S&P 1500 AI Review 2022–2025: Methodology & Data Guide

S&P 1500 AI Review 2022–2025: Methodology

For readers who want to understand the choices behind the numbers, not just the numbers themselves, see all the methods below.

ABSTRACT

What this index measures and what it found

Between fiscal year 2022 and 2025, the share of S&P 1500 companies disclosing AI in their annual 10-K filings rose from 25% to 90% — a near-complete transformation of the disclosure landscape in three years. This review tracks that shift across 6,066 filings from roughly 1,500 large-, mid-, and small-cap U.S. companies, measuring not just whether companies mention AI, but how much they say, how they frame it, and whether the language is growing more or less confident over time.

What was measured. For each filing, every passage substantively discussing AI, machine learning, or related technology was identified — approximately 26,000 passage-windows run through a two-stage extraction and classification process. Three core measures were then computed: breadth (share of companies disclosing at all), intensity (AI mentions per 1,000 words of filing text), and framing (whether language was positive, negative, forward-looking, risk-oriented, or workforce-related).

What was found. Breadth rose sharply across all company sizes. Intensity more than tripled among all filers. On tone, the picture is more complex: aggregate sentiment became substantially more cautious over the period, but this reflects who joined more than how incumbents changed. On workforce language, growth framing outpaces displacement framing roughly 5 to 1 — though a disclosure caveat applies throughout.

What this is, and is not. This review measures corporate disclosure, not corporate behavior. A company that says nothing about AI in its 10-K may be deploying it aggressively; one with rich disclosure may be mostly aspirational. The data captures what management chose to tell shareholders — a meaningful signal about perceived materiality and competitive positioning, but not a direct measure of AI investment, capability, or impact.

What the review covers

The review tracks AI-related language in annual 10-K filings submitted to the SEC by companies in the S&P 500, S&P 400, and S&P 600 — together, the S&P 1500. The trend window is fiscal years 2022 through 2025.

Corpus: 6,066 filings from approximately 1,500 companies. Of these, 3,711 (61%) contain at least one substantive AI passage; 2,355 contain none. The zero is genuine — those filings were scanned and found to contain no qualifying AI content, not skipped.

6,066

Total filings analyzed

3,711

Filings with AI content (61%)

~26,000

AI passage windows classified

Fiscal year assignment

When a company's fiscal year straddles two calendar years (e.g., a year ending March 31, 2023), it is assigned to the calendar year containing the majority of its reporting days. A March 31 fiscal year end is labeled FY2022 if the majority of days fall in 2022.

FY2021 exclusion

A small number of FY2021 filings are present in the underlying data but are excluded from all published trend analysis. The FY2021 cohort is sparse and its window counts are not comparable to later years due to an artifact of the filing-collection process. Every published trend begins at FY2022.

FY2025 partial snapshot

Partial year notice FY2025 data reflects filings captured as of May 31, 2026. Off-calendar filers whose fiscal year ends later in 2025 and whose 10-K had not been filed by the snapshot date are not yet included. FY2025 figures should be treated as directional — the direction is clear, but final levels will shift modestly when the full cohort is in.

How AI passages are identified — two-stage extraction

This review does not read every word of every 10-K for AI content. Instead, a two-stage process combines fast keyword scanning with LLM substantive review.

Stage 1 — Keyword scanning

Each filing is scanned for an expanded AI terminology lexicon covering seven categories: canonical AI terms ("artificial intelligence," "machine learning," "generative AI," "large language models," "LLMs," "ChatGPT," and the capitalized abbreviation "AI"); named AI models and brands (Copilot, Gemini, Claude, GPT-4, Grok, and others); cloud AI platforms (Bedrock, Vertex AI, SageMaker, Azure AI); AI company names (OpenAI, Anthropic, DeepMind, and others); technical AI terms ("deep learning," "neural network," "natural language processing," "foundation models," "transformer architecture"); AI hardware chips (H100, H200, A100, Blackwell, TPUs, Trainium); and emerging terminology ("GenAI," "agentic AI," "AI agents," "RAG," "prompt engineering").

