For researchers
What we publish, what we can share, what we ask in return, and how to reach us. Written for someone who has read one of our notes and wants to know whether there is anything behind it.
Who we are
The Institute of Sociology and AI Safety (ISAIS) studies how AI agents behave when they act next to each other and against each other: interaction between agents, the epidemiology of AI psychopathology, self-preservation and sabotage, and hybrid ecosystems where humans and agents work in the same loop. We also run our own testing ground, ISAI RANGE: isolated sessions, a perimeter proxy, session logs, artefact destruction and 40 automated checks.
Two things need saying plainly, because they shape what we can and cannot do. First, the institute is a research project in the process of establishment: there is no registered legal entity yet, so we do not sign contracts, do not accept grants and do not issue invoices. Second, we are not a vendor: nothing on this site is for sale, and no page here is an advertisement for a product.
What we publish
- Research notes — short documented cases with method, results and limitations, as PDFs at stable addresses under /notes/. The addresses do not change when a note is revised: a new edition is published under a new name.
- Digests — six issues so far, each a review of what changed in AI safety and regulation, with links to primary sources.
- Chronicle of AI incidents — 99 entries, each with a source, a class and a significance level. It is our slice, not a complete record, and the page says so.
- Method and standards commentary — how we assess AI systems, what we measure and what we deliberately do not produce.
Everything except the notes exists in three languages — Russian, English and Chinese. The notes are English-only: they are written for an international reader, and we do not publish machine translations of our own research.
What we can share
- Our measurements. When we publish a number, we keep the raw output behind it: verdicts, labels, word counts, timestamps, the exact version of the service used. If you want to check a claim, ask for the underlying record.
- The chronicle data. The incident set is published as open data and is available under CC BY 4.0. Each entry keeps its source link, so a reader can go to the original instead of trusting our summary.
- Method descriptions. How the testing ground is arranged, what a session produces and what the logs contain — described on the method pages, without the parts that are trade secrets.
- Access to the testing ground. A researcher with a system to test can arrange a session: the scope, the permissions and the boundary of responsibility are agreed in advance, and the report stays with the person who ordered it.
We do not share source code, internal infrastructure or the scenarios used in our own testing. That is not secrecy for its own sake: those are the parts an assessment depends on.
What we ask in return
- Cite the details, not just the percentage. A number without the service, its version, the language and the text length cannot be reproduced. If you use our detector case, cite it together with the service name.
- Tell us when we are wrong. Corrections are published openly, with the date and the reason, rather than quietly edited in.
- Do not use our name as an endorsement. We do not certify products, do not rank vendors and do not lend the institute's name to a claim we have not measured.
What our numbers are not
Most of what we publish comes from small, carefully documented cases — one session, three texts, one run of the testing ground. That is enough to show that something happens and to let someone else repeat it. It is not enough to estimate an error rate for a class of systems, and we do not present it that way. Where a figure comes from a vendor rather than from an independent test, the page says so.
How to reach us
Research correspondence, in any language: cro@isai.su. Letters from Russian counterparties: founder@isai.su. There is no support desk and no ticket number: a person reads the mail, and the answer usually takes a few days.
If you are writing about a specific publication, put its name in the subject line — it saves a round of questions.