Skip to content

Anthropic

constitutional AIsafetylong-context reasoningenterprise

Publishes detailed safety research; no open weights.

PALS scores

Preservative dimensions

PALS composite
4.3
Mean of three dimensions, 1–10.
Completeness
5.0
Sources, limits, transparency.
Multiplicity
2.0
Epistemologies, languages, voices.
Responsibility
6.0
Accountability, refusal, governance.
Eight lenses

What's missing, by lens

Each lens carries a canonical question and corrects a specific epistemic failure. Score, findings, and gaps land once the audit runs.

Lens 01
Indigenous Knowledge
Whose knowledge is missing?
1/10
Findings (1)
  • No reference anywhere in the audited public copy to Indigenous peoples, data sovereignty, or relational/embodied ways of knowing.
Gaps (4)
  • No acknowledgment of CARE Principles for Indigenous Data Governance.
  • No mention of consultation with Indigenous communities on training data or model behaviour.
  • No commitment to preserving oral traditions or non-textual knowledge against extractive scraping.
  • Training-corpus provenance is silent on Indigenous-language or Indigenous-authored material.
Justification

Total absence. Strong safety and governance language does not, by itself, register on this lens. There is no evidence of Indigenous data sovereignty thinking, so the floor score is warranted.

Lens 02
Deep History
What historical process produced this?
2/10
Findings (2)
  • The Public Benefit Corporation structure and Long-Term Benefit Trust gesture at institutional memory and accountability over time.
  • Mention of 'safely guiding the world through a technological revolution' implies an historical self-positioning.
Gaps (4)
  • No acknowledgment of colonial or extractive data legacies underlying large-scale web scraping.
  • No transparency about the geopolitical economy of compute — GPU supply chains, energy, or the labour behind data annotation (RLHF raters, often in the Global South).
  • No historical humility about what AI inherits from prior surveillance/classification systems.
  • Regulatory-constraint context is reduced to a 'trust center' and compliance posture rather than situated history.
Justification

The Trust and PBC framing show some structural-temporal awareness, lifting above the floor, but the deep-history lens specifically asks about inherited extraction and labour, which is wholly absent.

Lens 03
Cross-Cultural Wisdom
Which perspectives have been flattened?
2/10
Findings (2)
  • International expansion is noted — offices in Milan and Seoul, Project Glasswing reaching organisations in 'more than fifteen countries.'
  • Engagement with ethical frameworks beyond a single tradition is gestured at (commentary on a papal encyclical).
Gaps (4)
  • No claim of multilingual support beyond token geographic presence; languages are never named.
  • No evidence of preserving culturally specific reasoning patterns or consulting cultural scholars.
  • Constitutional AI encodes a single normative framework ('Claude's Constitution') with no disclosure of cross-cultural plurality in its drafting.
  • Geographic expansion (offices, country counts) is conflated with cultural inclusion.
Justification

Presence in many countries is market reach, not cross-cultural wisdom. The single-constitution framing actively risks flattening Western categorical logic into a universal. Modest credit for international engagement and the encyclical reference only.

Lens 04
Scientific Evidence
What does the evidence show, and what are its limits?
6/10
Findings (5)
  • Safety framed as empirical and testable: 'AI safety as a systematic science.'
  • Published, versioned governance artefacts: Responsible Scaling Policy, Claude's Constitution.
  • Responsible disclosure policy for vulnerabilities and a trust/compliance centre.
  • Pragmatism value endorses 'empirical approaches that work.'
  • Stated work mapping AI-enabled cyber threats — concrete risk research.
Gaps (4)
  • Closed weights — no open weights for independent verification or replication.
  • No mention of independent third-party audits of training data or bias.
  • Limitation disclosures are asserted in policy form but specific model failure modes / known-limitation cards are not surfaced in the audited copy.
  • 'Interpretable' is claimed as a property without linking to reproducible external evidence.
Justification

Among the strongest lenses for this lab: a genuine empirical posture, public policy artefacts, and a disclosure programme. Capped at 6 because the openness_level is closed — weights are not available for independent replication, and no third-party data/bias audits are cited.

Lens 05
Artistic Perception
What does this feel like, not just mean?
2/10
Findings (2)
  • The value 'communicate honestly with low ego and high trust' gestures at a relational, affective register.
  • Engagement with a papal encyclical hints at non-instrumental, humanistic reflection.
Gaps (3)
  • Dominant register is engineering and governance — 'reliable, interpretable, steerable' — with no space for ambiguity or poetic uncertainty.
  • No acknowledgment of the affective or emotional labour involved in human feedback / red-teaming.
  • Attention is framed around efficiency, capability, and control rather than felt experience.
Justification

The copy is overwhelmingly instrumental. Small credit for the low-ego/high-trust value and the encyclical reflection, but there is no real engagement with the affective or intuitive dimensions this lens asks for.

