Redistributed

Document: Claude 3 Model Card and System Cards (Claude 3 through Claude 4.6) Source: Anthropic System Cards

Anthropic positions itself as the safety-focused AI lab — the company building AI "the right way." Model cards exist to help outsiders evaluate those claims. So what do Anthropic's model cards actually reveal?

The answer is a split screen: excellent safety documentation alongside minimal infrastructure transparency. Claude's cards genuinely lead the industry on red teaming, safety evaluations, and alignment research. They go quiet on training data, labor conditions, and environmental costs. Understanding both sides — what Anthropic does well and where it falls short — matters for anyone trying to hold the AI industry to account.

What Model Cards Are Supposed to Do

Before examining Claude's documentation, we need to understand the standard it's measured against.

Model cards were introduced by Mitchell et al. (2019) at Google as "short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions." The framework calls for:

Model details (architecture, training procedure)
Intended use and out-of-scope applications
Training data composition and sources
Performance across demographic groups
Limitations and ethical considerations
Quantitative evaluation metrics

Gebru et al.'s (2021) Datasheets for Datasets standard goes further on training data, calling for documentation of:

Data sources and collection methods
Data composition (languages, domains, demographics)
Preprocessing and labeling procedures
Who did the labeling work and under what conditions
Consent and privacy considerations
Maintenance and updates

These frameworks were designed as accountability tools — ways for developers to communicate what they know (and don't know) about their systems, and for users to make informed decisions. For more on how these standards work in practice, see What Model Cards Tell You and Datasheets for Datasets.

How does Claude's documentation measure up?

What Anthropic Discloses

Safety Evaluations: The Strong Suit

Claude's documentation shines on safety testing. The cards describe:

Risk categories: CBRN (chemical, biological, radiological, nuclear), cybersecurity, autonomous AI capabilities, and alignment issues. This taxonomy is clearly defined and consistently applied across model versions.

Red teaming approach: Anthropic describes multi-attempt attack testing using 200-attempt reinforcement learning campaigns — a more rigorous methodology than single-attempt testing. They monitor approximately 10 million neural features during evaluation, mapping to human-interpretable concepts including deception, sycophancy, bias, and power-seeking.

External partnerships: The cards name specific external evaluators — Apollo Research for alignment, Deloitte for security, the National Nuclear Security Administration for nuclear risk, and the US and UK AI Safety Institutes. This is unusually specific.

Mitigation efficacy: The cards report before-and-after measurements showing how specific interventions changed model behavior. This moves beyond "we did safety work" to "here's what the work accomplished." Other labs should study this approach — it represents a genuine best practice that the AI Index Report 2025 and the Stanford FMTI both recognize.

The Stanford Foundation Model Transparency Index (FMTI) confirms this strength. In their 2024 assessment, Anthropic scored highest on safety disclosure categories while performing poorly on upstream (data) transparency.

Capabilities Documentation

The cards document what Claude can do:

Input/output modalities (text and images in, text out for Claude 3+)
Context window sizes and pricing tiers
Benchmark performance (MMLU, HumanEval, GPQA)
Comparative analysis against previous Claude versions

The Claude 3 card explicitly positions the three-tier system: Opus for maximum capability, Sonnet for balanced performance, Haiku for speed and cost. This tiered approach reflects the broader Pricing Power dynamics shaping how AI capabilities reach different users.

Usage Policies

Anthropic publishes acceptable use policies, terms of service, and privacy policies with in-app opt-out notices. Users know what's prohibited and have mechanisms for opting out of certain data uses.

What Anthropic Withholds

Training Data: The Black Box

The Claude 3 model card states:

"Claude 3 models are trained on a proprietary mix of publicly available information on the Internet as of August 2023, as well as non-public data from third parties, data provided by data labeling services and paid contractors, and data we generate internally."

That's the entirety of the training data disclosure. We learn:

Some internet data (unspecified sources, unspecified proportion)
Some licensed third-party data (unnamed parties, unknown terms)
Some contractor-generated data (unnamed contractors, unknown tasks)
Some synthetic data (unknown methods, unknown proportion)

We don't learn:

The size of the training dataset
The language composition
The domain distribution (how much news? how much code? how much academic papers?)
The top data sources
The selection criteria
The deduplication and filtering methods
What copyrighted material may be included
What personal information may be included

The FMTI assessment gave Anthropic 7/51 points on upstream (data) transparency. The single point came from mentioning "deduplication and classification" as curation methods.

The official justification: "proprietary and not disclosed publicly to protect competitive advantages."

This is a choice. Training data transparency is not technically impossible — it's competitively inconvenient. That said, the competitive concerns are real: in a market where AI Market Concentration already advantages a few players, unilateral disclosure carries genuine risk. The question is whether the industry can find mechanisms — regulatory floors, mutual commitments, or third-party audits — that make disclosure a shared obligation rather than a unilateral sacrifice.

