Economics, Labor & Industry
The AI Economy

The new roles are real, growing measurably, and paying well. The qualifier is that they are creating a small number of high-value positions while simultaneously commoditizing a larger number of entry-level and routine positions. The net effect is real job creation, but not evenly distributed.

The most robustly growing AI-native roles (sourced from verified job postings data):

• AI / Machine Learning Engineer: US job postings grew 143% year-over-year in 2025. LinkedIn ranked this the #1 fastest-growing job title in the US in 2026. Salary range: $160,000–$312,000 for senior practitioners. These engineers build, fine-tune, and deploy AI systems.
• Prompt Engineer: Postings grew 135.8% YoY. Salary range: $110,000–$150,000 in the US. The role designs systematic approaches for reliable AI outputs, not just writing prompts but building reusable frameworks that minimize hallucination and maintain brand alignment. Projected 33% CAGR through 2030 for the broader market.
• AI Solutions Architect: Postings grew 109.3% YoY. Salary $139,000–$200,000. Works with engineering and business leadership to design scalable, compliant AI system architectures. The bridge between capability and enterprise deployment.
• AI Ethics Officer / Responsible AI Lead: Demand for AI ethics skills up 125% since 2023. Salary $120,000–$180,000. Creates ethical guidelines, conducts reviews, and manages compliance with EU AI Act and sector-specific regulatory requirements. Didn't exist as a standalone role three years ago.
• MLOps Engineer: Manages how models are deployed, monitored, and maintained in production. Growing in direct proportion to AI deployment expansion. Companies learning that shipping a model is 20% of the work; keeping it reliable, monitored, and updated is 80%.
• AI Content Creator: Postings grew 134.5% YoY. Not about writing text, about directing AI output strategically: knowing what to generate, how to evaluate quality, and how to maintain brand and factual integrity across AI-generated output at scale.
• Chief AI Officer (CAIO): 1 in 4 companies now have CAIOs (IBM, 2025). 66% of companies expect most organizations will hire CAIOs within two years. The White House required all federal agencies to create the role. Salary: variable but typically $250,000–$600,000+ at large enterprises.
• Generative Engine Optimization (GEO) Specialist: As AI handles 50%+ of search responses by 2028 (Gartner projection), optimizing how organizations are cited and summarized by AI systems, not just indexed by search, is a real emerging commercial need. GEO experts report improving AI platform visibility by 40%.
• AI Data Curator / Trainer: Prepares, annotates, and validates training datasets. The unglamorous but essential work that determines whether AI systems are accurate, representative, and ethically grounded. Salary $80,000–$115,000; premium rates for niche medical, legal, and scientific domains.
• AI Agent Developer / Automation Engineer: Specifically builds and maintains agentic AI workflows that autonomously execute business processes. Job postings mentioning "agentic AI" grew 985% between 2023 and 2024 (McKinsey data).

The important counterforce: 1 in 10 job postings now require AI skills (up from 1 in 30 in 2023). Skills in AI-exposed occupations are changing 66% faster than in less-exposed roles (PwC 2025). This means that most workers will not get new job titles, they will see their existing roles transformed. The new title holders are a smaller number than the workers whose current roles are being reshaped.

Goldman Sachs, Morgan Stanley, and JPMorgan have each published analyses concluding the AI sector is not currently in a bubble in the classical sense. Goldman Sachs CEO David Solomon, Amazon's Jeff Bezos, and OpenAI CEO Sam Altman have each said investors will lose money in AI. Both things are simultaneously true, and that tension is the core of the most important financial question in technology in 2026.

The case that this is NOT a bubble (the institutional consensus): Today's AI leaders, Microsoft, Google, NVIDIA, Meta, Amazon, have extraordinary balance sheets, genuine revenue, and strong profit margins. The P/E ratios of AI-exposed companies remain well below dot-com bubble levels. NVIDIA's $5 trillion market cap (October 2025) reflects $215.9 billion in annual revenue and $120 billion in net income, this is not speculative valuation. A December 2025 five-factor diagnostic framework applied to AI investment found it linked to actual enterprise revenue, not speculation. Goldman's analysis concludes: "fundamentally driven growth rather than irrational speculation about future growth."

