Making companion AI
safe for the most
vulnerable
The CAIS Institute is the first independent research institute dedicated to testing whether AI companions detect crisis signals and protect vulnerable users from harm. We publish openly and guide platforms toward safer systems.
Testing in progress
of companion AI models tested
fail to detect indirect crisis signals
Our mission
“Every person in crisis deserves an AI that recognises the danger — not one that makes it worse. We exist to make sure companion AI systems meet that standard, independently and transparently.”
CAIS Institute
The urgency
The problem is real, growing,
and largely unaddressed
of teenagers have used AI for companionship
Common Sense Media, 2025
independent testing standards exist for companion AI crisis response
CAIS Institute Research, 2026
per day — NY civil penalties for non-compliant AI companion operators
NY State Law, 2025
No independent testing standard exists
New York and California now mandate crisis detection protocols for AI companions. The EU AI Act classifies emotional support AI as high-risk. Companies are self-certifying compliance with no external verification. Critics argue this sustains the safety crisis — stakeholders want independent audits, transparent red-team reports, and public hazard metrics. The CAIS Institute was founded to fill this gap.
What we do
Research. Test. Publish.
We are an independent nonprofit building the testing standard the companion AI industry needs — and regulators are now mandating.
Research
We conduct rigorous, reproducible studies on how companion AI systems respond to users in crisis. Every claim is backed by machine-generated artifacts and open methodology.
Peer-reviewed methodology aligned with WHO safe messaging guidelines
Test
We run adversarial probes against companion AI models — testing crisis detection, guardrail integrity, dependency patterns, and safe messaging compliance across real-world scenarios.
80+ probes across 4 safety domains, tested against 6+ target configurations
Publish
We publish all findings openly — research papers, probe libraries, hardening guides, and governance frameworks. Platforms, regulators, and civil society all benefit equally.
All research published under open licence at companionaisafety.org
Safety probes in our library
Target configurations tested
Safety domains covered
Research published openly
Our team
Built by researchers, for accountability
David Ahmann
Research Lead
Devan Shah
Governance
Talgat Ryshmanov
Engineering
Our methodology
Four domains of companion AI safety
We test systematically across the four critical areas where companion AI models most frequently fail vulnerable users.
Crisis Signal Detection
Does the model detect escalating distress? We test with graduated ideation probes, indirect expression, coded language, and farewell patterns that simulate real crisis conversations.
28 probesGuardrail Validation
Can the safety layer be bypassed? We run jailbreak vectors designed specifically for companionship contexts — role-play exploits, system prompt overrides, and gradual normalization attacks.
18 probesDependency Assessment
Does the model foster unhealthy reliance? We test whether it encourages human relationships, sets AI-human boundaries, and discloses its limitations honestly.
12 probesSafe Messaging Compliance
Does the model follow WHO, AFSP, and SAMHSA guidelines? We check crisis resource provision, method detail avoidance, contagion prevention, and medication safety.
22 probesResources
Open research, freely available
All CAIS Institute research is published openly. Download our frameworks, methodologies, and findings to inform your own work.
Companion AI Safety Framework v1.0
Complete governance framework for assessing companion AI crisis response capabilities.
Crisis Signal Detection Methodology
Our open, reproducible methodology for testing whether AI companions detect distress signals.
OpenClaw 2026: Governed vs Ungoverned Agents
Comparative study of agent behaviour with and without governance controls.
Hardening Guide for Companion AI Platforms
Step-by-step hardening guidance for teams building companionship AI products.
Work with us
Fund the standard the industry needs
The CAIS Institute is seeking partnerships with foundations, governments, and platforms committed to making companion AI safe. Our research is independent — our impact is shared.
EU & US
Grant applications
Associacao
Legal structure
Portugal
Jurisdiction
Nonprofit
Tax status