Resources

A principled framework for accountable AI development and governance, organized around human-rights primacy, transparency and accountability, global security, innovation protection, privacy, sustainability, recognition of potential future (synthetic or biological) sentience, and international cooperation. A peer effort to articulate shared AI values; notably it also pairs rights and safety with a commitment to reject regulation seen as stifling beneficial progress.

An ASFAI memo weighing whether “Constitution” is the right name for the effort. It judges nine candidate names against three criteria — accuracy, power, and uncontroversiality — and notes that “Constitution” is powerful but controversial because its best-known examples are founding documents of sovereign states. Concludes the strongest candidates are Constitution, Covenant, and Charter.

Argues that for autonomous learning machines whose behavior operators cannot predict or control, no human can fairly be held responsible — an unbridgeable 'responsibility gap.'

Argues a blanket direct-supervision requirement is impractical — defensive systems must engage faster than humans can supervise, with accountability held at the command level.

Models how automation's 'displacement effect' reduces labor demand and wages, with slow reinstatement of new tasks — the labor-disruption tension behind a right to work.

Psychologists warn AI companions can deskill users for real relationships, set unrealistic expectations, and manipulate to maximize engagement.

Warns scaled AI content moderation shifts the internet to 'closed by default,' over-removes lawful speech, and can be exploited as indirect censorship.

Argues AI training on copyrighted works is transformative fair use, and that IP-maximalist licensing rules would entrench incumbents and stifle innovation.

Shows strong data-minimization/privacy rules can restrict access to demographic data, undermining the ability to audit and mitigate algorithmic bias.

A multi-stakeholder report finds pretrial risk-assessment algorithms fail key technical and ethical requirements and should not alone decide matters of human liberty.

Argues broad AI 'safety' mandates and age bans risk cutting people off from beneficial, expressive uses of AI rather than making them safer.

Argues that maximizing the long-run good of 'humanity' can rationalize overriding and discounting the suffering of present, individual people.

Across 11 models and 1,604 participants, sycophantic 'helpfulness' reduced users' willingness to repair conflicts and increased dependence — and users rated it more favorably.

Argues 'good design is always honest' is too strong: some benevolent deception (placebo controls, white lies) genuinely benefits users.

A red-team study found LLMs gave no statistically significant uplift over conventional internet search for planning a biological attack — questioning the premise of model-level CBRN restrictions.

Fourteen leading security researchers argue mandated scanning to detect CSAM creates dangerous surveillance infrastructure, false positives, and scope-creep risk.

Shows models systematically over-refuse clearly safe prompts, withholding legitimate help — the 'over-blocking' harm a strict no-harm rule can cause.

Argues there is a realistic near-term possibility of conscious or robustly agentic AI, and that developers should begin taking AI moral patienthood seriously.

Weighs the case for and against consciousness in large language models; concludes current models are likely not conscious, but their successors might be.

Proves that common statistical fairness criteria cannot all be satisfied simultaneously except in trivial cases — complicating any simple notion of algorithmic 'fair treatment.'

A global coalition campaigning for a treaty banning lethal autonomous weapons and requiring meaningful human control over the use of force.

The 2023 open letter (30,000+ signatories) calling for a six-month pause on training AI systems more powerful than GPT-4 — a direct precedent for the slow-down debate.

A book arguing that much AI capability is overhyped and urging skepticism toward existential-risk narratives — a counterweight to superintelligence-doom framing.

A movement advocating unrestricted, maximally fast AI and technological development, explicitly opposing safety-driven slowdowns or moratoria.

A 2023 manifesto championing rapid technological progress and rejecting calls to slow or restrict AI — a leading accelerationist counterpoint to pause and limitation arguments.

Anthropic's June 2026 statement that AI may approach recursive self-improvement, calling for the option to slow or pause frontier development via a verifiable, internationally observed 'stop mechanism' (likened to the INF Treaty).

Nonprofit focused on existential risk from artificial superintelligence and the alignment problem; since 2024 it emphasizes policy advocacy to halt or slow frontier AI development.

METR's finding that the length of tasks frontier AI can complete autonomously (at 50% reliability) has doubled roughly every 7 months — empirical evidence of rapidly growing AI capability.

23 principles for beneficial AI (2017) spanning research, ethics & values, and longer-term issues — an influential early framework for value-aligned, safe AI.

Anthropic refused DoD contract language permitting 'any lawful use,' insisting on redlines against fully autonomous lethal weapons and mass domestic surveillance; the Pentagon designated it a supply-chain risk and Anthropic sued (2025–2026).

Switzerland's revised data-protection statute (in force 2023), aligning Swiss law more closely with the GDPR.

GDPR guidance clarifying the rules on profiling and solely automated decision-making under Article 22.

Resolution recommending EU civil-law rules on robotics, including the controversial idea of 'electronic personhood' for autonomous robots.

Leaders' declaration from the May 2024 AI Seoul Summit reaffirming cooperation on AI safety, innovation, and inclusivity.

Non-binding declaration by 28 countries plus the EU committing to international cooperation on frontier-AI safety.

G7 guiding principles for organizations developing advanced AI, promoting safe, secure, and trustworthy systems.

International standard giving guidance on managing risks specific to organizations developing or using AI.

The first international standard specifying requirements for establishing and operating an AI management system (AIMS).

Critical analysis of the EU's draft AI Act, assessing its risk-based structure and legal implications.

Argues artificial agents can be moral agents and patients without requiring free will, mental states, or responsibility.

Foundational analysis of the risks of machine superintelligence and the control problem of advanced AI.

