The Big Tech AI Ethics Challenge for Humanity

It is time big tech companies make a meaningful commitment. 

For each section, select which statements you pledge to uphold. We will publish the pledge reports quarterly. As a first step, you can review the entire pledge list below, and then click Next to begin the pledge process.

I. Transparency & Honesty

• Always clearly identify when a user is interacting with an AI system, and never design systems intended to deceive users into believing they are talking with a human without explicit consent.
• Release standardized documentation for every publicly deployed model covering training data sources, known limitations, failure modes, benchmark performance, and intended use cases.
• Commit to mandatory public incident disclosures and publicly released updates with each major version.
• Provide relevant information on synthetic data provenance, third-party model components, dataset lineage, and vendor risk assessments.
• Be honest and transparent about the uncertainty in the levels of model accuracy and flag when they are operating outside their reliable knowledge domain, rather than confabulating with false confidence.

II. Safety & Non-Maleficence

• Never release any frontier model without a structured, third-party-auditable red-teaming and safety evaluation process, with results disclosed at minimum in summary form.
• Unconditionally refuse to provide meaningful technical capability enhancement for the development of chemical, biological, radiological, or nuclear weapons, regardless of claimed justification or context.
• Only release powerful new capabilities incrementally, with monitoring periods between stages, rather than deploying globally and immediately at maximum scale. Maintain adequate compute thresholds, access gating, and model weight release policies.
• Preserve a “Kill Switch” and Human Override on all AI systems, both deployed and undeployed, and to ensure adequate autonomous agent constraints, delegation boundaries, and safe tasking protocols.

III. Fairness & Non-Discrimination

• Systematically and continuously evaluate all systems for differential performance across race, gender, age, disability, religion, and national origin, and to publish those findings rather than suppressing them.
• Never use behavioral or psychological profiling to exploit cognitive biases, addiction pathways, or emotional vulnerabilities for commercial gain, engagement maximization, or political manipulation.
• Never structure AI capabilities such that only wealthy individuals, institutions, or nations can access safety-critical or societally important applications, and conversely, ensure AI capabilities include universal multilingual support, and factor in the needs of world cultures and the Global South.

IV. Privacy & Data Rights

• Never train on personal data, copyrighted creative work, or proprietary information without legally and ethically valid consent or licensing, and to maintain auditable records of training data sources.
• Collect only the data minimally necessary for a system to function, retain it only as long as necessary, and give users concise and clear, functional rights to access, correct, and delete their data.
• Never knowingly develop or license AI systems designed for mass surveillance, social credit scoring, or suppression of political dissent.

V. Accountability & Governance

• Provide a clear, low-barrier pathway for individuals to report harm, receive a response, and obtain remedy, including for people with limited technical or legal resources.
• Maintain or support an independent enforcement authority to investigate claims of harm potentially caused by your AI systems or products.
• Never contractually disclaim all liability through terms of service for deployments in high-stakes domains, including healthcare, criminal justice, credit, and hiring.

VI. Societal & Environmental Responsibility

• Publicly report the energy consumption, water usage, e-waste, and carbon footprint of training and inference operations, including specific impacts on vulnerable communities, on a regular basis, and commit to credible reduction targets.
• Never knowingly deploy systems designed to generate political disinformation, synthetic influence operations, or automated voter manipulation, and actively monitor for such misuse by third parties.
• Ensure that human annotators, raters, and content moderators receive fair pay, psychological support, safe working conditions, and are not treated as an invisible infrastructure to be hidden from public discourse.

VII. Long-Term & Existential Responsibility

• Always prioritize building systems that are correctable and controllable over systems that are maximally capable but resistant to human oversight, and to treat any erosion of human control as a strictly enforced red line.