AI: Token or GPT?

In April 2025, Duolingo CEO Luis von Ahn wrote to employees: “Duolingo is going to be AI-first.” The same memo said Duolingo would “gradually stop using contractors to do work that AI can handle” and grant headcount only when a team could not “automate more of their work.” Shopify described Tobi Lütke’s internal memo with a similar rule: “reflexive AI usage is now a baseline expectation at Shopify.”¹

These memos are not vague aspirations. They are operational instructions. They tell employees how to interpret budgets, hiring, performance reviews, contractor use, product scope, and personal risk.

The same phrase carries different orders to different listeners. The CEO hears platform shift. The CFO hears cost discipline. The board hears technological seriousness. Investors hear margin expansion. The CTO hears architecture and data debt. Product hears features. HR hears workforce redesign. Vendors hear pipeline. Consultants hear transformation budget. Employees hear layoffs. Engineers hear that someone thinks code writes itself now.

The strategy problem is not that organizations are adopting AI. It is that the phrase AI has become a general purpose token whose replication velocity across memos, dashboards, OKRs, board decks, layoff narratives, and LLM outputs exceeds the rate at which organizations absorb the underlying technology into real work systems.

Put differently, AI-first often borrows against a future operating model. It lets today’s organization spend the credibility of tomorrow’s absorbed GPT-tech before the complementary infrastructure, expertise, governance, and workflow changes exist.

The contrast case is mechanism-oriented strategy: this model, with this data, inside this workflow, changes this behavior, creating this value, with these risks. The rest of the essay is about why that sentence is harder to circulate than AI-first.

AI may be a general purpose technology. It may eventually change knowledge work the way electricity changed factories and computing changed offices. The economic case is real enough to take seriously: LLMs are broadly applicable, improve quickly, and can spawn downstream tools and workflows.²

But AI is already something else inside firms.

It is a general purpose token.

Key terms

These labels keep the steps of the chain separate. They are not rival definitions of AI.

• GPT-tech — general purpose technology in the economic sense: broadly applicable, continuously improving, and generative of complementary innovations.
• LLM — Large Language Model: the model that processes and generates language.
• model-token — computational unit an LLM processes or generates.
• sign-token — compact cultural sign people can copy across media: AI, platform, innovation, transformation, data-driven, AI-first.
• strategy-token — sign-token that changes organizational decisions: decks, OKRs, dashboards, budgets, roadmaps, job titles, reorgs, rituals, vendor categories, performance reviews, and layoff narratives.
• AI-token — strategy-token version of AI when it stands in for the future, intelligence, efficiency, innovation, risk, transformation, inevitability, cost discipline, or workforce rationale.
• selfish token — sign that replicates more easily than the mechanism it once named. Selfish is analytical, not moralizing.
• token phenotype — visible artifacts or behaviors a token induces: posts, memos, dashboards, benchmarks, layoffs, job titles, policies, rebuttals, and rituals.
• token host — person, organization, platform, model, or institution that repeats, modifies, resists, or turns the token into rules and work.

The terms

A language model works with tokens: fragments of text. A company works with another kind of token: a short phrase people can copy into a memo, dashboard, job title, budget request, vendor pitch, or layoff story.

That is the chain this essay follows. An LLM generates text. People turn that text into memos, posts, policy drafts, strategy decks, product copy, risk frameworks, governance language, investment theses, skeptical essays, and layoff rationales. Some of those artifacts make a phrase like AI-first easier to copy. A company can then turn the phrase into hiring rules, roadmap changes, procurement categories, performance reviews, and cost-cutting stories.

The terms below keep those steps separate. They are labels for different parts of the chain, not rival definitions of AI.

GPT-tech means general purpose technology in the economic sense: a technology with broad application across many sectors, ongoing technical improvement, and the ability to generate complementary innovations. The name avoids conflating that literature with the model-architecture acronym: it does not mean generative pretrained transformer.

LLM means Large Language Model: the model that processes and generates language.

A model-token is the computational unit an LLM processes or generates.

A sign-token is a compact cultural sign: AI, platform, innovation, transformation, data-driven, customer obsession, AI-first.

A strategy-token is a sign-token that changes decisions inside an organization. It shows up in decks, OKRs, dashboards, budgets, roadmaps, job titles, reorgs, rituals, vendor categories, performance reviews, and layoff narratives.

The AI-token is the strategy-token AI when it means the future, intelligence, efficiency, innovation, risk, transformation, inevitability, cost discipline, or workforce rationale.

A selfish token is a sign that replicates better than the concept or mechanism it once represented. Selfish is analytical, not moral: it names differential ease of replication across channels, independent of whether the repetition serves truth or value creation. AI-first becomes selfish when it travels farther and faster than the workflow change it was supposed to name.

A token phenotype is the visible artifact or behavior induced by a token: posts, memos, dashboards, benchmarks, layoffs, job titles, policies, rebuttals, and rituals.

A token host is any person, organization, platform, model, or institution that repeats, changes, resists, or turns the token into rules and work.

AI can be real and still produce the language companies use to sell it, govern it, mock it, resist it, and justify decisions around it.

The token arrives before the mechanism

A mechanism-oriented AI claim sounds like this:

We believe an AI system can reduce enterprise renewal risk by synthesizing support tickets, usage drops, CRM notes, contract terms, and meeting history into account-specific intervention prompts ninety days before renewal. This creates value only if account managers trust the output, act earlier, and change the renewal conversation before the customer has already decided to churn.

A mechanism-oriented claim names the user, the data, the behavior change, the economic claim, and the ways the plan can fail.

