Generative AI

Confidence 0.95 · 22 sources · last confirmed 2026-05-28

AI systems that produce novel content (text, images, code, video, audio) by sampling from learned distributions over training data. Practically: large language models, image diffusion models, video generators, and the orchestrating systems built on top of them. As of 2024, GenAI represents >20% of all AI-related private investment ($33.9B globally) and is the dominant force behind the 2024 jump in enterprise-ai-adoption.

By 2026, the wiki tracks a parallel paradigm-vocabulary rename of generative-AI engineering: Andrej Karpathy’s software-3.0 (Sequoia AI Ascent, 29 April 2026) frames the same construct as the third paradigm of computing (1.0 = explicit rules / code; 2.0 = learned weights / datasets; 3.0 = prompts / context windows). Generative AI is what Software 3.0 produces and runs on; the practitioner-discipline that builds with it is agentic-engineering (professional quality bar) or vibe-coding (floor-raising democratisation).

Working definition

In the AI Index taxonomy, “Generative AI” is a sub-class of AI distinguished by content generation rather than pattern recognition or prediction. It typically rests on foundation-models — large pretrained models adapted via prompting, fine-tuning, or retrieval. Generative AI is also where agent workflows are increasingly built (Salesforce’s Agentforce, autonomous coding agents, etc.).

Key claims

Investment

• $33.9B private investment in GenAI in 2024, up 18.7% from 2023, 8.5× 2022 levels. Now >20% of all AI-related private investment.
• U.S. GenAI investment exceeded the combined total of China + EU + UK by $25.4B in 2024 (gap was $21.8B in 2023).
• Source: AI Index 2025 §4.3 Chapter Highlights 2–3.

GenAI as part of the “AI foundation” three-arena cluster (MGI 2026)

MGI’s The race takes off in the next big arenas of competition (March 2026) bundles GenAI into a three-arena “AI foundation” cluster with semiconductors and cloud services. The cluster is the largest single contributor to the 18-future-arena market-cap accretion since 2022:

Metric	Value (2022–25)
AI foundation theme market-cap accretion	+$10.77T (60% of the 18-arena total)
AI foundation theme revenue accretion	+$490B
AI software & services market cap 2022–25 CAGR	142%
AI software & services revenue 2022–25 CAGR	55%
AI software & services 2040 revenue projection	$1.5T–$4.6T (17–25% CAGR from $85B baseline)
Semiconductors market-cap accretion since 2022	+$7.39T alone — ~68% of the AI-foundation total
Cloud services market-cap accretion since 2022	+$1.83T
AI software & services market-cap accretion since 2022	+$1.55T

Two structural reads:

• GenAI is not an industry on its own in MGI’s frame; it is one of three arenas in the AI-foundation bundle, with the other two (semiconductors + cloud) capturing the infrastructure and deployment substrate. Semiconductors alone account for ~68% of the AI-foundation market-cap accretion — i.e., value-capture in the GenAI wave has so far accrued disproportionately to the firms that enable GenAI rather than the firms that deploy GenAI as an end product (consistent with the wiki’s existing automation-vs-augmentation and adoption-depth gradients showing firm-side returns are heavily concentrated at the substrate layer).
• The $1.5T–$4.6T 2040 revenue projection for AI software & services is the wiki’s longest-horizon empirically-grounded GenAI revenue number. The headline open question MGI itself raises: “Whether investor expectations for AI’s future will be supported by companies’ sustained returns on invested capital above the cost of capital is one of the biggest open questions in business today.”

Enterprise application spending: 22× in two years

Nishar & Nohria (HBR.org Digital, April 2026) — drawing on industry analyst data:

• Enterprise GenAI application spending: $1.7B (2023) → $37B (2025) — more than 20× in two years. SaaS took ~10 years to achieve a comparable level of penetration.
• ~40% of code is now AI-generated; vast majority of developers use AI coding tools daily.
• Substitution effects: >1/3 of companies report replacing at least one SaaS tool with a custom-built GenAI alternative; majority expect to build more in the coming year.
• Public SaaS valuations 30–60% below 2021 peaks as the standardized-tool bargain weakens.
• Tools enabling the shift: Cursor, Replit Agent, Claude Code, OpenAI Codex — make idea-to-functional-application possible in hours/days. Coined as “vibe coding” (a user describes what they want in plain language; the AI generates working software).