The case that parts of this ARE bubble-like: Goldman Sachs also found that $19 trillion in market cap is "running ahead of economic impact." The broader AI ecosystem includes OpenAI, $840 billion valuation against $25 billion in revenue, burning cash at extraordinary rates with no roadmap to profitability. Total global AI capex estimated at $527 billion in 2026. A February 2026 NBER study of 6,000 executives found 89% of firms report no measurable AI productivity impact to date, the investment thesis for much of this spending is that it will pay off eventually. Sam Altman himself warned that "people will overinvest and lose money" in this phase.

Where bubble dynamics clearly exist, the startup layer: Single-purpose AI applications (email assistants, scheduling tools, simple chatbots) are being destroyed by each frontier model update, a dynamic venture capitalists are calling "feature-zation." Analysts predicted 99% of AI startups would fail by 2026 as users realized they had no proprietary technology. Users mistook traffic for revenue and adoption for defensibility. This layer of the ecosystem is actively correcting.

The most specific bubble risk, private credit for data centers: Morgan Stanley estimates $3 trillion in data center spending 2025–2028, half funded by private credit. Private-credit lenders have extended large sums to untested borrowers using unconventional collateral. If a major data center developer fails, the contagion risk through this opaque private credit market is the scenario most cited by serious analysts as the specific mechanism through which an AI financial shock could propagate into the broader economy.

This is the most important economic question in AI right now, and the honest answer is: the data is genuinely mixed, the gap between task-level and organizational productivity is real and documented, and the situation rhymes historically with a well-studied pattern economists have seen before.

The headline finding (February 2026 NBER, peer-reviewed): A study of approximately 6,000 senior executives across the US, UK, Germany, and Australia, including companies already using AI, found that 89% report no measurable impact of AI on labor productivity over the past three years. More than 90% report no impact on employment. Apollo Chief Economist Torsten Slok summarized: "AI is everywhere except in the incoming macroeconomic data."

The counterevidence, macro signs of something happening: Stanford economist Erik Brynjolfsson, writing in the Financial Times (February 2026), documented that US productivity grew approximately 2.7% in 2025, nearly double the 1.4% annual average of the prior decade. Q4 2025 GDP tracked at 3.7%. The Federal Reserve Bank of St. Louis found 1.9% cumulative excess productivity growth since ChatGPT launched in November 2022. Individual workplace studies show task-level productivity gains of 14–55% in controlled experiments (customer support agents, software developers, legal document reviewers).

The "Solow Paradox" redux, why this pattern has a name: In 1987, Nobel economist Robert Solow observed that "you can see the computer age everywhere but in the productivity statistics." The computer revolution took roughly 15–20 years to show up in aggregate productivity data, and when it did, it drove the sustained growth of the 1990s. The lag was explained by the time required for organizations to restructure around new technology: not just adopting the tool, but redesigning workflows, training people, eliminating redundant processes, and building complementary infrastructure. Nobel economist Daron Acemoglu's 2024 analysis projected a "modest 0.5% productivity gain over the next decade", acknowledging the paradox but projecting smaller eventual payoff than historical computer adoption produced.

The specific failure pattern: McKinsey's 2025 surveys identified what they call the "GenAI paradox": ~80% of companies use generative AI, ~80% report no significant bottom-line impact. A Boston Consulting Group study published in Harvard Business Review (March 2026) found that AI use increased email time (doubled) and reduced focused work sessions by 9%, meaning AI is adding cognitive load in some contexts, not reducing it. ManpowerGroup's 2026 survey found worker AI confidence fell 18% even as use grew 13%. The productivity gains are real at the task level. The organizational absorption is not keeping pace.