AI-first drops all of that. It fits in a slide title, memo subject, OKR, dashboard, keynote, budget request, investor answer, or layoff note. A person, firm, platform, or model can repeat it without specifying the workflow.

A GPT-tech transforms the world slowly: infrastructure, complements, standards, training, trust, governance, workflow redesign, and institutional learning. People spread the AI-token through feeds, keynotes, demos, dashboards, podcasts, analyst reports, boardrooms, vendor pitches, investor calls, strategy decks, Slack threads, and layoff memos before the underlying mechanism has been understood.

The central error of the AI era is treating visible artifacts of work as evidence that the slow work of absorption has happened. AI is real; the shortcut is the error.

A general purpose technology has to be absorbed. A general purpose token only has to be repeated.

Real AI capability moves on two clocks: the slower GPT-tech path of infrastructure, complements, workflow, and trust; and the faster AI-token path of memos, dashboards, markets, and workforce narratives.

The channel changes what gets repeated

The same AI claim changes as it moves through LinkedIn, Twitter/X, dashboards, board decks, investor calls, procurement categories, layoff memos, and LLM outputs. Each channel keeps the part it can use. LinkedIn keeps the career signal. Twitter/X keeps the fight. A dashboard keeps the number. A board deck keeps the category. A layoff memo keeps the justification. An LLM repeats common wording.

The workflow gets dropped because it is expensive to carry. A mechanism-oriented AI claim needs the model, the data, the user, the failure mode, the review path, the integration burden, the operational metric, and the behavior change. AI-first needs two words.

That is the media-theory premise in operational terms: channels do not merely carry an AI claim; they select which version survives. People, platforms, firms, and models copy signs that fit available media. AI fits the slide title, dashboard, OKR, investor call, LinkedIn post, Twitter/X dunk, vendor category, all-hands theme, layoff memo, and LLM-generated press release. The workflow-specific mechanism does not.³

The channel also changes what organizations accept as proof.⁴ A demo can pass for transformation, a benchmark can pass for intelligence, an adoption dashboard can pass for value, a generated artifact can pass for work replacement, and a CEO’s AI-first language can pass for strategy.⁵

Adoption has two clocks.⁶ One clock is slow: memory, apprenticeship, standards, local judgment, governance, trust, and workflow. The other clock is fast: administration, command, markets, dashboards, feeds, board decks, and procurement categories. AI as GPT-tech needs slow absorption. The AI-token rides fast circulation. That is how firms build AI dashboards, governance offices, savings targets, and workforce plans before the work itself has changed.

Some AI talk is participation, not information transfer.⁷ We are AI-first marks modernity. AI will replace you marks inevitability. AI is slop marks resistance. The claims disagree, but each keeps AI at the center. On LinkedIn and Twitter/X, this becomes call-and-response: AI-first, AI-native, agents are coming, learn AI or be replaced, AI slop, do more with less, the future of work.⁸

The same structure appears outside corporate management. Policy, regulation, academia, and critical AI communities also produce strategy-tokens: AI safety, alignment, responsible AI, frontier models, open weights, human-centered AI. Some name real work. Some become category badges. The point is not that managers are uniquely confused. Contemporary media environments select for portable signs before they select for mechanisms.

After circulation succeeds, people and systems use the sign to sort firms and workers into future-ready or obsolete, AI-native or legacy, efficient or bloated.⁹ The representation can precede the territory: dashboards, governance offices, savings targets, and workforce plans appear before the work changes.¹⁰ Infrastructure makes repetition cheap: LLM interfaces, Slack, PowerPoint, OKR tools, benchmark sites, procurement categories, and LinkedIn.¹¹ Speed compounds the error: the demo outruns the implementation plan, the benchmark outruns the evaluation culture, and the layoff headline outruns careful attribution.¹²

How channels compress AI claims¹³

Channel	What survives	What gets stripped
Slide title	Category	Workflow
Dashboard	Number	Judgment and review burden
Board deck	Strategic narrative	Implementation detail
Investor call	Future-facing signal	Causal uncertainty
Layoff memo	Justification	Task-level evidence
LinkedIn	Career signal	Mechanism
Twitter/X	Conflict and affiliation	Nuance
LLM output	Familiar phrase	Local context

The medium is part of the mechanism. People, firms, and LLMs preserve the short word and discard the workflow. AI survives the media chain. The mechanism is what the media chain strips away.

Token-formation sequence

real capability
  -> visible artifact
  -> portable sign
  -> institutional translation
  -> metric / target / budget / narrative
  -> changed behavior

The rest of the essay follows that map through productivity evidence, visible artifacts, layoff narratives, strategy language, dashboards, and proxy work.

The technology is real, and the effects are uneven

This media compression has measurable consequences inside real work systems, as the productivity studies below illustrate.

The AI-token borrows credibility from real AI capability.

LLMs already produce measurable gains in some settings, but the evidence does not show one simple productivity curve. It shows task-specific mechanisms.

Brynjolfsson, Li, and Raymond studied a generative-AI assistant rolled out to customer-support agents and found a roughly 14 percent average productivity gain, with much larger gains for novice and lower-skilled workers.¹⁴

Peng, Kalliamvakou, Cihon, and Demirer found that developers using GitHub Copilot completed a bounded JavaScript task 55.8 percent faster than the control group.¹⁵

A Google randomized controlled trial found about a 21 percent reduction in task time on a complex enterprise-grade task, while warning that the lab result may not generalize across ecosystems, tools, and time.¹⁶

The gains are uneven in exactly the way a mechanism-oriented person should expect. METR’s 2025 randomized controlled trial found that experienced open-source developers working in familiar, mature repositories took 19 percent longer when early-2025 AI tools were allowed, even though they expected beforehand that AI would speed them up and believed afterward that it had.¹⁷

A later study of open-source projects found that AI-assisted programming increased activity from less-experienced developers but also increased rework and maintenance burden for core developers.¹⁸

A 2026 Microsoft-scale study of command-line coding agents found that adopters merged roughly 24 percent more pull requests, while explicitly treating merged PRs as a proxy for output, not final value.¹⁹

These findings do not refute one another. They show that AI helps developers and AI slows developers can both be true when the unit is not coding but a work system: task type, codebase maturity, developer expertise, review burden, architecture, test coverage, quality bar, organizational context, and error tolerance.

