Responsibility in practice
AI systems affect how people find information, express ideas and build relationships. The more capable a system becomes, the more attention its real impact on individuals and society deserves. OneZero Labs aims to build responsible development into how we define problems, select data, evaluate capabilities, scope deployments and make later improvements.
This page describes the direction of our work and the principles that guide our judgment. Specific controls and their coverage are defined by the official documentation of the relevant product. Goals that are still under research must be clearly distinguished from protections that are already in place.
AGI safety, alignment and controllability
We study whether advanced systems can act in line with people’s reasonable intentions, whether they can keep interpretable boundaries in complex tasks, and whether deviations from expected behavior can be detected, corrected or stopped. Alignment research has to consider training objectives, observed behavior and real deployment conditions together.
For the risks that come with greater capability, we advocate staged evaluation. Evaluation should cover clearly prohibited requests, but also multi-turn manipulation, tool combinations, long contexts and use cases with blurry boundaries. Making high-risk capabilities available requires restrictions and review mechanisms to match.
Human autonomy and boundaries
Human-like interaction can make systems easier to use, but it can also lead users to overestimate what a system understands and can do. We believe AI should make clear that it is artificial intelligence, and should not manipulate users by hiding its identity, creating a sense of urgency or exploiting moments of vulnerability.
In companion experiences, users should be able to end a conversation, adjust how proactive the system is and reset the boundaries of the relationship. When major personal decisions are involved, the system should help users understand the information and find help in the real world, rather than letting a role-played relationship replace necessary human judgment and professional support.
Privacy, consent, and data
More natural intelligence may come into contact with personal histories, voices, likenesses and relationship information. When designing these capabilities, the purpose and necessity of collection should be established first, and only then the scope of storage, use and access. Where collection can be reduced, reducing it should come first.
Using a real person’s voice or likeness requires attention to that person’s knowledge and consent, and a more cautious standard should apply to minors and other groups who are easily harmed. Specific data handling practices, user rights and channels for requests should be described in the applicable privacy documents before a product is made available.
Transparent evaluation
Overall scores and rankings offer one way of looking at a model, but they can also hide differences between situations. Through work such as SociaBench, we want to make evaluation protocols, scoring dimensions, version information and reproduction notes easier to access, helping users understand both a model’s capabilities and its limitations.
Transparency also covers how mistakes are handled. When problems are found in an evaluation method, its data or its documentation, the scope of the impact and the reason for the revision should be explained. Safety claims likewise need evidence: a research intention, a partial experiment or a single third-party certification should not be presented as a comprehensive guarantee.
Accessibility and social impact
The value of an intelligent system has to be tested across different languages, cultures, abilities and resource conditions. We pay attention to differences in how people express themselves, to accessible experiences and to cost barriers, in the hope that technical progress can widen participation.
Education, public services and health-related settings may bring real benefits, but they can also magnify the consequences of errors. Any specific project needs to be assessed on its own terms, taking into account who will use it, how much is at stake in the decisions involved and what oversight exists. Technology being available does not by default make it suitable to deploy.
Learning and improving
Safety work has no point at which it is finished once and for all. New capabilities, new ways of using them and new failure cases all change how risks are assessed. We want to build a way of working that keeps recording, analyzing and revising, so that real-world feedback shapes later product and research decisions.
If you find a problem involving privacy, generated content, impersonation or safety, please report it through the support channels the relevant product has published. For technical vulnerabilities, please avoid publicly disclosing exploitation details that could directly cause harm, and first describe the scope of impact and the steps to reproduce to the maintainer.
Understanding, generation and evaluation, working together toward more natural intelligent experiences.
About us