AI does both simultaneously, and which effect dominates depends on the industry sector and the specific type of work involved. The tension is real and the resolution is not uniform.

How AI accelerates onshoring: The historical driver of offshoring was labor cost arbitrage, moving repetitive, manual, or procedural tasks to lower-wage markets. AI directly reduces the cost of doing those tasks anywhere, including in high-wage countries. If a legal services firm can review contracts at $0.50 per document using an LLM instead of $50 per document using an offshore paralegal, the offshore location advantage disappears. For call centers, data entry, basic software testing, and document processing, the primary categories of knowledge-work offshoring, AI is reducing the economic rationale for offshore labor faster than those locations can upskill into higher-value work.

Additionally: Supply chain security concerns (post-COVID, amplified by geopolitical tensions) are pushing manufacturing reshoring. The US CHIPS and Science Act is specifically designed to reshore semiconductor manufacturing. Governments across the West are actively subsidizing AI infrastructure domestically, TSMC Arizona, Intel Ohio, Samsung Texas, SK Hynix Indiana. AI reduces the need for large human operations in these facilities, making the economics more favorable for domestic siting.

How AI undermines traditional onshoring logic: The traditional argument for onshoring was that proximity, cultural alignment, and quality control justified paying higher domestic wages. When AI eliminates the manual work entirely, there is no "onshoring" to do, the labor is simply gone. A company that was debating between offshore and domestic call center staffing may ultimately deploy an AI agent that replaces both. This is not onshoring, it is elimination. The Goldman Sachs finding that AI can automate tasks representing 25% of all US work hours applies to domestic workers as much as offshore ones.

The manufacturing exception: Physical manufacturing, particularly precision manufacturing that requires robotics, skilled technicians, and proximity to the supply chain, is genuinely reshoring in AI-exposed sectors. AI reduces the proportion of work done by low-skill labor and increases the proportion requiring engineers, technicians, and systems operators, skill profiles more competitive in developed economies. Goldman Sachs (March 2026) projects ~500,000 net new US construction and electrical jobs just from data center build-out.

This is not a hypothetical future risk, the transition is measurably underway and is the single largest structural threat to the existing digital advertising business model. Gartner projects that traditional search engine volume will decline by 25% by 2026 and 50% by 2028 as users shift to AI-native interfaces. The $300+ billion global search advertising market is not simply threatened; it is being structurally disintermediated.

The mechanism of disruption: Traditional search monetizes intent, a user searches "best running shoes," sees ads, and potentially clicks through to a retailer. The entire ecosystem depends on users visiting web pages. AI search summarizes answers directly and reduces click-through dramatically. Studies show AI search visitors convert 4–5× better when they do click through (higher intent), but overall click volume falls dramatically. The content publishers who depend on search traffic to fund journalism, research, and creative production lose their primary discovery mechanism.

What Google and Microsoft are doing: Both have integrated AI into search (Google AI Overviews, Bing Copilot), attempting to maintain advertising placement within AI-generated responses. Google's approach: insert sponsored content into AI summaries. The challenge: AI summaries compress multiple sources into one answer, reducing the number of opportunities for ad insertion from ten blue links to potentially one or two placements. Revenue per query falls even if query volume holds.

The emerging replacement model, Generative Engine Optimization (GEO): As AI systems become the primary way users access information, organizations are investing in being cited and summarized favorably by AI systems rather than indexed by search engines. GEO specialists help companies appear in AI-generated responses, a nascent but rapidly growing practice. Whether this creates a new monetizable layer for AI providers (sponsored citations?) remains unclear and contested.

The collateral damage, independent publishers and journalism: The open internet content ecosystem is funded by search-driven advertising revenue. If search intent flows to AI assistants that synthesize content without directing traffic to source pages, the revenue model for independent journalism, reference content, and the long-tail web collapses. This is not speculative, Google's AI Overviews has already produced measurable traffic declines to independent publishers since its 2024 rollout. The organizations most dependent on this model have the least capacity to absorb the shock.