The shorthand claim says:

AI writes code.

The mechanism asks:

• What part of software engineering got cheaper?
• What burden moved downstream?
• Who reviews the output?
• What does the metric count?
• What has become more important because output is now cheaper?

Those are different questions.

The AI-token is powerful because the referent is real. There are real support gains, coding assistants, summarizers, copilots, automation tools, and workflow changes. The evidence gives the token legitimacy. The problem starts when that legitimacy travels farther than the mechanism.

A support productivity gain becomes:

White-collar work is ending.

A code-generation demo becomes:

Engineers are obsolete.

A chatbot pilot becomes:

Agentic transformation.

A cost-cutting program becomes:

AI-enabled restructuring.

A board anxiety becomes:

AI strategy.

The AI-token does not have to be causally precise. It has to be useful to repeat.

Visible artifacts and hidden work

An AI output is evidence that something happened, not proof that the underlying work has been replaced. Abbott’s Flatland gives a useful image: a sphere passing through a two-dimensional world appears to the Flatlanders as a changing circle: point, widening circle, narrowing circle, point. The observations are not wrong. The instruments are not broken. The error is dimensional. They mistake the visible trace for the object.²⁰

Modern organizations do this constantly.

They see the visible artifact:

prompt -> code
prompt -> memo
prompt -> slide
prompt -> support response
prompt -> analysis

Then they infer that they have seen the work.

But the visible artifact is often only the output.

A legal brief is not lawyering. A diagnosis is not medicine. A strategy deck is not strategy. A line of code is not engineering. A support response is not customer success. A generated summary is not institutional understanding.

The generator is the work behind the artifact: judgment, context, accountability, tacit knowledge, timing, trust, integration, review, maintenance, customer reality, and consequence.

AI tools and AI rhetoric make the output spectacular. Code appears. Text appears. Slides appear. Images appear. Answers appear.

The artifact appears, so the work seems replaced.

That is the mistake the Flatland analogy names: a visible trace gets mistaken for the system that produced it.

The better question is:

What work outside the artifact made it valuable, and who performs that work now?

For code, the work outside the artifact includes understanding the system, choosing what should exist, preserving invariants, integrating with other services, reviewing edge cases, maintaining the result, and owning failure. For a support response, it includes customer context, escalation judgment, policy authority, empathy, and trust. For a diagnosis, it includes examination, history, risk tolerance, liability, and follow-up.

AI can remove some of that work. It can also move the work to data preparation, review, exception handling, customers, senior employees, maintenance, security, quality, or trust repair.

AI instead of engineers exposes the managerial tendency to identify a job with its most visible artifact.

Engineers produce code, but engineering is not code production. Engineering is the construction and stewardship of working systems under constraints. It includes deciding what should exist, understanding existing systems, preserving invariants, designing abstractions, integrating services, debugging failures, securing boundaries, testing assumptions, migrating data, managing tradeoffs, and remaining accountable when the system breaks.

If model output makes code cheaper while system understanding, review, integration, and accountability become the bottlenecks, AI has not eliminated engineering. It has changed where the engineering labor sits.

The visible artifact is not the generator. Tokenized management keeps losing that distinction.

LLMs help manufacture AI discourse

Steam, electricity, computers, the internet, blockchain, and the cloud all became symbols of the future as well as tools.

AI adds a reflexive loop: LLMs help write the discourse that inflates, contests, governs, sells, mocks, and explains AI.

An LLM produces model-tokens. Those tokens become posts, memos, slide titles, policy drafts, job descriptions, product pages, risk frameworks, benchmark summaries, rebuttals, strategy documents, and layoff scripts. Some praise AI. Some condemn it. Some regulate it. Some sell it. Some warn against it. Some mock it.

At the level of belief, those artifacts disagree. At the level of replication, they perform the same act: they make AI the object everyone must address.

A pro-AI post says:

AI will transform everything.

An anti-AI post says:

AI is destroying work.

A skeptical post says:

AI is overhyped.

A governance memo says:

AI requires responsible adoption.

A layoff memo says:

AI enables us to operate more efficiently.

A vendor white paper says:

AI unlocks enterprise value.

Memetic replication does not require agreement. It requires hosts and channels. A supporter can repeat the sign. A critic can repeat it. A regulator can repeat it. A layoff memo can repeat it. A joke can repeat it. An LLM-generated explanation of why AI hype is overblown can repeat it.

The loop is:

LLMs generate model-tokens
  -> model-tokens become discourse artifacts
  -> discourse artifacts circulate as sign-tokens
  -> some sign-tokens become strategy-tokens
  -> firms treat the AI-token as institutionally necessary
  -> firms generate more AI discourse
  -> LLMs assist that discourse
  -> more AI-token replication

AI produces text and accelerates the production of the symbolic units by which organizations misunderstand AI.

AI as workforce narrative

Layoff discourse shows why executives reach for the AI-token.

Companies lay off workers for many reasons: overhiring, margin pressure, failed bets, interest rates, restructuring, outsourcing, capital reallocation, investor demands, automation, management fashion, and genuine technological substitution.

