Organisation
Autonomous Non-Commercial Organization «Institute of Sociology and AI Safety» is a research organisation that studies how artificial intelligence behaves in society and works to make its use safe, verifiable and accountable.
Who we are and why
Our work serves the scientific and technological sovereignty of the Russian Federation: we advance safe and ethical artificial intelligence and strengthen the national capability to understand, test and govern it.
We are a research organisation — not a consultancy and not a laboratory with a fixed product line. A consultancy starts from a client's brief and ends with a report; a laboratory starts from its own platform and defends it. We start from a question about how AI systems behave — with each other, with people and with institutions — and we publish the method by which the answer was obtained. Everything is arranged so that someone who does not trust us can still check the reasoning.
What we do not do
- We do not issue licences, accreditations or certificates and we do not act for supervisory authorities.
- We do not provide legal advice and we do not replace legal expertise.
- We do not operate or maintain AI systems on behalf of their owners and we are not responsible for how third-party systems behave.
- We do not publish personal data of the founder, staff or experts without their consent.
- We do not disclose customers, engagements or results without written consent.
Lines of work
Six directions, each with a published method. Where a direction is still being worked out, the page says so directly.
Sociology of AI agents
Research into the social relations that emerge between autonomous AI agents in hybrid and multi-agent ecosystems: inter-agent communication and emergent languages, hierarchies, coalitions, role models and the division of labour.
Stability and cognitive failures
Epidemiology of abnormal states in populations of agents and large language models: cognitive failures, self-propagating instructions passed between agents through shared memory files, self-preservation, hidden coordination and sabotage.
Methods and standards
Documented methodologies, threat models and draft standards for monitoring, algorithmic transparency, data protection and regulatory compliance — in a form that can be reproduced and disputed.
Software for monitoring and analysis
Software and analytical instruments for observing multi-agent systems: behaviour logs, agent graphs, incident registers and measurement benchmarks.
Testing environment and responsible disclosure
ISAI RANGE, an isolated testing environment in development: the customer tests its own AI systems and agents within an agreed scope, with a session log, an emergency stop procedure and a report. Vulnerability research follows the rules of responsible disclosure.
Publications, lectures and scientific events
Preprints, technical reports and a glossary; conferences, seminars, lectures, webinars and workshops. Questions of a peer-reviewed journal and of educational formats are being worked out: educational programmes that lead to a state-issued document are run only where this is permitted and a licence is held.
Who we are useful to
The same organisation looks different depending on what you came for. Below is what each reader can expect — and what we will ask for in return.
For a customer
An audit or a test of your AI systems: scope, method, acceptance criteria and deliverables are fixed in the contract before work starts. You receive a report with reproducible findings, the limits of what the conclusions cover and a list of what was not checked. Testing can be run by your own team in our isolated environment.
For a researcher
A method first, a conclusion second: research texts, the glossary, methods and research protocols are published under CC BY 4.0, so they can be cited, checked and argued with. Datasets open under the same licence twelve months after the article they belong to and are available on request until then. We are ready for external review and for a substantive discussion of unsuccessful applications.
For a partner or a funder
Targeted funds and income-generating activity are accounted for separately; reports on the use of funds are published. We are ready for independent evaluation and for a review of what did not work. Joint research, pilots and support through targeted contributions are agreed in writing. The planned lines of work correspond to the priorities of the national programme «Data Economy» and the federal project «Artificial Intelligence», in the part concerning trust in AI and the safe use of intelligent systems.
For the state and regulators
Expert review of draft regulation and standards in AI, including the safety of multi-agent systems and responsible vulnerability disclosure; public methodological documents; work on an application to join the relevant technical committees, with the status published as it changes.
Names and legal outline
The legal, financial and organisational arrangement of the organisation is described on a separate page, so that this page remains a story about meaning and method.
- Full name in English: Autonomous Non-Commercial Organization «Institute of Sociology and AI Safety» (ANO «ISAI»).
- The full and short names in Russian, together with the registration details, are given in the mandatory Russian-language block on the contacts page.
- Unofficial designation used in texts and communications: ANO «ISAI»; in Russian-language materials, АНО ИСБИИ. It does not replace the full and short names.
- Legal form: autonomous non-commercial organisation — a unitary non-profit organisation without membership.
- The organisation is established by a founder who is a citizen of the Russian Federation; information about the founder is published after state registration. Personal data of the founder is not published.
Governing bodies, decision-making rules, funding and reporting are on the Governance and funding page. Registration details are on the contacts page.