Enterprise use

• 71% of orgs use GenAI in at least one business function in 2024, up from 33% in 2023 — more than doubled. Source: AI Index 2025 §4.4.
• U.S. AI-job postings citing GenAI skills tripled YoY in 2024 (Lightcast data).
• Top use cases (% of orgs deploying GenAI for this function): marketing strategy content (27%), knowledge management (19%), personalization (19%), design (14%), code creation (13%).
• Only 1% of C-suite executives describe their GenAI rollouts as “mature.” Most are still capturing value at small scale. See enterprise-ai-adoption.

Return on iteration: the compounding-benefits reframing (Kiron & Schrage 2026)

A reframing of the AI-ROI question, distinct from but adjacent to the productivity studies above. The premise: in a GenAI era, the marginal cost of a first attempt (drafts, code, prototypes, analyses) has dropped sharply — what remains expensive is evaluating what AI produces and capturing what was learned. Productivity is therefore measured not just per unit of input but per unit of learning per unit of interaction.

The economic framing: AI deployments are either consumption economics (asset depreciation: task in, output out, loop closes; knowledge evaporates after each interaction) or compounding economics (asset appreciation: each cycle improves the next). Two structural dynamics make compounding the imperative:

1. The Polanyi tacit-dimension breach — for decades, tacit knowledge (“humans know more than they can articulate”) protected knowledge workers from automation. LLMs now breach that moat not by codifying tacit knowledge but by inferring it from behavioral traces at scale. Legal reasoning in briefs, financial judgment in analyst reports, strategic thinking in board presentations — as these become legible to AI models, the tacit expertise embedded in them becomes machine-readable.
2. Jevons paradox amplification — efficiency gains stimulate demand. As tacit expertise becomes machine-readable, sophisticated capability gets cheaper, and projects previously too expensive to prototype proliferate. Iteration cycles compress from months to hours.

Empirical anchors:

• Orgs that build systematic feedback loops between humans and AI are 6× more likely to derive substantial financial benefits from AI (Ransbotham et al. 2024).
• Orgs investing in learning with AI are 73% more likely to achieve significant financial impact.
• Orgs combining strong organizational learning with AI-specific learning are up to 80% more effective at managing uncertainty.

Operational mechanism — the verification → evaluation → learning capture flywheel; see micro-productivity-trap and enterprise-ai-adoption for the full treatment. The wiki’s coining of “return on iteration” captures the shift: ROI in a GenAI era is the rate at which captured learning changes subsequent practice, not the static ratio of value-generated to cost.

Learning curves: model selection becomes a skill (Anthropic Economic Index, 5th report)

The fifth Economic Index report extends the four-report longitudinal series with a learning-curve analysis (Feb 5–12, 2026 sample). Two findings worth promoting up here from the ai-employment-effects page because they shape how GenAI productivity is realized:

1. Model selection slopes are steep and asymmetric: users select premium-tier models (Opus over Sonnet/Haiku) differentially for higher-value tasks. +1.48 pp Opus share per +$10/hour task value on Claude.ai; +2.79 pp per +$10 on the 1P API (about twice as steep on the enterprise surface). Choosing the right model tier per task is a learnable, measurable skill.
2. Tenure earns ~3–4pp success advantage after controls — high-tenure Claude users reach the same model with the same task and still succeed more, controlling for task type, model, and conversation complexity. This is the wiki’s first quantified evidence of AI literacy as productivity differentiator at population scale.

These two findings reposition the productivity story: headline speedups (9-12× for HS / college tasks) are realized only when users learn the surface — model picker, prompt structure, recovery from failure. GenAI’s productivity gain is partly endogenous to AI-literacy investment.

Measurement framework: economic primitives (Anthropic Economic Index, 4th report)

The fourth Anthropic Economic Index report introduces five economic primitives for measuring real-world GenAI use, derived from privacy-preserving classification of Claude conversations:

1. Task complexity — human time without AI; multi-tasking within a conversation
2. Human and AI skill level — years of education needed to understand prompts and responses
3. Use case — work / education / personal
4. AI autonomy — degree of delegation, from collaboration to fully directive
5. Task success — whether the task was completed

The framework lets the wiki track GenAI productivity, complexity, and autonomy at population scale (~1M Claude.ai conversations + ~1M API transcripts per sample). Headline numbers from the November 2025 sample:

• Tasks needing high-school education: ~9× speedup on Claude.ai.
• Tasks needing a college degree: ~12× speedup on Claude.ai (greater on API).
• Success rate falls slightly with complexity (70% → 66% from sub-HS to college-degree tasks).
• Aggregate productivity contribution: +1.0 to +1.2 pp/yr (reliability-adjusted) — see ai-employment-effects.