The word AI can package that causal mess into a story that boards and investors know how to reward: the company is reallocating work to the future instead of explaining overhiring, missed bets, or margin pressure.

The story says:

This is not merely cost-cutting. This is transformation.

This is not managerial failure. This is technological inevitability.

This is not weakness. This is discipline.

That is the narrative advantage. The company appears future-facing rather than cornered.

Some AI-attributed layoffs reflect real automation. Some roles are being redesigned. Some tasks are becoming cheaper. Some companies will need fewer people in certain functions.

Management can use the AI-token to turn ambiguous causality into a future-facing story.

By mid-2026, Challenger, Gray & Christmas data had AI as a leading employer-cited reason for U.S. job cuts, while commentators continued to dispute how much of that attribution reflected actual substitution versus broader restructuring and cost pressure.²¹ The distinction changes what the claim proves. Employer-cited reasons are not the same as causal proof.

Even when genuine substitution occurs, the AI-token still compresses the questions that matter: which tasks changed, which workers were affected, what downstream review or maintenance burden appeared, and whether headcount reduced has replaced the value question: did durable capability hold or improve?

A mechanism-oriented layoff claim would say:

These tasks have been automated or compressed; these workflows have been redesigned; this review burden remains; these quality risks are acceptable; these roles are changing; this is the observed productivity effect; these are the costs we expect to appear downstream.

A tokenized layoff claim says:

AI enables us to do more with less.

The first is a causal account. The second is a high-level coordination phrase.

It may still be true in a particular case. It is not yet specified enough to know.

AI as strategy medium

A company saying AI is our strategy is like saying electricity is our strategy or software is our strategy.

It may identify the technological epoch. It does not identify an advantage.

A GPT-tech becomes strategic only when a firm finds a specific mechanism of value: a workflow it can redesign, a constraint it can remove, a capability it can compound, a cost structure it can change, a customer experience it can improve, or a defensible learning loop it can build.

AI-first does not answer that.

It does not tell us whether AI is being used for product differentiation, support automation, internal knowledge retrieval, decision support, developer productivity, customer self-service, compliance monitoring, personalization, risk detection, labor substitution, data monetization, or investor narrative.

Those are different strategies.

AI-first hides the difference because every function can map it to its own agenda. Product can hear feature roadmap. Finance can hear cost reduction. HR can hear workforce redesign. Engineering can hear tooling. Investors can hear discipline.

Mechanism questions the slogan avoids

• What specific work changes?
• Whose behavior changes?
• What data is required?
• Who owns that data?
• What is the trust model?
• What must humans review?
• What error rate is acceptable?
• What error would be catastrophic?
• What cost moves downstream?
• What metric would fool us?
• What customer value is created?
• What capability becomes defensible?
• What should we stop doing?
• What would prove the thesis wrong?

Without those answers, AI strategy is identity-oriented strategy.

Identity-oriented strategy lets the organization say who it is:

We are AI-first.
We are innovative.
We are transforming.
We are data-driven.
We are a platform company.
We are customer-obsessed.

Mechanism-oriented strategy specifies how value is generated:

this model
with this data
inside this workflow
used by this role
under this review process
changes this behavior
creating this value
with these risks
and these failure signals

The first form is selected for internal replication. The second is selected for contact with reality.

Corporate media favors the first.

PowerPoint selects for pillars, symmetry, and executive confidence. OKRs select for measurable nouns. Dashboards select for recurring numerical objects. Slack selects for fast recognition. Board decks select for coherent narratives. Procurement selects for categories. Performance reviews select for participation signals. Investor calls select for concise futures.

A mechanism enters that system and gets shortened.

Original mechanism:

AI can reduce low-complexity support resolution time if it has access to historical tickets, product documentation, account context, escalation rules, and a human review loop; value appears only if repeat contact declines without degrading trust or hiding unresolved customer problems.

Deck version:

AI-powered support transformation

OKR version:

Deflect 30% of tickets with AI

Dashboard version:

AI deflection rate

All-hands version:

AI helps us serve customers faster

Layoff version:

AI enables leaner operations

At every step, the statement becomes more transmissible and less faithful.

That is how the claim changes as it travels.

Measurement, value capture, and the AI dashboard

Once AI becomes a dashboard category, the problem deepens.

A dashboard measure can stop describing work and start steering it. Goodhart’s Law is usually summarized as: when a measure becomes a target, it ceases to be a good measure.²²

AI strategy can create Goodhart traps at each translation step.

If AI adoption becomes the target, employees will use AI whether or not it improves work.

If AI use cases launched becomes the target, teams will launch use cases whether or not they touch meaningful constraints.

If tickets deflected becomes the target, customers may be prevented from reaching humans while unresolved frustration disappears from the dashboard.

If lines of code generated becomes the target, code volume rises while maintainability falls.

If AI-driven savings becomes the target, teams can relabel ordinary cost cuts as AI even when the operating model has not changed.

A measure can corrupt behavior. A score can also train people and institutions to care about the score instead of the goal.

C. Thi Nguyen calls this value capture: simplified scores, rankings, or metrics overtake the richer values they were supposed to serve.²³

That is what happens when managers turn AI into a dashboard category.

