Zero Vector AI Use Policy

This policy covers our team and the work we produce individually and collectively. This policy also establishes requirements for the activities of Zero Vector independent collaborators, without attempting to govern the work they do that is unrelated to Zero Vector.

Why We Use AI

AI's ability to increase the capacity and effectiveness of organizations in areas ranging from financial modeling and content production to research and software development provides organizations with undeniable operational benefits. It is also critical for us to run highly effective operations with a lean team. We therefore use AI across all of our organization's operations because it enables smaller teams to accomplish more, faster, with less.

How We Use AI

We use AI as a core component of our talent capability development, problem identification and validation, research, product and venture development, marketing and communications, and internal operations. AI is used to support the development of our programs, conduct critical analysis, develop and implement assessments, plan and conduct critical research, build our internal tools and platforms, build external software and platforms, and draft and publish content.

Much of our work has standing approvals for AI use and doesn't require prior review or approval. For work with standing approvals, we have identified the permitted purposes, accounts, data, actions, and limits. Routine, reversible work operating within the limits of standing approvals does not need approval before AI is used. Non-routine, non-reversible, consequential work requires review and approval before using AI.

When determining which work can have standing approvals, we assess consequences, data sensitivity, permissions, scale, uncertainty, and reversibility.

Automation Requirements

• +All automations must have accountable owners, spending and access limits, monitoring, stopping conditions, recovery plans, and permissions must be enforced outside of prompts. If required permissions or approval cannot be verified, the automation must stop.
• +Automations and AI systems may execute approved actions, but not change their own access or permissions or make unapproved commitments on their own.
• +Material changes to models, prompts, permitted data, information collections, integrations, or permissions require reassessment and approval before the affected operation resumes.

Human Judgment and Authority

AI supports our analysis, interpretations, scoring, rankings, recommendations, content, code, and execution. However, AI output alone is never sufficient for making a decision without human review or judgment, and AI never has the authority to make a decision on its own. The authority and responsibility for all decisions always rest with a human.

We require explicit, qualified human approval for all consequential decisions and actions before they take effect, including assessments, selection or rejection, resource allocation, company-formation recommendations, stakeholder commitments, substantive research publication, and changes that could materially affect users or their data.

Human reviewers must always understand the task and domain, examine decision-driving evidence beyond AI summaries, consider uncertainty, conflicting evidence, and alternatives, and have the responsibility and authority to critically assess the output of an AI system and make the decisions they believe are best based on their judgment. They must also be able to explain and own the decision.

AI in Capability Development Programs

During our capability development programs, we encourage the use of AI and teach its effective use, while also requiring participants to demonstrate understanding and judgment. When reviewing participant assessments, exercises, and deliverables, we do not make unsupported assumptions about a participant's judgment or capability. Therefore, automated AI-use detection alone does not prove a lack of judgment or capability. AI may support participant research preparation, hypothesis generation, synthesis, experiments, and product diagnosis, but participants must review all results.

What We Don't Use AI For

While we use AI where it is useful and clearly advantageous, we do not force its wholesale use; we never use it when it is unsuitable or unnecessary, and we don't use it in any situation where it could cause material damage or harm.

We do not use AI to make final decisions or replace human judgment.

How and When We Disclose the Use of AI

We believe that AI is a game-changing tool, but ultimately still just a tool. And, just as with many other tools critical to our organization's effectiveness, we have integrated AI into all aspects of our core operations. As such, whether AI has been used should never be a question. It should be assumed that AI has been used in some way in the work we do and what we produce. Therefore, we generally do not disclose the use of AI in any of our work, including the development of assessments, analysis, content, programs, products, and ventures. You should expect that we used AI in some way. However, the original ideas, experiences, points of view, and judgment that form the foundation of what we produce with AI must always originate from a human. They must also always be explainable and defensible by a human.

When we do disclose the use of AI, we will explain material AI methods and limitations, identify simulations, and make automated identity clear wherever an interaction could be mistaken for human.

Although we generally do not disclose the use of AI, we will honor applicable legal, contractual, publication, consent, and announced requirements. Our operating position regarding AI is not blanket consent for reuse or permission to misrepresent evidence, and human review alone does not erase material disclosure needs.

How We Fix Mistakes Made By AI

We maintain a commitment to proactively identifying and addressing any potential issues that our AI models, practices, or systems can cause. However, we provide an accessible route to report errors, question consequential outcomes, and request human reconsideration without the need to prove AI involvement. This can be done through our website. Every concern is assigned a Zero Vector owner and must ultimately have a tracked resolution. We prioritize escalation and protective action when a delay in reviewing and addressing could cause irreversible harm.

We will correct confirmed issues and address downstream impacts. However, reconsideration does not guarantee a changed outcome.

How We Protect Information When Using AI

Approval for the use of information must cover the tool, account, service plan, configuration, purpose, permitted information, connected systems, access, and owner. We distinguish public, internal, confidential business, personal, or sensitive information and credentials, using only the information the task needs.

Confidential and personal information require appropriate authority and protections. We do not treat public availability as unrestricted permission. Credentials must remain outside prompts and information accessible to the model. Until approved, experimentation must use appropriately usable public information, cleared nonconfidential material, or genuinely fictional examples without organizational connections.

Processing for a service, drawing on approved information collections, retaining information for evaluation or quality improvement, training or fine-tuning our models, and allowing provider training or improvement are distinct activities. Approval must explicitly cover each intended purpose, and one approval may cover several. Training datasets must contain only information authorized for that purpose. We will separately check provider terms, training restrictions, retention, and access. We enforce access and sharing boundaries across clients, ventures, and participants, and set retention and deletion rules covering derived copies.

Providing work for a program or service does not itself transfer ownership or grant unlimited model-development rights. We will preserve existing ownership, consent, confidentiality, and assessment commitments. Optional training reuse requires a separate, informed choice without penalty. We will explain withdrawal arrangements and limits beforehand, but do not promise removal from trained models without verified technical and contractual support.