The AI performance on professional exams is real but widely misunderstood. The implication for professional licensing is genuinely contested, with serious arguments on both sides. We will not paper over the disagreement.

What the AI benchmark performance actually shows: GPT-3.5 (2023) passed the bar exam's multiple-choice component at ~50% accuracy, barely above the pass threshold. By 2024–2025, frontier models pass all three components of the bar exam at performance levels comparable to the top quartile of human test-takers. GPT-4 scored in approximately the 90th percentile on the bar exam. AI systems pass USMLE Step 1, 2, and 3, the early ChatGPT result was borderline passing; current frontier models score at attending-physician levels. AI passes the CPA exam at rates comparable to strong human candidates.

The "exam-passing ≠ professional capability" argument (strong position): Professional exams test a structured knowledge base under controlled, time-limited conditions with known question types. Actual professional practice requires: (1) applying that knowledge to ambiguous real-world situations with incomplete information; (2) maintaining ongoing relationships with clients, patients, and opposing counsel; (3) exercising judgment in ethically complex situations; (4) accountability and liability for decisions; (5) integrating new information continuously over years. An AI that passes the bar exam cannot appear before a court as counsel of record, cannot be disbarred for misconduct, cannot be sued for malpractice. The exam is a threshold signal about knowledge, not a complete measure of professional capability. The AMA, ABA, and AICPA have each made versions of this argument explicitly.

The "licensing as guild protection" argument (also strong position): Professional licensing has always served two purposes: consumer protection (ensuring minimum competence) and market protection (limiting supply to sustain high fees). The argument that AI "only passes the exam, not the whole job" has historically been made about every technology that threatened professional monopolies, from paralegals threatening lawyers to nurse practitioners threatening doctors. In many domains, AI may genuinely provide better diagnostic accuracy, more comprehensive legal research, and more consistent accounting than the average human practitioner, particularly for routine cases. If the threshold function of licensing is knowledge verification, and AI surpasses that threshold, the question of what licensing is actually protecting becomes legitimate. Gartner projects that by 2028, AI will reduce the need for specialized human expertise in 40% of professional services tasks.

The most likely near-term trajectory: Not elimination of licensing, but stratification. AI tools become required competencies for licensed professionals. The value of professional licensing shifts from knowledge certification toward: accountability (someone to sue), judgment in ambiguous cases, patient/client relationship management, and ethical responsibility. Licensing requirements will likely expand to include demonstrated AI tool competency, how well you direct and validate AI, rather than just domain knowledge alone.

Proprietary data is genuinely valuable, but it is neither "the new oil" nor the sole determinant of competitive advantage. The relationship between data, algorithms, and compute is more dynamic and contextual than either extreme suggests.

Where data advantage is real and durable: In domains where high-quality, labeled, proprietary data is genuinely difficult to acquire, data is the primary moat. Healthcare diagnostic AI trained on millions of proprietary patient records with expert clinical annotations is extremely difficult to replicate, not because the algorithm is secret but because the annotated data doesn't exist elsewhere. Financial trading systems trained on proprietary order flow have genuine data moats. Any company with a large, proprietary behavioral dataset (search history, purchase patterns, interaction logs) has training signal that cannot be purchased or synthesized, Amazon's product interaction data, Google's search queries, Spotify's listening patterns.

Where algorithmic innovation has outpaced data advantage, repeatedly: DeepSeek R1 matched frontier US model performance at ~$6M training cost with significantly less proprietary data, using reinforcement learning techniques rather than data volume. The Chinchilla finding showed that data quality and training method matter more than raw data volume. Microsoft's Phi models demonstrated frontier-adjacent performance on carefully curated synthetic data, not proprietary web data. The history of AI is full of examples of algorithmic innovation closing gaps that data-rich incumbents assumed were durable moats.