Field deployment evidence: Bain client EBITDA + Lowe’s case (Dutt, Chatterji et al. 2026)

A practitioner-vantage HBR article co-authored by Bain & Company partners and OpenAI’s Economic Research team argues most firms get stuck in a micro-productivity trap — task-level gains that fail to translate to firm-level results. Firms that escape it via a four-step transformation framework see:

• Bain client EBITDA gains: 10–25% (continuing to scale as programmes scale).
• Lowe’s + OpenAI partnership (Mylow / Mylow Companion launched March 2025): online conversion >2× when customers engage Mylow; +200 bps customer satisfaction when associates use Mylow Companion; deployed across 1,700+ stores.
• FabricationCo (anonymized Fortune 1000 manufacturer): ~$30M additional profit on track; quote generation ~15× faster; +10pp win rate in 3 months.

The article frames “process redesign — not the technology” as the most challenging part of GenAI deployment and the source of most of the value.

Field-experimental evidence of GPT-4 in creative idea generation (Boussioux et al. 2024)

A field study comparing human crowdsourcing vs human-AI prompt-engineered solutions on a circular-economy business-ideation challenge:

Dimension	Human Crowd	Human-AI
Novelty	Higher (3.508 vs 3.230 / 3.469)	Lower
Strategic viability, environmental + financial value, overall quality	Lower	Higher (single-instance differentiated higher than multi-instance independent)

Cost / time profile — HAI cost $27 + ~5.5 hours vs HC cost $2,555 + ~2,520 human-hours (~94× cost reduction, ~458× time reduction).

The paper argues HAI solutions concentrate in incremental search space (proximal to existing ideas) while HC retains an advantage at the extreme top of the novelty distribution.

Field-experimental evidence of GPT-4 in knowledge work (Dell’Acqua et al. 2026)

A preregistered RCT with 758 BCG consultants:

• Inside the AI capability frontier: AI users +12.2% completion, +25.1% faster, +33.9% quality (Cohen d ≈ 0.96).
• Outside the frontier: AI users 19 pp less likely to be correct on a deliberately AI-defeating task.
• Both bottom-half-skill (+31%) and top-half-skill (+11%) consultants gain; AI is an equalizer even among elite professionals.

This paper introduces the jagged frontier concept to describe the unevenness of GPT-4’s capability across tasks of similar perceived difficulty — a per-task property, not a uniform feature.

2025–26 numbers from the AI Index 2026

• Population adoption: 53% within three years — faster than the PC or the internet.
• Adoption correlates with GDP per capita; outliers: Singapore 61%, UAE 54%; U.S. ranks 24th at 28.3%.
• Estimated value of GenAI tools to U.S. consumers: $172B annually by early 2026; median value per user tripled between 2025 and 2026.
• U.S. private AI investment: $285.9B in 2025 — ~2.6× the $109.1B reported in 2024 by the prior edition — and ~23× China’s $12.4B.
• 1,953 newly funded AI companies in U.S. in 2025 — more than 10× the next country.
• Organizational adoption: 88% (up from 78% in 2024).
• >80% of U.S. high school and college students now use AI for school-related tasks; 4 in 5 university students use generative AI.

Capability gains in 2024

• AI video generation breakthroughs: OpenAI’s SORA, Stable Video Diffusion 3D and 4D, Meta’s Movie Gen, Google DeepMind’s Veo 2.
• Test-time compute (e.g., OpenAI’s o1, o3) — models that iteratively reason — dramatically improved performance: o1 = 74.4% on IMO qualifier vs. GPT-4o’s 9.3%, but at ~6× the cost and 30× the latency.
• Inference cost crashed: GPT-3.5-equivalent cost dropped 280× in 18 months (Nov 2022 $20/M tokens → Oct 2024 $0.07/M tokens via Gemini-1.5-Flash-8B).
• Smaller models matching big ones: Microsoft’s Phi-3-mini (3.8B params) matched the 60% MMLU threshold previously held by PaLM (540B) — 142× reduction in 2 years.
• AI agents show early promise: RE-Bench results — top systems score 4× human experts in 2-hour budgets, but humans win 2:1 at 32 hours. Already match human expertise on select tasks (e.g., writing Triton kernels).