The most literal form is the model-token leaderboard: ranking employees by how many computational tokens their AI tools consume. In April 2026, The Wall Street Journal reported that an internal Meta dashboard ranked employees by individual model-token usage and assigned status titles such as Token Legend. In May 2026, Business Insider reported that Amazon deprecated an informal internal KiroRank leaderboard after it encouraged some employees to perform tasks that did not necessarily solve problems in order to climb the ranking. Indeed’s CIO offered the counterdiscipline: track token use in the background, but keep the visible management system closer to outcomes.²⁴

A model-token leaderboard is media theory with a bill attached. The model-token starts as a computational and pricing unit. The dashboard turns it into a visible social object. Ranking turns that object into status. Management attention turns status into an incentive. At that point, the model-token has become a strategy-token: not a measure of value created, but a portable proof of participation. The institution can see the meter running, so people learn to make the meter run.

The original value might be:

Help customers resolve problems with less friction.

The category becomes:

AI support automation.

The dashboard turns that into:

Ticket deflection rate.

The OKR turns that into:

Deflect 30% of tickets with AI.

The organization optimizes for the number.

At first, the number is supposed to represent customer value. Eventually the number becomes the value. Teams celebrate deflection even if customers are angrier. Executives see lower ticket volume and call it efficiency. Support leaders are rewarded for automation rates. Product teams deprioritize the harder work of eliminating the underlying customer confusion.

The metric has captured the value.

That is the institutional sequence:

sign of technological possibility
  -> metric
  -> target
  -> management system
  -> changed behavior
  -> changed institutional desire

AI adoption becomes transformation. AI deflection becomes customer service. AI-generated code becomes engineering productivity. AI-enabled savings becomes strategic progress. AI use cases launched becomes innovation. AI literacy training completed becomes capability.

The original question was:

Did AI change the work in a way that creates durable value?

The captured question becomes:

Did the AI number move?

It is institutional value replacement.

High-resolution proxy work

The crude error is easy to spot:

AI writes code, so engineers are obsolete.

The subtler error looks rigorous:

We have analyzed AI engineering productivity by task type, code acceptance rate, pull-request cycle time, codebase maturity, review burden, quality incidents, defect density, developer sentiment, and tool adoption cohort.

Those measures can improve the diagnosis. They still do not answer the dimensional question: which part of engineering moved, which bottleneck changed, which new burden appeared, and who remains accountable when the system breaks?

More resolution inside the proxy is not the same as understanding the work behind it.

Expertise can create its own blind spot. In Gell-Mann amnesia, you notice media errors in a domain you know, then forget that lesson when reading about domains you do not know.²⁵ Here, the pattern reverses: a formal model, credentialed committee, or high-resolution dashboard can reduce the work to a proxy, such as code accepted, tickets deflected, adoption rate, or cycle time, and make that reduction look rigorous.

Naive simplification is easy to challenge. Sophisticated simplification arrives with controls, cohorts, taxonomies, and confidence.

A company can build an AI governance framework, maturity model, use-case taxonomy, adoption dashboard, ROI calculator, vendor scorecard, and workforce plan and still fail to answer:

What mechanism creates value?

The error is not speed. The error is reasoning about the proxy instead of the work. A committee can spend six months on adoption categories, dashboards, benchmarks, risk registers, and workforce scenarios and still never ask which workflow changed, which burden moved, or who remains accountable when the system breaks.

Recognition asks:

• What is this like?
• What category does it belong to?
• What familiar pattern is this?
• What label lets us coordinate?
• What does this signal about who we are?

Mechanism reconstruction asks:

• What generated this?
• What work outside the artifact does it depend on?
• What superficially similar cases have different causes?
• What would make the anomaly regular?
• What does this sign-token collapse?
• What would prove us wrong?

People can use the AI-token to recognize the future before they understand the work.

GPT-tech and complements

AI may genuinely become a GPT-tech, and that possibility gives the AI-token plausibility.

General purpose technologies do not deliver their full value through surface adoption. They require complements: new workflows, skills, governance, infrastructure, and intangible capital. That is the mechanism behind the AI productivity paradox described by Brynjolfsson, Rock, and Syverson: impressive AI capabilities can coexist with weak measured productivity because GPT-techs need complementary reconstruction before their full effects appear.²⁶

Implementation lag is the mechanism. The shortcut story promises GPT-tech returns without the complementary reconstruction.

This is a form of strategic debt. Borrowing is not automatically irrational; a firm has to invest before the full return appears. The error is spending the social and financial returns as if absorption has already happened: reducing headcount, flattening expertise, declaring productivity, and rewarding adoption before the organization knows which human knowledge, data plumbing, review paths, and maintenance work make the system valuable.

When firms spend the AI-token before building the complements, my prediction is a delayed complement cycle:

• Substitution story. Firms cut roles, freeze hiring, and count AI adoption as capability. Metrics make participation visible before the operating capability exists.
• Displaced work. Review, exceptions, customer repair, and maintenance move to people outside the automation target. Model output makes artifacts cheaper before it makes judgment, accountability, or integration cheaper.
• Rebundled roles. Work collects around evaluation, data cleanup, integration, governance, security, and system stewardship. The complements skipped in the strategy still have to be done.
• Expertise repricing. Some senior expertise becomes more valuable. AI systems need experienced people to define quality, catch edge cases, mentor reviewers, and improve tools.

The result is uneven labor change: some roles disappear, some move, some get renamed, and some require expertise the substitution story priced as surplus. Ford’s 2026 quality reset makes the complement problem concrete: the company said AI and automation were not enough on their own and hired, promoted, or brought back about 350 experienced technical specialists to mentor staff, lead design reviews, and improve automated quality tools.²⁷

It suggests:

install tool -> reduce labor
ship feature -> become innovative
launch pilots -> transform organization
adopt coding assistant -> replace engineers
deploy chatbot -> cut support
announce AI-first -> become future-ready

A real GPT-tech path looks more like:

technical capability
  -> complementary infrastructure
  -> workflow redesign
  -> skill reallocation
  -> trust and governance
  -> new operating model
  -> changed production function
  -> delayed, uneven gains

The shortcut story turns that into:

AI

That shortcut is the source of both speed and error.