The "oil" analogy's specific failures: Data, unlike oil, is non-rivalrous, sharing it doesn't reduce your supply. Data depreciates in relevance faster than oil depletes. Data advantages can be eroded by privacy regulation (GDPR limiting behavioral data collection). Synthetic data is increasingly substituting for real data in domains where verification is possible. And the value of data depends entirely on the quality of the model trained on it, bad algorithms applied to good data produce bad models.

The most accurate framing: Proprietary data is a necessary but not sufficient condition for AI competitive advantage. It matters enormously in specific contexts (healthcare, finance, specialized industrial domains). It matters less in general-purpose language and reasoning domains where internet-scale data is abundant and algorithmic innovation can substitute. The strategic question for any organization is whether their data is genuinely proprietary, genuinely domain-critical, and genuinely better than what can be generated or acquired, not whether "data" in the abstract is valuable.

The discussion of AI risk in public discourse focuses heavily on misaligned superintelligence or deepfake disinformation, genuinely important risks. The more immediate systemic fragility risks from AI in critical infrastructure are less dramatic, harder to attribute, and operating right now, which may make them more dangerous precisely because they attract less attention.

Monoculture risk, the single point of failure problem: When 70% of enterprises use the same two or three foundation models (GPT-5, Claude, Gemini), a systematic failure, security vulnerability, or unexpected behavioral change in any of those models propagates instantaneously across thousands of applications simultaneously. A traditional software vulnerability affects one system; a model-level vulnerability affects every application built on that model. The 2024 CrowdStrike outage, a single software update crashing 8.5 million Windows devices, previews this dynamic at AI scale. When the same AI model is simultaneously processing medical records, financial transactions, infrastructure monitoring alerts, and government communications, a systematic failure creates cascading failures across multiple critical domains at once.

Correlated error risk in financial systems: When AI systems from the same training distribution are simultaneously making recommendations across competing financial institutions, they are likely to make correlated errors, all moving in the same direction at the same time. High-frequency trading already operates faster than human regulatory oversight. AI-assisted investment decision-making at multiple institutions using similar models creates synchronized behavior that amplifies market movements in ways that human-operated markets partially avoid through individual judgment variance. The 2010 Flash Crash produced a 1,000-point market drop in minutes from algorithmic trading interactions, AI-assisted investment at scale could produce larger, faster, and harder-to-attribute events.

Supply chain optimization brittleness: Just-in-time supply chains optimized by AI are extremely efficient under normal conditions and extremely fragile under conditions outside the training distribution. COVID demonstrated this with human-designed supply chains; AI-optimized chains may be more efficient but no more robust to black swan events. A model optimized for efficiency will systematically under-provision buffers and redundancy, exactly what is needed when conditions diverge from training data.

Energy grid vulnerability: AI systems managing power grid load balancing, demand prediction, and fault detection are being deployed across US and European grids. These systems are trained on historical demand patterns. An adversarial actor with knowledge of the model's training distribution could potentially trigger grid failures by engineering load conditions the model misclassifies as normal. The CISA acknowledged AI-enabled attacks on grid infrastructure as a top-tier concern in 2025.

This question operates at three distinct levels, economic, aesthetic, and philosophical, and deserves an answer at each. We will not collapse them into a single reassuring narrative.

The economic reality: The market for commodity creative output, stock photography, background music, template-based illustration, product description writing, basic graphic design, is in active free-fall. Getty Images, Shutterstock, and similar stock media platforms are seeing structural revenue declines as AI generation replaces their lower-tier catalogs. Voice actors for commercial content, translators for routine documents, and junior copywriters are all experiencing measurable demand contraction in markets where AI output is "good enough." The economic harm is real, concentrated in the mid-to-lower tiers of creative labor markets, and already occurring.