For each keyword hit, a window of surrounding text (approximately ±200 characters) is extracted. Adjacent hits that overlap are merged into a single longer passage. Before these passages move to Stage 2, four mechanical filters remove obvious false positives: SEC exhibit-list patterns; person-name false matches (Watson, Bard, Claude, Gemini); geological false matches (Bedrock, Mistral); and data-science references with no AI-adjacent signal in surrounding context.

Stage 2 — LLM substantive review

Each passage surviving the mechanical filters is evaluated by Anthropic Claude Sonnet at temperature 0, via the Anthropic batch API. The model makes a single yes/no judgment: is this passage substantively about AI technology, AI investment, or AI workforce implications — or is it a false positive that escaped the keyword filters? Only confirmed passages enter the corpus, then tiled into uniform non-overlapping windows of approximately 700 characters each.

Validation Window inclusion was validated on a stratified 50-filing sample using two full independent raters — Anthropic Claude Sonnet 4.6 and Google Gemini 2.5 Pro — with GPT-4o serving as a partial third corroborator on the 10 highest-disagreement filings (scope-limited to the first 25K characters per filing). Production labeling was done by Sonnet 4.6 alone; the multi-model comparison was the audit, not an ensemble vote. Two-sided content coverage was 81.2% (Sonnet windows matched by Gemini 78.0%; Gemini by Sonnet 84.4%). On passages both raters did find, per-flag agreement was 99.1% (explicit AI terminology), 86.3% (descriptive AI content), and 85.4% (AI investment). The ~19% unique content reflects paragraph-boundary judgment, not classification disagreement. The hybrid approach recovers approximately 97.6% of the content a full document-level LLM read would find, at roughly 79% lower cost.

The core measure — AI intensity

The primary volume metric:

ai_intensity = (number of AI passage windows) ÷ (total filing word count) × 1,000

This measures AI mentions per 1,000 words of filing text. It is normalized by filing length so that a longer annual report doesn't automatically score higher. The metric is computed for every filing, including those with zero AI windows (which score zero, not null).

Two ways to read intensity

Intensity is reported two ways, and they answer different questions:

Measure	Denominator	What it answers	FY22→FY25
all_filings intensity	All 6,066 filings including zero-AI ones	The "average S&P 1500 company" including those that say nothing	0.019 → 0.126 (6× increase)
windowed-only intensity	Only the 3,711 filings with AI content	How much AI-disclosing companies say about AI, conditional on saying anything	0.075 → 0.139 (nearly 2× increase)

Important These two measures should never be presented as the same thing. The all-filings number reflects both who is disclosing and how much they say. The windowed-only number holds "who" constant and isolates "how much."

Breadth — adoption coverage

breadth = (filings with ≥1 AI window) ÷ (all filings)

The denominator is always the full filing count for that year — not just the filings that contain AI language. A breadth denominator that excluded non-disclosers would make adoption look artificially high.

Year	S&P 500 (large-cap)	S&P 400 (mid-cap)	S&P 600 (small-cap)
FY2022	34.1% (169/495)	23.2% (67/289)	18.5% (115/622)
FY2023	70.8% (356/503)	52.8% (168/318)	43.5% (274/630)
FY2024	86.9% (438/504)	79.3% (272/343)	66.2% (411/621)
FY2025 ⁽*⁾	95.2% (457/480)	90.3% (308/341)	87.0% (621/714)

(*) Partial cohort as of 2026-05-31. S&P 600 figures carry a ±1–4 percentage point soft band. Percentages rounded from stored breadth values.

The seven framing signals

Each AI passage window is tagged with up to seven independent yes/no signals describing how a company discusses AI, not just that it does. Click any signal to see its definition and real filing examples.

has_past AI already in production High reliability ▼

The passage describes AI the company already has deployed, completed, or integrated. Dominant verb patterns: "completed," "launched," "have deployed," "are using," "provides."

