How Many Artificial Intelligence Are There: A Measurement Guide

When people ask how many artificial intelligence are there, they usually want a simple number. The true answer depends entirely on what you count and how often it changes. Media reports often cite large figures with zero methodology. Practitioners need defensible, replicable counts. We must separate models, deployed systems, agents, companies, and research outputs.

We define five distinct counting lenses below. We also show current indicators for each category. This independent analysis focuses heavily on multi-agent systems. Every figure we present includes a date and a verifiable source. You can track these shifts through our latest independent coverage of multi-agent AI systems.

This framework provides a clear way to measure the field. You can stop relying on vague vendor claims. You can build your own tracking methodology.

Defining What Constitutes an AI System

No authoritative registry exists for these technologies. This absence causes inconsistent counts across news reports and market studies. To build a defensible count, we must define specific boundary rules. We separate the field into five clear categories.

• Lens A: Model families versus variants, checkpoints, and fine-tunes.
• Lens B: Deployed software systems running in production environments.
• Lens C: Agents operating within multi-agent applications.
• Lens D: Organizations, companies, and research labs building these tools.
• Lens E: Research outputs including papers, datasets, and code repositories.

Each category requires a different measurement approach. Conflating these lenses leads to wildly inaccurate totals. A single company might publish ten papers and release fifty model checkpoints. We cannot count all of these as a single metric.

Tracking Current Counts Across Five Analytical Lenses

Each lens requires a specific measurement method. We provide current indicators and replication steps below. You must apply these methods consistently to track growth over time.

Lens A: Foundation Models and Variants

We define this category as foundation model families and notable public variants. You can track these numbers using public model catalogs and official releases. Platforms like Hugging Face host hundreds of thousands of files. Most of these files are minor variations rather than distinct systems.

To replicate this count, snapshot a major catalog. Record the exact date of your query. You must filter your search carefully.

• Check the total number of foundation models.
• Filter out minor fine-tunes to avoid double counting.
• Exclude inactive or deprecated checkpoints.
• Remove duplicate uploads from unofficial accounts.

Forks and fine-tunes heavily inflate these numbers. Private models remain largely uncounted in public catalogs. You must state your exclusion criteria clearly.

Lens B: Deployed AI Systems in Production

This category covers productized features, APIs, and internal services. These represent tools actively working in business environments. A deployed system differs entirely from a raw model weight file.

You can measure this by tracking vendor documentation and changelogs. Monitor product release notes from major software providers. Sample the top vendors and document your inclusion criteria.

1. Identify the top fifty enterprise software vendors.
2. Scan their official release notes for new autonomous features.
3. Count discrete products rather than underlying API calls.
4. Log the date of each feature deployment.

Private deployments make exact counting difficult. Internal corporate tools rarely appear in public documentation. You can only estimate the private enterprise market.

Lens C: Agents in Multi-Agent Applications

This lens focuses on autonomous agents instantiated per task. It also includes persistent agents within orchestration frameworks. This represents the most fluid category of measurement.

You must analyze specific frameworks and estimate agents-per-task multipliers. Ephemeral agents are event-driven and unbounded. They exist only for the duration of a specific task. Read more about these frameworks in our recent multi-agent AI news and analysis.

Count these as rates rather than static numbers. A single application might spin up thousands of temporary agents daily.

• Track the number of active orchestration frameworks.
• Estimate the average number of agents per task.
• Calculate the daily instantiation rate.
• Separate persistent manager agents from temporary worker agents.

Lens D: Organizations Building AI

This category counts companies, labs, and research groups. We look for entities with active, verifiable development. This metric helps track industry consolidation and growth.

Use public firm lists and open datasets with disclosed methodologies. Look for organizations publishing code or releasing models. Financial databases provide a good starting point for this count.

Varying definitions of a technology company create counting challenges. Datasets often lag behind actual market conditions.

• Query startup databases for specific machine learning tags.
• Filter out companies with no shipped products.
• Verify active academic labs through university directories.
• Cross-reference company lists with public code repositories.

Lens E: Research Outputs

Research outputs include papers, datasets, and code repositories. These metrics show the raw academic and engineering momentum. They serve as leading indicators for future commercial products.

Watch this video about how many artificial intelligence are there:

Track publication submission categories and major conference proceedings. Monitor code platforms for topics related to machine learning.

1. Count primary paper submissions per month.
2. Track active open source repositories.
3. Monitor new dataset releases on public platforms.
4. Record the number of accepted conference papers.

Preprint duplication can skew paper counts. Inactive code repositories often inflate open source numbers. Always filter for activity within the last ninety days.

How to Replicate and Update These Measurements

You can reproduce these counts using transparent methods. Follow these steps to build your own tracking system. We recommend updating your counts quarterly.

Define your exact scope and exclusion criteria first. Write simple scripts to pull counts from public APIs. Create a methodology appendix with dates and sources.

• Use official API endpoints instead of web scraping.
• Store snapshot data with exact timestamps.
• Document every exclusion rule you apply.
• Share your raw data alongside your final counts.
• Verify your numbers against primary sources only.

Report dynamic agent counts as hourly or daily rates. A static number fails to capture the reality of multi-agent systems. Always document your exact search parameters. This transparency allows other researchers to verify your findings.

Avoiding Common Counting Failures

Many industry reports fail to provide accurate numbers. They fall into several predictable traps. You must actively avoid these errors when building your methodology.

Double counting represents the most common failure. Researchers often count a single model multiple times. This happens when a model appears in different catalogs.

• Do not count a fine-tuned model as a new foundation model.
• Do not count an API wrapper as a distinct intelligence.
• Do not include abandoned code repositories in your totals.
• Do not accept vendor marketing numbers without verification.

You must apply strict filtering to your datasets. Look for duplicate entries across different platforms.

Frequently Asked Questions

How many artificial intelligence are there globally?

No single number exists. The total depends on whether you count foundation models, deployed software features, or active research projects. You must define your counting lens first.

Why do different reports show conflicting numbers of AI systems?

Reports use different measurement criteria. Some track every minor code fork. Others only count major foundation models. This creates massive discrepancies in published figures.

How do we measure autonomous agents?

We measure agents as a rate of instantiation. A system might create hundreds of temporary agents to solve one complex task. Static counts do not work for these dynamic systems.

Establishing a Reliable Measurement Framework

Counting these systems requires strict boundaries and clear definitions. You cannot rely on a single static number. The technology moves too fast for simple tallies.

• A five-lens framework yields defensible measurements.
• Dynamic agent counts require rate-based reporting.
• All figures need a declared date and scope.
• Fine-tunes and forks require careful filtering.

You can now compute your own counts with transparent methods. This approach removes the ambiguity from vendor claims. You can build a rigorous tracking system.

Learn more about MAIN’s independent editorial mission. This framework was developed by the authors behind MAIN.