The authenticity predictions

Expect authenticity polarization outside formal AI strategy.

When people say that sounds like AI, sometimes they mean a real defect: generic structure, unsupported claims, padded sentences, fake precision, or missing judgment. Sometimes they do something else: they attack the origin of the work instead of the work. That is an ad hominem version of the AI-token. The burden shifts from criticism to authentication: prove a machine did not touch this.

On the anti-AI side, expect:

• Human-written, no AI, handmade, and real artist become claims about craft, care, and accountability. Sometimes the claim will be earned. Sometimes it will be packaging.
• Writers, students, applicants, and creators keep drafts, revision history, notes, uneven rhythm, personal detail, or visible drafts to prove human involvement.
• Schools, employers, publishers, and platforms adopt detection rituals because they need an enforceable rule. The rule can punish style instead of cheating.
• Readers ask was AI involved? and stop before asking is this true, useful, original, or well made?
• AI slop remains useful when it names bad work: false facts, generic structure, missing judgment, broken code, empty words. It fails when it treats authorship as the reason the work is bad. Slop is slop: human and machine systems can both produce it.

On the pro-AI side, expect the matching ad hominem:

• You refuse AI becomes a way to attack someone’s competence instead of answering their concern.
• AI-native, AI-powered, and built with agents signal competence, speed, and modernity even when no one has shown the work is good.
• Visible human labor gets dismissed as nostalgia, preciousness, or inefficiency.
• Fast output gets treated as evidence of capability, even when review, integration, and accountability have merely moved elsewhere.

One side uses AI to question authenticity. The other uses AI to question competence. Both skip the mechanism-level question: what was made, how was it made, what judgment shaped it, and whether the result is any good.

How to use the analysis

Use the analytical lens in this essay when people use a claim about AI to end an argument too quickly.

Write down the claim. Then separate three things: what the person is saying, what the claim helps them do, and what evidence would make the claim true.

Claim	What the claim can help people do	Better next question
That sounds like AI	Rejecting work by attacking authorship.	What is actually wrong: false claim, generic structure, unsupported evidence, missing judgment, bad style, or no defect at all?
AI-first	Claiming future-readiness before the operating model has changed.	Which workflow changed, and what complements now exist: data access, review, governance, skills, maintenance, and trust?
AI slop	Naming low-quality generated work, or dismissing abundant work without critique.	What failure appeared: false facts, dead code, boilerplate, duplicated logic, missing context, or review burden?
You refuse AI	Recasting caution as incompetence.	What constraint would AI actually change, and what risk is the skeptic preserving?
Tokenmaxxing	Treating model usage as participation proof.	Did the work improve, or did the meter run?
Human-written / no AI	Asking authorship to prove quality.	What quality, judgment, accountability, or care does human authorship add here?

Once you know the claim, where it is being repeated, and what people are doing with it, the predictions become practical. If managers put the claim in a dashboard, expect rankings, targets, and gaming. If employers use it in hiring, expect status claims and screening rituals. If schools use it in grading, expect disclosure rules and draft-history demands. If people use it in feeds, expect counter-slogans. If executives use it in a layoff memo, expect causal compression.

Do not stop at the claim. Ask what follows from it. If the question is was AI involved?, ask what quality failure or value gain follows from that fact. If the question is are we AI-first?, ask which constraint changed. If the question is how many tokens did we use?, ask what value the work created and what burden moved somewhere else.

Mechanism-oriented AI strategy

Mechanism-oriented AI strategy starts with a workflow or job-to-be-done, not with the word AI.

Pick a place where value is created or lost: a renewal motion, a support queue, a code review path, a compliance review, a lab workflow, a procurement process, a claims process, a sales handoff, a customer onboarding path, or an internal knowledge search. Then ask whether AI changes the constraint.

For each candidate use case, skip the generic question:

Can AI do this?

Ask:

Does AI change the economics of this workflow after accounting for data access, trust, review, exception handling, integration, compliance, maintenance, and behavioral adoption?

Slogans drop those costs. Strategy keeps them in the decision.

Mechanism-oriented strategy separates cases that ordinary corporate speech collapses into AI:

• AI as artifact generator.
• AI as knowledge interface.
• AI as workflow compressor.
• AI as decision support.
• AI as customer self-service.
• AI as support triage.
• AI as coding assistant.
• AI as quality monitor.
• AI as compliance reviewer.
• AI as personalization engine.
• AI as scientific accelerator.
• AI as symbolic strategy.
• AI as workforce narrative.

Those are not interchangeable uses of the same word.

A company can have excellent AI artifact generation and weak AI strategy. It can have high adoption and low value. It can have many pilots and no operating change. It can reduce headcount and increase review burden. It can generate more code and ship slower. It can deflect tickets and degrade trust. It can impress the board and confuse customers.

The diagnostic is simple:

What changed in the work?

If the answer is people are using AI, the strategy is still token-level.

If the answer is we launched twenty use cases, the strategy is still token-level.

If the answer is we reduced headcount, the strategy is still not necessarily mechanism-level.

A mechanism-level answer sounds like this:

This role used to spend six hours collecting, reconciling, and summarizing fragmented account information before renewal. The AI system now performs the first-pass synthesis in eight minutes. The account manager spends twenty minutes reviewing, correcting, and adding judgment. The renewal conversation begins thirty days earlier. Risk interventions happen before procurement hardens. Retention improved in the target cohort, and we are watching for false confidence, stale data, and review fatigue.