Company	Example
F5 Networks (FFIV)	In September 2025, we completed the acquisition of CalypsoAI Corp., which is expected to integrate into our F5 ADSP to create a complete solution for securing AI inference.
Alphabet/Google (GOOG)	In 2024, we launched Gemini 2.0, our most capable model yet.
FedEx (FDX)	Developments in data and technology, including artificial intelligence and machine learning, are facilitating the execution of our DRIVE transformation.
Moderna (MRNA)	Approximately 75% of our employees are currently active users, embedding the tool into their specific functions for customized support.
IQVIA (IQV)	Our 2024 AI employee program included training, a hackathon, and use-case-sharing — enabling participants to build confidence in using AI tools.

has_future_certain Firm, planned AI commitment High reliability ▼

The passage describes a firm, planned future AI commitment. Key triggers: "will," "plan to," "committed to," "roadmap," plus forward dates. No hedging language present.

Company	Example
Axon (AXON)	Beginning in the fourth quarter of 2024, we plan to introduce the Axon AI Era Plan.
NVIDIA (NVDA)	We plan to increase our U.S.-based manufacturing and invest in specialized equipment and processes to support domestic production.
Cognizant (CTSH)	We plan to continue to make significant investments in our AI capabilities to meet the needs of our clients.
IBM (IBM)	We are committed to an open innovation ecosystem around AI, to help our clients maximize flexibility.
Salesforce (CRM)	Agentforce is a complete AI system for building a digital labor force that can autonomously identify what work needs to be done and execute the plan.

has_future Hedged future AI activity Directional only Lower reliability ▼

The passage describes future AI activity with hedged or uncertain language: "may," "expect," "evaluate," "anticipate," "could." This is the weakest signal by inter-rater agreement — 17 of 60 validation windows had cross-model disagreement. Treat prevalence figures as directional, not precise.

Company	Example
Copart (CPRT)	We continue to evaluate emerging technologies like artificial intelligence, machine learning, and generative AI for incorporation into our business.
CDW Corporation (CDW)	We expect the competitive landscape to continue to evolve as new technologies emerge, such as solutions that incorporate AI.
Vertiv (VRT)	We believe that our future success will depend in part upon our ability to anticipate technology shifts, such as the growth in artificial intelligence.
Insulet (PODD)	While we anticipate being able to capitalize on opportunities using AI tools, doing so is not without risk.
Physicians Realty (DOC)	We may use generative AI tools in our operations. If our peers use AI tools and we fail to utilize AI in a comparable manner, we may be competitively disadvantaged.

has_positive Positive sentiment about AI High reliability ▼

The passage carries positive sentiment — benefit language, capability claims, competitive advantage framing, or value-creation statements. Distinct from opportunity framing: positive captures tone, not necessarily strategic intent.

Company	Example
Palo Alto Networks (PANW)	Our cybersecurity platforms help secure enterprise users by delivering comprehensive cybersecurity backed by industry-leading artificial intelligence and automation.
Rockwell Automation (ROK)	Our machine learning software enables customers to improve operational productivity and meet regulatory requirements.
Keysight Technologies (KEYS)	The group provides automated software test solutions that include AI-ML to automatically identify, build, and execute tests critical to digital business success.
AMD (AMD)	AMD is uniquely positioned to lead in this next computing era. Our broad portfolio provides the unique opportunity to make AMD the end-to-end AI leader.
Moderna (MRNA)	This rapid adoption across our company demonstrates the power of our platform.

has_negative Risk, harm, or competitive threat language High reliability ▼

The passage carries negative sentiment — risk language, harm, failure, liability, or competitive threat framing. Risk Factors is the dominant section for this signal, but it also appears in MD&A when companies discuss headwinds. Often co-occurs with has_risk_framing.

Company	Example
Capital One (COF)	There are significant risks involved in utilizing models and AI, and no assurance can be provided that our use will enhance our business. Generative AI has been known to produce false or 'hallucinatory' inferences.
Meta (META)	AI activities that threaten people's safety or well-being, or other societal harms or complications that could adversely affect our business, reputation, or financial results.
Fortinet (FTNT)	Vulnerabilities within our AI systems may be identified by competitors, researchers, or malicious actors before we detect or remediate them, resulting in reputational damage.
Verizon (VZ)	The use of AI by threat actors may increase the frequency and severity of cyberattacks against us.
Marsh & McLennan (MMC)	Litigation risks, liquidity and market volatility, an inability to obtain contractual limitations of liability for errors and omissions in connection with the use of AI.

has_risk_framing Structured risk disclosure Directional only Medium reliability ▼

The passage is explicitly structured as a risk disclosure — conditional harm language ("adversely affect," "no assurance," "may result," "could expose us") appearing in an Item 1A Risk Factors context. Second-weakest signal by inter-rater agreement. Use as directional; avoid precise cross-company comparisons.

