (Smith, Greaves, and Panch, PLOS Digital Health, 2023) publish formal objections: applying a psychiatric term to a system failure anthropomorphizes the error and obscures developer accountability.

  4. 2023–present

    The Liability Argument

    Legal and regulatory scholars note the liability implications: "hallucination" frames false AI outputs as analogous to human perceptual disorder — involuntary, pathological, not attributable to design decision. Alternative terms proposed by critics include "confabulation" (plausible fabrication), "fabrication," and "factual error" — each of which more accurately describes a design artifact and more readily supports product liability analysis. Industry retains "hallucination."

Key institutional moments

  1. 2000First AI use of "hallucination" in IEEE computer vision paper — positive technical meaning.
  2. 2021Migration into NLP research — pathological framing imported.
  3. 2022November: ChatGPT launch — consumer-scale deployment.
  4. 2023May: OpenAI formally defines "hallucination" in consumer documentation.
  5. 2023Cambridge Dictionary updates to include AI usage. PLOS Digital Health publishes formal critique.

Drift path

Transparency

Tech company "Transparency Reports," 2010–2013 → Google AI Principles, 2018 → Model Cards, 2019 → Gebru firing, December 2020

Open lexicon entry →

  1. 2010–2013

    The Transparency Report Template

    Technology companies establish the practice of publishing "Transparency Reports" beginning in 2010–2013, following government data request controversies. These reports disclose government requests for user data — a narrow category chosen by the company. The format establishes a lasting template: a document called a "transparency" document that discloses selected information, creating the appearance of openness without the substance.

  2. 2018–2019

    AI-Specific Adoption

    Google publishes its AI Principles in 2018, listing transparency as a named value. Margaret Mitchell and Timnit Gebru at Google Brain publish the Model Cards framework in 2019 — a genuine transparency effort. The industry adopts the format while choosing what to disclose: capabilities and limitations, but typically not training data composition, fine-tuning procedures, or RLHF reward model specifications.

  3. December 2020

    The Timnit Gebru Firing

    Google fires Dr. Timnit Gebru, co-lead of its Ethical AI team, after she refuses to retract a paper specifically concerning transparency about training data and environmental costs of large models. The firing occurs at the same company that lists "transparency" as an AI principle. The institutional contradiction is documented publicly. Google continues using "transparency" as a core AI value term.

  4. 2023–present

    Structural Opacity as Industry Standard

    Every major model is released with a model card or system card — documents that describe what the company chooses to describe. Training data sources remain undisclosed or described only at category level. Reward model specifications are proprietary. Fine-tuning datasets are trade secrets. "Transparency" describes the existence of a disclosure document, not the completeness of disclosure.

Key institutional moments

  1. 2010Tech company "Transparency Reports" establish template for selective disclosure.
  2. 2018Google publishes AI Principles including transparency as a value.
  3. 2019Model Cards framework published by Mitchell and Gebru — genuine tool; industry adopts format without substance.
  4. 2020December: Google fires Timnit Gebru after she raises training data transparency concerns.
  5. 2023System cards become industry standard — selected disclosure while training data and RLHF specifications remain proprietary.

Drift path

Frontier

Anthropic founding documents, 2021 → industry prestige term, 2022 → self-regulatory body, 2023

Open lexicon entry →

  1. 2021–2022

    Term Introduction

    Anthropic, founded 2021 by former OpenAI researchers, introduces "frontier AI" and "frontier models" to describe the most capable systems at the leading edge of development. The geographical metaphor's governance implication is immediate: frontier territory is beyond settled governance, not yet subject to established rules. Anthropic explicitly positions itself as operating in this space.

  2. July 2023

    The Frontier Model Forum

    Anthropic, Google, Microsoft, and OpenAI found the "Frontier Model Forum" — a self-selected industry group that defines what "frontier" means for regulatory purposes, placing the companies most subject to potential regulation in the position of defining the regulatory category. The self-selection is structurally identical to an industry writing its own safety standards.

  3. 2023–present

    Regulatory Exploitation

    As AI regulation develops in the EU, UK, and US, "frontier AI" becomes a regulatory category. The UK's AI Safety Institute (est. 2023) focuses specifically on "frontier AI systems." Companies self-identifying as "frontier labs" use the term simultaneously to claim technical prestige and to argue for specialized regulatory treatment. The metaphor licenses operating at the edge of what governance frameworks have reached.

  4. 2024

    The Governance Circularity

    By 2024, "frontier AI" is used in government policy documents, legislative proposals, and regulatory frameworks — in language largely supplied by or shaped by the industry that coined and self-applied the term. The word performs the governance gap it describes: by calling themselves "frontier," labs claim to be at the edge where rules have not yet reached, while simultaneously participating in the processes that write those rules.

Key institutional moments

  1. 2021Anthropic founded — "frontier AI" enters institutional vocabulary.
  2. 2022Term adopted across OpenAI, Google DeepMind communications.
  3. 2023July: Frontier Model Forum founded by Anthropic, Google, Microsoft, OpenAI — self-selected companies define the category.
  4. 2023UK AI Safety Institute established with mandate focused on "frontier AI systems."
  5. 2024"Frontier" appears in EU AI Act negotiations and US policy documents in language shaped by industry participants.

Drift path

Bug

Engineering slang, 1878 → CVE catalog, 1999 → Codex Security count as vigilance, March 2026

Open lexicon entry →

  1. 1878–1947

    Defect as Invader

    Edison uses "bugs" for faults in apparatus. The word frames a defect as an unwelcome living thing in an otherwise sound machine. The 1947 Harvard Mark II moth becomes the origin story computing prefers — folklore that keeps the invader metaphor intact.

  2. 1999–2010s

    The Shared Catalog

    MITRE's CVE program (1999) makes openings common knowledge: one identifier, many defenders. "Vulnerability" still names a wound. The catalog's purpose is memory, not a league table of finders.

  3. 2010s–2025

    Scanner as Proof of Control

    Automated SAST/DAST and, later, LLM-assisted auditing convert openings into volume metrics. Severity bands, false-positive rates, and "alert noise" become the language of quality. The company that owns the detector also owns the labels.

  4. March 2026

    The Count as Vigilance

    OpenAI's Codex Security announcement reports 11,353 findings on 1.2 million commits (792 critical, 10,561 high, 14 CVEs) plus large drops in "alert noise" and false positives. The count is presented as evidence the system is watching. The word "bug" still cannot name a structural door, a trust assumption, or a person in the substrate.

Key institutional moments

  1. 1878Edison — "bugs" as defects in apparatus; invader metaphor set.
  2. 1947Harvard Mark II moth — computing's preferred origin story for "bug."
  3. 1999MITRE CVE catalog — openings named as common memory.
  4. 2026March: Codex Security — 11,353 findings presented as vigilance; "noise" and "false positive" used to license discard.