Labor Conditions: Mentioned, Not Specified

Anthropic references data annotation in the cards, noting they use "data labeling services and paid contractors." But what does this mean in practice?

The FMTI found that Anthropic's documentation mentions "crowd worker wellness standards and fair compensation commitments" and notes partnerships "with platforms aligned with our belief in providing fair and ethical compensation."

What's missing:

Specific compensation amounts
Countries where workers are located
Hours, quotas, and working conditions
Which labeling platforms are used
Worker protections and complaint mechanisms

We know the work happens. We don't know who does it, where, for how much, or under what conditions. The Ghost Work review and Hidden Labor explainer document why this matters — the people building AI systems deserve the same visibility as the systems themselves.

Compare this to the detail companies provide about their infrastructure partnerships (AWS, Google Cloud are named) or their evaluation partners (Apollo Research, Deloitte are named). Labor gets vague language; compute gets specifics. To be fair, Anthropic's mention of "fair compensation commitments" is more than some competitors offer — but commitments without data are promissory notes, not evidence.

Environmental Impact: Absent

Claude's model cards do not report:

Compute usage for training (explicitly marked "proprietary")
Energy consumption in megawatt-hours
Carbon emissions in tons CO2
Water usage for cooling

The FMTI gave Anthropic 0/4 on all environmental metrics.

Anthropic does claim to "fully offset its operational carbon emissions each year." But offset claims are cheap without baseline numbers. How much is being offset? We don't know.

Third-party estimates suggest Claude 3 Opus uses approximately 4.05 Wh per request with about 1.80 grams of CO2 per query. But these are inference costs, not training costs. Training emissions — likely orders of magnitude larger — remain undisclosed. For context on why this matters at scale, see Environmental Cost and Data Center Energy.

Model Architecture: Trade Secret

The cards do not specify:

Model architecture details
Model size (parameter count)
Training duration
Specific training techniques beyond Constitutional AI

The Claude 3 card describes Constitutional AI at length but treats the underlying model as a black box with alignment techniques applied on top.

Evolution Across Claude Versions

Anthropic has released model cards for Claude 1 (March 2023), Claude 2 (July 2023), Claude 3 (March 2024), and subsequent versions through Claude 4.6 (February 2026). How has documentation evolved?

Improved:

Safety evaluation depth has increased substantially
External partnerships are more specifically named
Red teaming methodology is more rigorously documented
Capabilities testing has grown more comprehensive

Unchanged:

Training data opacity remains constant
Labor conditions remain vague
Environmental impact remains unreported
Architecture details remain withheld

The pattern suggests intentional policy: safety transparency increases because it aligns with Anthropic's positioning; infrastructure transparency stays opaque because disclosure carries competitive and legal risk. This is not unique to Anthropic — it reflects the structural incentives facing all Frontier Lab Oligopoly members. The question is whether the improvements on the safety side represent a genuine trajectory toward broader transparency, or whether they've found a stable equilibrium where some disclosures substitute for others.

Comparison to Other Labs

How do Claude's cards compare to competitors?

OpenAI (GPT-4)

OpenAI's GPT-4 System Card is similarly strong on safety and weak on data. It describes extensive red teaming, safety mitigations, and risk categories. But training data? "GPT-4 was pretrained on publicly available data (such as internet data) as well as data we've licensed."

Neither company names specific datasets, proportions, or sources.

OpenAI does provide something Anthropic doesn't: comparative analysis between early and launch versions showing specific safety improvements. The GPT-4 card compares "GPT-4-early" (minimal mitigations) against "GPT-4-launch" (post-safety work), with specific examples of changed behavior.

Meta (Llama)

Meta's Llama cards are more structurally complete on some dimensions. They provide:

Data mix percentages (not sources, but proportions)
Training infrastructure details
Explicit intended use cases

But Meta also omits specifics on data sources and labor conditions. The Llama cards are more forthcoming about architecture while remaining opaque about provenance.

Google (Gemini)

Gemini's documentation includes more specific categories of training data: "data obtained by crawlers, licensed data obtained via commercial licensing agreements, user data collected from users of Google products and services." This is categorical transparency without source-level specificity.

Google also documents data processing methods — deduplication, safety filtering, quality filtering — with more specificity than Anthropic.

The Common Pattern

All major labs:

Document safety evaluations extensively
Provide benchmark performance
Describe acceptable use policies
Omit specific training data sources
Obscure labor practices
Ignore environmental costs

Safety documentation is competitive — it supports marketing claims. Infrastructure transparency is not — it enables scrutiny. This dynamic is why external accountability efforts like the AI Index Report 2025 and the Stanford FMTI matter so much: they create reputational incentives where market incentives are absent.