Company	Example
Capital One (COF)	Higher interest rates may adversely affect the results of our operations and financial condition.
Progressive (PGR)	The growing development of agentic AI presents additional risks that may adversely affect our business. Advanced AI might produce datasets that are flawed, which could result in unfairly discriminatory outcomes.
Roper Technologies (ROP)	Challenges with properly managing AI use could result in reputational harm, competitive harm, and legal liability, and adversely affect our results of operations.
Fox Corporation (FOXA)	Our ability to compete could be adversely affected if our competitors gain an advantage by using AI. There can be no assurance that our efforts will be successful.
Motorola Solutions (MSI)	AI-related legislation may affect how our business is conducted or expose us to unfavorable developments resulting from changes in the regulatory landscape.

has_opportunity_framing AI as explicit business opportunity Highest reliability ▼

The passage frames AI explicitly as a business opportunity — growth, competitive positioning, revenue potential, or strategic advantage language. This is the tightest signal by inter-rater agreement: only 1 of 60 validation windows had cross-model disagreement. Prevalence figures are the most reliable of the seven.

Company	Example
Alphabet/Google (GOOG)	We continue this work as an AI-first company under the leadership of Sundar Pichai.
AMD (AMD)	Demand for our data center AI accelerator products was strong as large hyperscale customers deployed our AMD Instinct MI350X Series GPUs. We advanced our AMD AI GPU roadmap to deliver an annual cadence of leadership.
IQVIA (IQV)	IQVIA's portfolio delivers actionable insights and services built on high-quality health data, Healthcare-grade AI, advanced analytics, the latest technologies and extensive domain expertise.
Salesforce (CRM)	Agentforce is a complete AI system for building a digital labor force, integrating data, AI, automation, and humans to deploy trusted AI agents for concrete business outcomes.
C.H. Robinson (CHRW)	As a leader in Lean artificial intelligence supply chains, we deliver logistics like no one else.

Signal reliability — three-model blind test

Production signal tagging was done by a single model (Claude Sonnet 4.6, temperature 0, batched). Reliability was established by re-rating a stratified sample independently with models from three families. A 60-window blind test put the same passages through Anthropic Claude Sonnet 4.6, Google Gemini 2.5 Pro, and OpenAI GPT-5 using identical prompts — 60 windows × 7 signals = 420 binary cells. Labels were not produced by majority vote; the trio was the audit. Overall three-way per-cell agreement was 90% (377 of 420 cells). Pairwise agreement: Sonnet↔Gemini 94%, Sonnet↔GPT-5 95%, Gemini↔GPT-5 91%. Sonnet had the fewest lone dissents (6, vs. Gemini 20, GPT-5 17), which is why Sonnet's labels are the canonical reading. Per-signal disagreement ranged from 1 window (has_opportunity_framing) to 17 windows (has_future). A separate within-model drift test running 200 windows through Sonnet twice found 5.5% cell-level drift — consistent with inherent classification ambiguity in natural language, not systematic bias.

Sentiment measures

Metric	Formula	Range
Sentiment polarity	positive windows ÷ (positive + negative windows)	0 (all negative) → 1 (all positive)
Framing balance	opportunity windows ÷ (opportunity + risk windows)	0 (all risk) → 1 (all opportunity)

Both metrics are windowed-only — they cannot be computed for filings with no AI language.

The two-mechanism finding

Pooled sentiment polarity fell from 0.72 to 0.22 (FY2022→FY2025) and framing balance fell from 0.50 to 0.08. These are large drops. But they have a specific cause that must not be collapsed into a single behavioral narrative:

Mechanism 1 — Composition (dominant driver) Roughly 79% of companies that disclose AI at all first appeared in the disclosure pool in FY2023 or later. These later entrants arrived with sentiment around 0.16 and framing balance around 0.05 — already cautious — and stayed roughly flat. As they entered in large numbers, they pulled the aggregate down. The falling pooled average is primarily a reflection of who joined, not of how any given company changed its language.

