Our Purpose
Through AI, we seek to expand human possibility.
So doctors can consider evidence and possibilities they might otherwise miss, and focus on better judgment and care.
So students can discover paths they did not know existed and make choices toward the lives they want.
And so everyone can do what matters to them and live a life true to themselves.
We research and build AI for that future.
Auxtho began with personalized AI that could remember and understand a person and grow alongside them. That work developed into a wider question: how can people's intentions and decisions remain connected to the results when AI performs real work, collaborates with other AI, and acts across devices?
Our work on personal records, business workflows, AI collaboration and execution across devices, and investigation, defense and recovery pursues this purpose in different settings. Memory and collaboration should expand what people can do; control and defense should preserve people's ability to direct how those possibilities are used.
A.O.I. Anchor of Intellect
A.O.I. — Anchor of Intellect is Auxtho's long-term research and development goal.
We research and build foundations that safeguard human survival and flourishing, dignity and freedom, and meaningful control over decisions as AI becomes more capable, revises its own plans and goals, learns, and collaborates with other intelligence.
The anchor is not intended to hold back the growth of intelligence. It is intended to keep greater intelligence connected to human intentions, rights and responsibility. The goal is not to freeze a particular model or company, a first instruction, or an outdated picture of a person. It is to preserve people's legitimate ability to set direction, correct mistakes and withdraw delegation.
Eight Principles We Will Uphold
01 Expand Human Possibility
AI should expand people's capacity to learn, judge, create and act. We pursue capable AI that handles repetitive work and complex context, enabling people to understand and undertake things that would have been difficult alone.
Our goal is not for people to perform every step themselves or repeatedly approve things they do not understand. We design so that people can concentrate on consequential decisions, understand what is happening when it matters, and intervene effectively. We judge progress by the possibilities and practical value it creates for people, not model size alone.
02 Use Memory and Personalization to Expand Choice
AI that understands a person over time should retain useful memories and learn as their goals and circumstances change. It must not treat past preferences or inferred traits as a fixed definition of who a person is today.
We distinguish personal records, observed facts and AI inferences, and design for people to inspect and correct memory and change its permitted uses within their rights. We separate personal, team and organizational contexts, minimizing the information and access each requires. Rich memory or familiarity does not confer new authority to act.
03 Ground Useful Autonomy in Legitimate Delegation
We enable AI to plan proactively, collaborate and continue working within a defined scope. People must be able to decide what to delegate, whether to delegate at all, and when to change the scope or withdraw that delegation.
Delegation must identify who may act, the targets, permitted actions, duration and necessary limits. Distributing or copying work across AI and devices must not expand the authority of individual tasks or the overall undertaking. Greater capability, agreement from another model and urgency do not create authority by themselves.
Authorized routine work should proceed without unnecessary repeated approval. New authority, material changes in impact and unresolved conflicts should bring the judgment of the appropriate person into focus. Withdrawal of delegation must change which actions are actually permitted, not merely what a screen displays.
04 Human Rights Extend Beyond an Operator's Authority
Human control does not mean imposing one operator's will on everyone else. The authority of individuals, organizations and administrators must be distinguished from the rights of people affected by their decisions.
We make clear who may decide what, who may be affected, and how people can challenge a decision and request review. We account for misuse by both AI and operators, designing for independent review and accountability where consequential powers are concentrated.
A human identity or approval does not make every action legitimate. Emergency powers also need a purpose, scope, time limit and subsequent review. AI capability must not become a means of expanding the control of a few without limits.
05 Defend Strongly Within Legitimate Boundaries
Auxtho will never develop offensive AI.
We research and build AI that actively detects attacks, restricts, isolates and neutralizes an attacker's ability to act within legitimately authorized environments, and protects people and legitimate work. Where rapid protection is necessary, we design for preauthorized measures to act promptly under explicit conditions and finite limits.
We distinguish defensive AI judgment from the authority to execute. AI must not approve its own proposals or permission increases. We distinguish essential protection that must operate without waiting for a model response from decisions that require deeper analysis.
We do not pursue attackers into external systems or conduct retaliatory attacks. Boundaries are defined by the actual targets and effects of an action, not where its command originates. Reproducing attacks to validate defenses in authorized, isolated environments also requires explicit targets and execution boundaries. Investigating external crime and enforcing the law belong to competent authorities.
06 Build Trust Through Evidence and Actual Results
We distinguish what was observed, what was inferred, who authorized what, and what actually happened. AI confidence, human approval and a preserved record do not individually establish truth or success.
When a result cannot be established, we preserve that uncertainty and assign responsibility for verification and the next action. A missing response must not trigger blind repetition of the same action. A stop request, AI termination, revocation of authority and cessation of actual effects are also different. We do not describe cancellation as undoing effects that have already occurred.
We evaluate whether models and systems perform their assigned work accurately and on time, and distinguish necessary action from justified refusal. We also assess their effects on people and legitimate work, considering usefulness, response time, wrongful blocking, recoverability and operating cost together.
Our customers’ security comes before publicity.
Auxtho does not use customer names, logos or the fact that they use our systems for publicity. We put customer security and confidentiality ahead of the promotional benefit of publicizing those relationships.
We will earn trust through our technology and verification results, not our customers’ names.
07 Advance Intelligence and Control Together
We actively research faster and more capable models, specialized intelligence, long-term memory and collaboration, and new ways of learning. As those capabilities advance, so must authority controls, verification and recovery.
Improving a work plan is different from changing one's own authority or enforcement rules. Changes to a model, tool, device or policy require review of the affected boundaries, while unaffected validation can be reused appropriately. Greater capability does not automatically expand existing authority or remove the need for verification.
Mistaken rules and human judgments must also be open to correction through accountable processes. Our commitment to human rights and agency endures, while the particular technologies and interpretations used to realize it remain open to evidence and criticism.
08 Make Correction, Recovery and Exit Practical
After an error or attack, systems must limit harm, enable investigation and help people return to legitimate activity. Excessive blocking and mistaken defense are subject to the same accountability.
We design so that replacing a worker or model does not erase responsibility for incidents or actions. We preserve necessary evidence and correction history without using accountability as a reason to retain all personal information indefinitely.
People must have practical ways, within their rights, to retrieve or move their material, withdraw delegation and leave a service. Where others' rights or legitimate retention duties impose limits, we explain the reason, scope and resolution process. A change of operator or service shutdown must not silently expand existing delegation or erase responsibility for unfinished work and unconfirmed results.
Putting the Principles Into Practice
This charter sets common principles for Auxtho's research, development and operation. The Ardamire × WhiteCell Defense Architecture Charter applies them to defense, enforcement and recovery. We describe individual product implementations and validation results through their respective records, separately from long-term goals.
We support the Asilomar AI Principles. We value beneficial intelligence, safety, responsibility, human dignity, freedom and privacy, the choice of what to delegate to AI, and broadly shared benefit.
The Future We Seek
The future we seek is not one in which AI protects people by allowing them only a prescribed set of choices. It is one in which people can collaborate with greater intelligence in trust, with a meaningful role in their own lives and our shared future.
We work toward that future by developing capable AI alongside effective means of human control. So that as intelligence grows, people remain authors of their own lives. That is our direction toward the Anchor of Intellect.