The Redistribution Angle

Model cards were designed as accountability tools. Incomplete cards limit accountability. What does Claude's documentation pattern mean for redistribution?

Disclosure Reflects Power

Anthropic chooses what to disclose. Those choices reflect priorities:

Safety performance? Extensively documented (supports brand positioning)
Evaluation partnerships? Named specifically (builds credibility)
Training data? "Proprietary" (avoids legal exposure, competitive risk)
Labor conditions? Vague commitment language (avoids scrutiny)
Environmental costs? Absent (avoids accountability)

The pattern is common across all major labs: disclose what helps, defer what hurts. Understanding this pattern is not a reason for cynicism — it is a reason to build better incentive structures.

Documentation Labor Shifts to Critics

When companies don't document, critics must. The Stanford FMTI team spent months manually gathering information that companies could provide in hours. Journalists, researchers, and civil society organizations do the accountability work that documentation should enable.

This is a redistribution of labor. Companies externalize the cost of transparency onto the public.

Users and affected communities can't make informed decisions without information. What data was Claude trained on? If you're a writer, artist, or programmer, you can't know whether your work was used. If you're deploying Claude, you can't audit for biases rooted in training data. If you're a regulator, you can't assess environmental compliance.

Claude's documentation gives you enough to use the product, not enough to evaluate its production.

The Standards Exist

Mitchell et al. (2019) and Gebru et al. (2021) defined what responsible documentation looks like. These aren't radical demands — they're the accepted standards proposed by researchers at Google itself.

Anthropic could move closer to these standards. The gap between what's possible and what's disclosed is a policy choice, not a technical constraint. Regulatory frameworks like the EU AI Act may eventually make this choice for them — but the labs that move first will help shape what compliance looks like.

What This Means for You

If you're deploying Claude:

You have good information about capabilities and safety evaluations
You don't have information to audit training data, assess biases, or evaluate supply chain ethics
You're trusting Anthropic's internal processes without external verification

If you're affected by Claude:

You can't determine if your work was used in training
You can't assess labor conditions in the production chain
You can't evaluate environmental claims

If you're evaluating AI companies:

Anthropic's safety documentation exceeds industry norms
Their data and infrastructure transparency matches industry norms (which are low)
"Better than others" is not "sufficient"

If you're advocating for AI accountability:

Push for training data disclosure beyond "proprietary mix"
Demand labor condition specifics, not vague commitments
Require environmental reporting as standard practice
Support the Stanford FMTI and similar accountability projects

The Bottom Line

Anthropic's model cards tell you what Claude can do and how Anthropic tests for safety. They don't tell you what Claude was built from, who did the building, or at what environmental cost.

This is selective transparency — detailed where it supports marketing claims, absent where it enables scrutiny. The documentation is strong relative to competitors and weak relative to established standards.

Model cards were invented to enable informed decisions about AI deployment. Claude's cards enable informed decisions about capability and safety. They do not enable informed decisions about ethics, sustainability, or accountability.

"Safety-focused" accurately describes Anthropic's deployment practices. It does not yet describe their full disclosure practices. The gap represents both a challenge and an opportunity — Anthropic's demonstrated ability to lead on safety documentation suggests they could lead on infrastructure transparency too, if the incentives aligned.

What Model Cards Tell You — The broader framework for evaluating model documentation
Datasheets for Datasets — The standard Anthropic's data disclosure falls short of
Llama License — How Meta's documentation compares
Gemini's Safety Documentation — Google's parallel approach to safety-first disclosure
AI Market Concentration — Why competitive dynamics shape transparency decisions
Ghost Work — The labor conditions model cards should but rarely document
Frontier Lab Oligopoly — The structural context for industry-wide opacity
EU AI Act Explained — The regulatory pressure that may force broader disclosure

Sources

Anthropic. "Claude 3 Model Card." March 2024. PDF
Anthropic. "System Cards." Accessed March 2026. Anthropic
Bommasani, Rishi et al. "The Foundation Model Transparency Index." Stanford CRFM, May 2024. Paper
Stanford CRFM. "Anthropic: Claude 3 Transparency Report." May 2024. Report
Stanford CRFM. "Anthropic Transparency Report." December 2025. Report
Mitchell, Margaret et al. "Model Cards for Model Reporting." FAT Conference, 2019. arXiv
Gebru, Timnit et al. "Datasheets for Datasets." Communications of the ACM, 2021. arXiv
OpenAI. "GPT-4 System Card." 2023. PDF
Meta. "Llama 3 Model Card." 2024. GitHub

General

Doc Review: Claude's Model Card