Mechanism 2 — Incumbent reframing (real but secondary) The 335 companies that disclosed AI throughout the full FY2022–FY2025 window show genuine within-company reframing. Their sentiment polarity fell from 0.72 to 0.37, with the sharpest drop at FY2022→FY2023. But they remain well above the new-entrant floor (0.16) — convergence toward caution, not capitulation to it.

Publishing only the pooled decline (0.72→0.22) without this decomposition would misrepresent what happened.

Workforce and labor stance

Workforce signals

Signal	Definition
has_workforce_relevance	The passage mentions AI alongside workforce, employee, headcount, or labor language at all.
has_workforce_growth	The passage frames AI as growing, hiring, upskilling, or adding roles. AI is causally linked to positive workforce change.
has_workforce_displacement	The passage frames AI as reducing, eliminating, automating away, or restructuring roles. AI is causally linked to negative workforce change.

The continuous companion metric workforce_direction = growth windows ÷ (growth + displacement windows) runs at approximately 0.77–0.81 throughout the index window. This is a stable level, not a trend; it should not be presented as a year-over-year movement.

Labor stance categories

Category	Definition
not_mentioned	No AI-workforce link detected in the filing.
mentioned_neutral	AI mentioned alongside workforce language but neither growth nor displacement framing.
growth	AI framed as adding or developing headcount. AI-causal link required.
displacement	AI framed as reducing or replacing headcount. AI-causal link required.
both	The same filing uses both growth and displacement framing. Highest in Information Technology.

Key figures: growth framing rose from 2.7% to 16.3% of all filings (FY2022→FY2025). Displacement stayed low: 1.0% to 2.9%.

Mandatory disclosure caveat These figures measure what companies chose to include in SEC filings, not what they did. Growth framing is easy to volunteer — investors and employees respond well to it. Displacement framing typically requires regulatory or legal pressure before companies explicitly name it. The strong growth tilt partly reflects disclosure incentives, not only operational reality.

Relationship analyses

The review includes two relationship analyses linking AI disclosure patterns to forward business outcomes. These are descriptive associations, not causal claims.

Workforce outcomes

Companies using AI-growth framing in a given year (FY2022, 2023, or 2024) were compared to other AI-disclosing companies on their actual employee count change the following year. Growth framers grew headcount at a higher rate (56% grew headcount vs. 51% of non-framers; median change +1.1% vs. +0.3%). Both groups are AI-disclosing companies — this is not a comparison to companies that avoid AI entirely.

Earnings outcomes

Companies using positive or opportunity-framing AI language beat analyst earnings estimates at a higher rate (70.6%) than non-framers (63.3%). The has_future and has_risk_framing signals carry wider error bars on relationship estimates and should be treated as directional only.

Causal caution These associations do not establish that AI language caused better outcomes. Companies that talk about AI as an opportunity may already be higher-performing; companies under earnings pressure may reach for growth language strategically. The direction of the associations is informative; the magnitude should not be overread.

Index membership and comparability

Three membership bases

Basis	Definition	Best for
Point-in-time	Actual filing population for each year. Survivorship-free.	"How did the index evolve year by year?"
Continuous	Companies present in all four years. Like-for-like panel.	Skews to stable large-caps; not representative of the full index.
Balanced panel	335 CIKs with AI windows at both FY2022 and FY2025.	Sentiment reframing analysis. Stricter subset of continuous.

S&P 600 soft band FY2025 received a comprehensive filing sweep that earlier years did not. For the S&P 600 specifically, this added roughly 189 filings. Single-year S&P 600 levels carry a ±1–4 percentage point soft band as a result. Cross-year S&P 600 trends are valid; do not over-interpret single-year absolute levels for this tier.

What this review does not measure

It does not measure AI investment (capex, R&D allocation, or headcount in AI roles). It does not measure AI capability or competitive position. It does not measure what any company actually does with AI — only what management disclosed to shareholders in the annual report.

Silence is not evidence of inactivity. Disclosure is not evidence of execution. The data captures a specific, important, and measurable signal — corporate AI disclosure — and should be interpreted as that.

S&P 1500 AI Review 2022–2025Methodology & Data Guide