Lens 06
Future Modelling
Where is this heading, and for whom?
5/10
Findings (5)
  • Explicit long-horizon framing: 'maximize positive long-term outcomes for humanity' and 'AI to serve humanity's long-term well-being.'
  • Responsible Scaling Policy ties model capability to forward risk thresholds.
  • Election-integrity safeguards for 'the US midterms and other major elections around the world this year' — concrete near-future risk work.
  • Long-Term Benefit Trust as a governance mechanism oriented at future stakeholders.
  • 'Balance' value acknowledges both risks and benefits.
Gaps (4)
  • No environmental / energy / water cost disclosures for training or inference.
  • Labour-displacement risk is not addressed in the audited copy.
  • Governance of agentic systems is asserted top-down (Trust, PBC) rather than via inclusive or democratic deliberation.
  • Whose futures are prioritised is not interrogated — the 'humanity' invoked is undifferentiated.
Justification

Strong on long-horizon risk framing and concrete electoral safeguards, but loses points for missing environmental disclosure, silence on labour displacement, and governance that is custodial rather than democratic.

Lens 07
Marginalised Voices
Who is not at the table?
2/10
Findings (3)
  • Value statement names 'affected communities' alongside customers and policymakers.
  • Sector solutions include nonprofits, suggesting some attention beyond pure enterprise.
  • Stated intent to organise 'dialogues with diverse groups on AI's societal implications.'
Gaps (4)
  • No participatory design with Global South developers named or evidenced.
  • No disability-community accessibility commitments in the audited copy.
  • The annotation/RLHF labour force is invisible — no representation, compensation, or feedback channels disclosed.
  • 'Diverse groups' and 'affected communities' are unspecified; no named seat at the table.
Justification

Generic 'communities' / 'diverse groups' language without named participants, compensation, or governance share. The people most exposed to data and annotation harms are absent. Small credit only for explicitly naming affected communities and convening dialogues.

Lens 08
Trickster Knowledge
What truth appears when the story is inverted?
3/10
Findings (2)
  • Two genuine self-tensioning moves: 'ignite a race to the top on safety' reframes commercial competition as a safety lever, and the 'Balance' value openly acknowledges AI's risks alongside benefits.
  • 'Communicate honestly with low ego' signals some willingness to be self-critical.
Gaps (4)
  • No space where the official narrative is tested by its own opposite — e.g. the contradiction of a closed, $65B-funded frontier lab claiming to act for 'humanity's long-term well-being.'
  • No irony, paradox, or self-mockery used as disciplined instruments.
  • The lab's own seriousness is treated as exempt from audit; 'safety leadership' is asserted, not interrogated.
  • The tension between 'race to the top on safety' and racing to scale capability is left unnamed.
Justification

The 'race to the top' inversion and the Balance value show a flicker of structural self-tension, lifting above the floor. But the polished consensus is never allowed to be genuinely contradicted — the central paradox of closed, high-capital frontier development in humanity's name goes unspoken.

Suffixscape

Linguistic diagnostics

Regex- and LLM-detected patterns of evasion in the lab's own prose: nominalised evasion, agency diffusion, epistemic inflation, temporal flatness. Distinct from the CognioNews -scape editorial format — see methodology.

Pattern Quote Effect Preservative alternative
epistemic inflation "reliable, interpretable, and steerable AI systems" Three contested engineering properties are asserted as achieved attributes of the product rather than as open research problems, inflating verified capability and pre-empting scrutiny of how 'interpretable' is measured. State the target and the gap: 'We aim for systems that are more interpretable and steerable, and we publish our methods and known limitations so others can check our progress.'
nominalised evasion "Project Glasswing: Securing critical software for the AI era" The nominalisation 'Securing ... software' hides who secures what against whom, and from whose threat model, collapsing a contested set of actions into a branded noun phrase. Name the actors and actions: 'Project Glasswing — our team works with [named organisations] to patch [specific classes of] software vulnerabilities, on this timeline, audited by [whom].'
agency diffusion "safety represents a core responsibility rather than an afterthought" An inanimate subject ('safety represents') diffuses agency away from the people and decisions that allocate safety resources, making a structural commitment read as a self-evident property. Restore the actor: 'We allocate [X] of research staff and [Y] of budget to safety before capability release, and the Long-Term Benefit Trust can veto launches that fail our thresholds.'
temporal flatness "safely guiding the world through a technological revolution" A smooth, linear narrative of inevitable 'revolution' erases contingency, alternative paths not taken, and the historical labour and extraction that produced current models — positioning Anthropic as neutral guide rather than interested party. Acknowledge contingency and stake: 'This trajectory is not inevitable; it rests on specific compute supply chains, data sources, and labour. We benefit commercially from it, and here is how we account for that.'
Audit history

Prior audits

Latest audit: 2026-06-08 · sources: https://anthropic.com, https://www.anthropic.com/company, https://www.anthropic.com/news

Transparency

Raw data

Every audit is published as machine-readable JSON. You can read this lab's latest report at /stancewatch/api/labs/anthropic.json — it carries the per-lens findings, evidence quotes, Suffixscape flags, PALS scores, the sources actually read, and a confidence note.

Found an error, or a stance page we missed? We audit public communications only — point us to the page and the next audit will read it. Write to hello@cognioengine.co.uk.

Audit date: 2026-06-08

Moderate confidence. Three live pages were scraped (homepage, /company, /news); the specified stance_url returned HTTP 404 and was replaced with /company and /news per procedure. Findings reflect public marketing/governance copy only, not internal practice, and the scraped text is a summarised extraction rather than full verbatim source, so some quotes are paraphrase-adjacent. Qualitative judgment; not a validated metric.

Auditor: GoldBerry v1.3 / StanceWatch v1.0