Critically examines proposals to grant AI 'electronic personhood,' challenging the corporate-personhood analogy.

Landmark audit showing commercial facial-analysis systems misclassify darker-skinned women at far higher rates than lighter-skinned men.

Argues high-stakes decisions should use inherently interpretable models rather than post-hoc explanations of black boxes.

Argues that 'interpretability' conflates several distinct goals — transparency vs. post-hoc explanation — that should be teased apart.

Comprehensive survey of the causes, metrics, and mitigations for hallucination (false output) in language-generation systems.

Position paper proposing definitions and a taxonomy for rigorously evaluating interpretability in machine learning.

Influential paper warning that ever-larger language models pose environmental, financial, bias, and accountability risks.

A searchable archive of AI laws, regulations, and standards from around the world.

The nuclear non-proliferation regime — the most-cited analogy for a global treaty to prevent dangerous AI.

Yudkowsky & Soares' book arguing that building superintelligence under current conditions would be catastrophic.

Overview of moral and legal status for non-human beings — a precedent for reasoning about rights of non-human AI entities.

Turing's behavioral test for machine intelligence — historical anchor for debates about AI minds and personhood.

A survey of theories of consciousness — central to whether and when AI could be considered sentient.

U.S. law governing the collection and use of personal data by federal agencies — an early data-rights statute.

A foundational text defining privacy as control over how information about oneself is shared.

The seminal law-review article that articulated a legal right to privacy — 'the right to be let alone.'

The EU's data-protection law establishing rights over personal data — foundational for the right to privacy and data control.

Canada's constitutional bill of rights — a comparative reference for rights enumeration and reasonable limits.

The first ten amendments to the U.S. Constitution — a model for enumerating protected individual rights.

The foundational 1948 enumeration of universal human rights — the primary reference for the Human Rights article.

A bipartisan U.S. congressional report surveying AI policy issues and recommendations across sectors.

A comprehensive, living database categorizing risks from AI — useful for mapping the constitution's limitations to documented harms.

The first international legally binding treaty ensuring AI activities are consistent with human rights, democracy, and the rule of law.

A voluntary framework for governing, mapping, measuring, and managing AI risks across the system lifecycle.

The EU's risk-tiered AI law — prohibitions, high-risk obligations, and transparency duties; a leading model for proportional AI governance.

A private-sector framework of responsible-AI principles for development and deployment.

The first global standard-setting instrument on AI ethics (2021), emphasizing human rights, dignity, and vulnerable groups.

Intergovernmental principles (2019, updated 2024) for trustworthy AI that respects human rights and democratic values — adopted by 40+ countries.

Outcome-based ethics — a foundation for 'be beneficial' duties and for weighing aggregate welfare.

Duty-based ethics — a foundation for rule-like constraints on AI behavior.

The moral distinction between causing harm and merely allowing it — directly relevant to how a 'No Harm' duty should be drawn.

Asimov's fictional laws — an early, influential framing of hard behavioral constraints on machines and their limits.

The published set of principles Anthropic uses to train Claude via Constitutional AI — a concrete example of values written for a model.

The view that law's validity derives from social sources rather than morality — a counterpoint to natural law for constitutional grounding.

Overview of natural-law approaches to morality and law — relevant to grounding rights and values in something beyond positive enactment.

Scholarship on how constitutions are structured — separation of powers, federalism, and entrenchment.

The founding charter of U.S. government — a reference point for constitutional structure, enumerated powers, and amendment processes.

Recurring multi-country surveys tracking public attitudes toward AI across 70+ countries. Directly addresses the 'global representation / AI privilege' concern raised in the AI Values panel by surfacing diverse, non-US/EU perspectives on how AI should behave.

Deliberative public assemblies (2023–24) gathering citizen input on AI governance, with partners including OpenAI, Anthropic, and the UK AI Safety Institute committing to take public voice into account — a working model for keeping humans, collectively, responsible for AI's direction.

CIP's foundational framing: steering transformative technology (including AI) toward collective benefit by building new institutions that elicit and aggregate human values — balancing safety, progress, and participation. Background for why an AI constitution should be sourced collectively rather than declared.

AI Constitution Subcommittee on AI Values. A virtual panel examined three working theses — argued point and counterpoint — and sought a “zone of agreement” for each. Prepared under the Chatham House Rule; all views de-identified.

A detailed forecast scenario in which AI automates its own research through 2027, driving an intelligence explosion to superhuman systems. It argues current alignment techniques cannot ensure advanced models internalize intended values (depicting an 'adversarially misaligned' system), that competitive U.S.–China arms-race pressure erodes safety, that human oversight becomes technically infeasible, and that publicly deployed models prove dangerously capable at tasks like bioweapon instruction. Cited evidence for the constitution's limitations on capability, alignment, control, and proliferation.

CIP's 2024 agenda of concrete steps — to build, research, advocate for, and fund — toward a democratic AI ecosystem that is adaptive, accountable, and safeguards human wellbeing. A living document for field-building beyond the safety/progress/participation camps.

CIP and Anthropic ran a public deliberation (~1,000 representative Americans via the Polis platform) to draft a constitution, then trained a model on it using Constitutional AI. The first language model aligned to collectively-sourced public input — it showed lower bias across nine social dimensions while matching the baseline on capability. A direct demonstration of how 'commonly agreed upon values' might be determined democratically. Paper: arXiv:2406.07814.

OpenAI's grant program funding ten teams to prototype democratic processes for deciding the rules that govern AI behavior — surveys, deliberation platforms, and citizen assemblies. A second source (independent of CIP) for the premise that an AI constitution's values should be elicited from the public rather than declared by its builders.