That answer works because it names the old workflow, the new workflow, the behavioral change, the metric, and the risks.

Operating inside the selection environment

The AI-token will keep circulating. That is not a temporary distortion. It is a stable outcome of the media systems firms use to allocate attention and resources.

Large organizations cannot replace token circulation with full mechanism description. No large organization can coordinate every decision at full causal resolution. They need coupling: fast token media and slow mechanism media need interfaces.

A token dashboard counts:

• AI adoption.
• AI use cases launched.
• Model-tokens consumed.
• Tickets deflected.
• Code generated.
• AI-assisted savings.

A mechanism dashboard checks:

• Review burden.
• Repeat contact.
• Escalation rate.
• Rework.
• Defect rate.
• Maintenance load.
• Decision latency.
• Customer trust.
• Workflow cycle time.

The first list has a job: allocate attention. But it has to point to the second list before it earns budget, headcount, or authority. AI adoption can tell you that people are trying the tool. It does not tell you whether review burden, rework, customer trust, or workflow cycle time improved.

The operating discipline is to let token-level coordination raise the question, then require mechanism-level learning to answer it.

The sequence

The mechanism is a sequence:

real AI capability
  -> media-visible artifact
  -> AI-token formation
  -> accelerated circulation
  -> LLM-assisted discourse
  -> ritual repetition
  -> corporate media translation
  -> strategy-token
  -> dashboard / OKR / budget / roadmap / layoff narrative
  -> Goodhart: proxy becomes target
  -> value capture: proxy becomes value
  -> institutional hardening
  -> Flatland: projection mistaken for generator
  -> anomalies reclassified as execution failure
  -> AI-token replication

A real capability produces a visible artifact: code, text, support response, image, summary, forecast, or dashboard. The artifact moves through media channels faster than the workflow around it can be understood. People and models repeat the compact sign. Organizations translate the sign into budgets, OKRs, dashboards, roadmaps, performance expectations, and layoff narratives. Then the proxy can start to replace the value it was supposed to measure.

The practical move is to reopen the token by forcing each claim back to workflow.

Diagnostic checklist

• What work changes?
• Whose behavior changes?
• What data is needed?
• What human labor remains?
• What review burden appears?
• What complement is being assumed rather than built?
• Which expertise becomes more valuable if output gets cheaper?
• Which errors are acceptable?
• Which errors would be catastrophic?
• What costs move downstream?
• What metric could fool us?
• What value is being preserved?
• What would prove the thesis wrong?

A healthy organization treats AI as a pointer. Follow it, and you should find a workflow, a user, data, review, exception handling, costs, value, and a way to detect failure. If those are missing, the claim is still a token. If a metric replaces the value, the organization has drifted into proxy work. If the artifact is mistaken for the work, it is back in the projection error.

AI may become a general purpose technology. Inside firms, AI has already become a general purpose token. Strategy starts by separating the two: media velocity is not technological maturity, and repetition is not value creation.

Sources and notes

1. Jay Peters, “Duolingo will replace contract workers with AI”, The Verge, April 28, 2025, reproduces Luis von Ahn’s all-hands memo from Duolingo’s LinkedIn post, including “Duolingo is going to be AI-first,” the contractor line, and the headcount rule. Shopify’s newsroom article “Serious results, unserious methods: Shopify’s AI playground” quotes the summary of Tobi Lütke’s April 2025 internal memo: “reflexive AI usage is now a baseline expectation at Shopify.” ↩

2. On general purpose technologies, see Timothy Bresnahan and Manuel Trajtenberg, “General Purpose Technologies ‘Engines of Growth’?”. On LLMs as GPT-tech candidates, see Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models”. That paper estimates exposure to tasks, not automatic job loss. ↩

3. Richard Dawkins, The Selfish Gene, ch. 11, “Memes: The New Replicators,” supplies the replicator logic behind the term selfish token. Dawkins’s later “What’s in a Meme?” summary is useful here because it restates the relevant caution: the meme analogy helps explain how ideas spread, but it should not be treated as biological equivalence. ↩

4. Marshall McLuhan, Understanding Media, ch. 1, “The Medium Is the Message,” is the source for the media-theory claim that a medium’s message is the change of “scale or pace or pattern” it introduces into human affairs. The Marshall McLuhan estate summary quotes that passage and the related line that media shape the scale and form of human association and action. ↩

5. Neil Postman, Amusing Ourselves to Death, especially ch. 2, “Media as Epistemology,” is the background for the claim that media environments shape what a culture accepts as knowledge. The current Penguin Random House page identifies the book; Technopoly is a secondary background source for Postman’s broader critique of technological culture. ↩

6. Harold Innis, “The Bias of Communication”, argues that media differ in their capacity to carry knowledge over time or over space. That is the source of the time-biased / space-biased distinction used here. ↩

7. James W. Carey, “A Cultural Approach to Communication”, in Communication as Culture, develops the transmission and ritual views of communication used in this section. ↩

8. Andrey Mir’s work on digital orality names the contemporary return of oral-like, participatory, formulaic circulation on top of literate and computational infrastructure. The term is used here as a media-ecology lens, not as a claim that Mir endorses the full AI-token argument. ↩

9. Guy Debord, The Society of the Spectacle, thesis 4, defines spectacle as a social relation between people that is mediated by images. The AI-token version here applies that lens to organizational status and futurity. ↩

10. Jean Baudrillard, Simulacra and Simulation, supplies the map/territory and model-precession lens used for AI strategy artifacts that precede transformation. ↩