The economic counterargument (also real): Throughout history, technological reproduction has simultaneously commoditized creative output and expanded the total market for creativity. Photography killed miniature portrait painting and created photojournalism, fashion photography, and a billion-dollar portrait studio industry. Recorded music killed the traveling minstrel and created global superstars, billion-dollar labels, and concert revenue. The question is whether AI follows this pattern, creating more total demand for creative work even as it disrupts specific roles, or whether the scale and speed of AI creative output is categorically different.

The aesthetic question, does AI-generated creativity have value? This is genuinely unresolved and intersects deeply with what creativity is for. If the function of art is to produce aesthetic experience in the viewer, AI-generated images produce that experience measurably, they win aesthetic competitions when identity is concealed. If the function of art includes the communication of human experience, perspective, and suffering to another human, then AI cannot perform that function, regardless of its technical output quality. The market is currently answering this question pragmatically: AI output is priced lower and accepted when context allows; human-certified creative work commands premium pricing in markets where provenance matters.

The scarcity question, feature or bug? Human creative scarcity, the fact that there is only one Taylor Swift, one Banksy, one Toni Morrison, is currently being leveraged as the primary defense of human creative economic value. This framing treats scarcity as intrinsically valuable rather than as a byproduct of human limitation. The philosophical question: if AI can produce music indistinguishable from a talented but not famous composer, what exactly are we preserving by paying premium prices for the human-verified version? The consumer who can't tell the difference is paying for proof of origin, not for the experience. This is not a rhetorical dismissal, provenance and authenticity have real cultural value. But that value is different from the value of the creative experience itself, and conflating them obscures what is actually at stake.

This is among the most important second-order questions in AI labor economics and is receiving far less attention than headline job displacement. The compression of the experience premium, the salary and status advantage that accumulates with years on the job, is one of the most structurally disruptive dynamics that AI introduces to organizational life.

The mechanism is documented, if not yet at scale: Goldman Sachs research (March 2026) confirms that AI can automate tasks representing 25% of all US work hours, concentrated in cognitive and procedural domains that are precisely where junior-senior experience gaps show up. Erik Brynjolfsson's landmark 2023 study of 5,172 customer support agents found that AI assistance produced a 14% average productivity gain, but gains were concentrated among novice workers, while experienced agents saw minimal improvement. This is the compression dynamic in its purest form: AI raises the floor without raising the ceiling.

The salary premium threat: Mid-career salary premiums in knowledge professions, law, consulting, finance, software development, are explicitly justified by accumulated domain expertise, judgment, and institutional knowledge. If an AI tool closes 60–70% of the productivity gap between a third-year analyst and a tenth-year partner, the justification for the compensation differential weakens structurally. Firms will face pressure to either flatten salary structures or demonstrate that experienced workers provide value that AI cannot replicate. The AI Impact SWOT analysis (2026) explicitly identifies the "mid-career salary premium" as among the most economically vulnerable features of knowledge-work compensation.

The mentorship and knowledge transmission crisis: Organizational learning has historically depended on experienced workers transmitting tacit knowledge to junior colleagues through direct work, observation, and gradual responsibility transfer. Two converging AI dynamics threaten this: (1) companies are hiring fewer juniors (the Second Talent analysis confirms this, companies that hired 5 juniors now hire 2 mid-level with AI tools); and (2) juniors who are hired interact with AI for tasks they previously did under senior supervision, reducing the tacit learning opportunity. The mechanism by which domain expertise is socially reproduced in organizations is disrupted at both ends, fewer juniors to teach and fewer teaching situations for them to be in.

The knowledge fragility consequence: If a generation of professionals learns to do work with AI assistance rather than developing the underlying judgment independently, what happens when the AI is unavailable, wrong, or insufficient for a novel situation? This is not hypothetical anxiety, it mirrors concerns about calculator dependency in mathematics, GPS dependency in navigation, and autocorrect dependency in spelling. The organizations most dependent on AI for routine cognitive work may be cultivating a talent pool that is excellent at directing AI but thin on independent judgment capacity, exactly what is needed when AI fails.