Agents (now its own concept page)

GenAI is the substrate for autonomous agents — see the dedicated ai-agents page covering the chatbot → agent → multi-agent progression, RE-Bench results, deployment examples (Salesforce Agentforce, Italgas DANA, Harvey, GitHub Copilot), and the cross-source debate over where agents fit in org maturity. MIT CISR places autonomous agents as a Stage 3+ attribute; Anand-Wu places them in the “no regrets zone” of the task-suitability 2×2; Cisco frames them as the near-term productivity story for everyone.

The access democratization (Anand-Wu)

Anand-Wu argue GenAI’s most underrated change is access, not intelligence:

1. Nontechie employees can use GenAI without expert support. For decades AI was the domain of engineers, programmers, data scientists. ChatGPT changed that with natural-language interaction.
2. GenAI is increasingly embedded into existing tools — email, videoconferencing, spreadsheets, CRM, ERP — lowering adoption barriers further.

Anand-Wu compare this to the MS-DOS → GUI transition of the 1980s: not necessarily more powerful, but dramatically more accessible. The strategic implication: competitive advantage will not come from access to the technology (everyone has it) but from complementary assets — proprietary data, unique people/processes/culture. See enterprise-ai-adoption.

Tools and embeddings (mentioned across sources)

Tool	Function	Citation
ChatGPT (OpenAI)	The democratization breakthrough	Anand-Wu
Harvey	Legal contract drafting; quality-control zone	Anand-Wu
GitHub Copilot	Code generation, debugging	Anand-Wu
Salesforce Agentforce	Business operations agents	AI Index 2025
Microsoft Copilot	Embedded productivity assistant	AI Index 2025 (job-postings data)
Italgas DANA	Generative-AI network control	MIT Sloan
Vantage / Executive LLM (Google Research, Gemini 2.5 Pro + Gemini 3)	Scalable assessment of human durable-skills (collaboration / creativity / critical thinking); LLM-as-conversational-teammate-and-autorater; Pearson 0.88 vs human experts	Globerson et al. 2026
Harness frameworks (LangChain / LangGraph, Microsoft Agent Framework v1.0, Google ADK, LlamaIndex, CrewAI, Haystack, DSPy)	Pre-assembled agent harness kits — chains, agents, memory modules, tool wrappers; sit at Layer 2 of the standard 4-layer agent stack	Kokane 2026, Chatterjee 2026

Limits

• Complex reasoning remains brittle. AI excels at IMO problems but fails PlanBench (logical planning) even when provably correct solutions exist. Limits effectiveness in high-stakes settings where precision is critical.
• Bias persists. GenAI continues to exhibit implicit biases (race, gender, occupation) despite explicit-bias mitigation. See responsible-ai.
• The data commons is shrinking — domains restricting AI training scrapers jumped from 5–7% to 20–33% in one year. Implications for training data quality and diversity.

Automation vs. augmentation: a load-bearing distinction

The strategic choice for any GenAI deployment is whether it automates work (substitutes for labor) or augments it (complements labor). automation-vs-augmentation tracks this as a standalone analytical lens. The empirical record:

• Augmentation effects (productivity): Brynjolfsson, Li & Raymond 2025 (QJE) is the canonical primary source. Customer-support AI built on GPT-3, designed explicitly to augment (not replace), produced +15% productivity with equalizing effect on low-skill workers (+30% RPH) and small quality decline at the top.
• Automation effects (employment): Brynjolfsson, Chandar & Chen 2025 shows the labor-market correlate: declining entry-level employment in occupations with high automation (not augmentation) shares of AI use, per ADP payroll data covering ~25M U.S. workers.

The choice between automation and augmentation thus has measurable consequences on both productivity (augmentation positive) and labor markets (automation contracts entry-level employment). Most enterprise deployments today are augmentative; agents are increasingly automative — see ai-agents.

Debates and supersession

• Will inference-cost decline continue? Hardware (-30%/yr) and energy efficiency (+40%/yr) trends support it; data-commons shrinkage (see responsible-ai) cuts the other way; energy-supply constraints (driving nuclear partnerships) are a third force.
• Test-time compute as a new scaling axis. Promising but expensive (o1: 6× cost, 30× latency vs. GPT-4o). Open question: economic viability for routine enterprise use, or only for high-value reasoning tasks.
• Hype vs. value gap. 71% adoption + 1% maturity + revenue gains mostly <5% per function. The dollar productivity story is real but smaller than the discourse implies.

Related concepts

• foundation-models — the substrate
• enterprise-ai-adoption — the deployment context
• responsible-ai — the governance overlay
• ai-benchmarks — how capabilities are measured