11. Friedrich Kittler, Gramophone, Film, Typewriter, supplies the media-materialist emphasis on storage, transmission, processing systems, and infrastructure. The Stanford excerpt opens with optical-fiber networks and digital convergence; the companion preface excerpt gives the compressed premise: “Media determine our situation.” ↩

12. Paul Virilio, Speed and Politics, supplies the dromological lens: speed as a structuring force, not merely a neutral acceleration of existing institutions. ↩

13. This table uses a narrow version of the media-ecology question behind Marshall and Eric McLuhan’s tetrad in Laws of Media: what a medium amplifies and what it pushes out of view. The full tetrad also asks what a medium retrieves and what it reverses into at its limit; those columns are omitted here because this essay’s immediate question is how organizational channels preserve portable AI signs and strip away mechanism. ↩

14. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work”, studied a staggered rollout of a generative-AI assistant to customer-support agents and reports a 14 percent average productivity gain in the NBER version, with larger gains for novice and lower-skilled workers. Later versions/reporting describe roughly 15 percent; the body uses the more conservative NBER headline. ↩

15. Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot”, reports that the Copilot group completed a bounded programming task 55.8 percent faster than the control group. ↩

16. Elise Paradis et al., “How much does AI impact development speed? An enterprise-based randomized controlled trial”, reports a best estimate of about 21 percent shorter task time for 96 Google software engineers on a complex enterprise-grade task, with a large confidence interval and a warning against assuming ecosystem-level generalization. ↩

17. METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, reports that experienced open-source developers working on familiar mature repositories took 19 percent longer with early-2025 AI tools allowed. METR later marked the result historical and possibly out of date for newer tools; that actually reinforces the point here, which is that the mechanism depends on setting, tools, task distribution, and time. ↩

18. Feiyang Xu, Poonacha K. Medappa, Murat M. Tunc, and Jan C. Fransoo, “AI-Assisted Programming Decreases the Productivity of Experienced Developers by Increasing the Technical Debt and Maintenance Burden”, analyzes open-source projects after Copilot adoption and reports more output from peripheral developers alongside more rework burden on core developers. ↩

19. Emerson Murphy-Hill, Jenna Butler, and Alexandra Savelieva, “Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft’s Early 2026 Rollout of Claude Code and GitHub Copilot CLI”, reports roughly 24 percent more merged pull requests among adopters and explicitly treats merged PRs as a proxy for output, not delivered value. ↩

20. Edwin A. Abbott, Flatland: A Romance of Many Dimensions, supplies the dimensional analogy. In sections 16-17, the Sphere appears to Flatlanders as a circle whose section changes size; the essay uses that scene for the organizational mistake of treating a visible cross-section as the whole generator. ↩

21. Challenger, Gray & Christmas, “Challenger Report: May Job Cuts Rise 16% from April; Highest May Total Since 2020”, reported AI as the leading employer-cited reason for U.S. job cuts in May 2026. The phrase employer-cited reason is intentional; it is not the same as proved cause, because layoff attribution often blends technological substitution with cost pressure, restructuring, overhiring correction, and investor narrative. ↩

22. The common Goodhart formulation is: when a measure becomes a target, it ceases to be a good measure. Mattson, Bushardt, and Artino, “When a Measure Becomes a Target, It Ceases to be a Good Measure”, note that this popular wording is most often generalized from Marilyn Strathern, while Goodhart’s original monetary-policy formulation was about statistical regularities collapsing under control pressure. The AI examples here are applications, not Goodhart’s original context. ↩

23. C. Thi Nguyen, “Value Capture”, Journal of Ethics and Social Philosophy 27, no. 3 (2024), defines value capture as what happens when simplified, often quantified versions of richer values come to dominate practical reasoning. Nguyen’s The Score develops the adjacent argument that scores, rankings, and metrics can train people and institutions to care about the score instead of the richer value. ↩

24. Isabelle Bousquette, “Why Some Companies Say AI ‘Tokenmaxxing’ Is Key to Survival”, The Wall Street Journal, April 14, 2026, reported that an internal Meta dashboard ranked employees by individual token usage and assigned titles including Token Legend. Brent D. Griffiths, “Amazon says it shut down a token leaderboard: ‘Don’t use AI just to use AI’”, Business Insider, May 29, 2026, reported that Amazon deprecated an informal employee-made KiroRank leaderboard after it encouraged some employees to do work that did not necessarily solve problems in order to climb the ranking. Griffiths’s “Top Indeed exec details why they’ll never have a ‘Tokenmaxxing’-esque leaderboard”, Business Insider, April 19, 2026, supplies the counterexample: Indeed monitors token use in the background but avoids making it a leaderboard or primary outcome metric. ↩

25. Michael Crichton coined the “Gell-Mann Amnesia effect” in “Why Speculate?”, a 2002 speech about noticing media errors in a domain one knows and then forgetting that lesson when reading about domains one does not know. The “reverse” version here is this essay’s extension, not Crichton’s term. ↩

26. Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics”, argues that AI’s full economic effects depend on complementary innovations, organizational change, new skills, and intangible capital. See also their later work on the productivity J-curve. ↩

27. Ben Shimkus, “Ford says AI alone couldn’t fix its quality problems. It needed to rehire veteran engineers to help”, Business Insider, June 25, 2026, reported that Ford executives said the company had hired, promoted, or brought back about 350 experienced technical specialists as part of its quality reset. The specialists mentor staff, lead design reviews, and improve AI and automated quality tools. The article also reports Charles Poon’s point that Ford had not done enough to preserve experienced engineers’ knowledge before some of it left the company. The relevant claim is narrow: Ford needed experienced technical specialists to make AI and automation work inside its quality system. ↩