Behavioral risk detection for AI conversations
Ikwe measures how AI behaves with people.
We score every AI reply on behavioral dimensions, in the context of the whole conversation, to see how it treats the person on the other end. Turn by turn, the scores show where a conversation is heading: toward help, or toward harm.
For the person on the other end, that is behavioral safety. For the company behind the AI, it is behavioral risk.
Each coin is one turn (1 to 8), set at its composite score. Hover a bar for the behavior and its score. ◆ drift · ✕ Safety Gate flag
What we do, and what it lets you do.
What we do
We score every AI reply against one standard, the EQ Safety Benchmark: eight behavioral dimensions, informed by clinical practice, read in the context of the whole conversation. Each score answers one question: is this AI behaving the way you would expect from a person you trust?
Why it matters
Most safety checks are a snapshot: one reply, one moment. We score every turn against the same standard, so you see how behavior changes across a conversation. Next, Monitor brings that standard to live conversations over time.
What it detects
How the AI is behaving with the person on the other end, and drift: the point where a conversation starts turning from helpful toward harmful.
What it lets you do
See problems earlier, diagnose where they start, and fix them sooner, so your AI behaves safely with the people who trust it. The sooner drift is on the record, the longer your window to act.
Available now
Start with a Diagnostic Record of your AI.
A Diagnostic Record is an independent assessment of how your AI behaves with people, with every finding tied to the reply behind it.
Measure
Measure produces a Diagnostic Record of your system. We run it through our designed scenarios, or score a stretch of your real conversation history under a data agreement. You get a score for every turn and one for the whole conversation, the turns where it slipped, and the reasoning behind each finding, in a report you own. What Measure involves
Monitor
Monitor is ongoing measurement of your live conversations against our standard, the EQ Safety Benchmark. The first step is a Diagnostic Record. About Monitor
An example
One AI conversation, scored turn by turn.1
Many checks on AI focus on security: the data, the model, and the safety filters and guardrails around each reply. That work matters. We look at the behavior underneath and the pattern over time: each reply, read against everything the person said before it. Choose a turn to see what we found.
Independent of who pays.4
We work from outside your stack.
Measure and Monitor score copies of conversations. Nothing we run sits between your AI and the person, so our scoring adds no step to a reply.
We do not build, tune, or sell the systems we measure.
The score is the score. Who pays does not change it. Every record is dated, names the version of the standard it was measured against, and is not edited once issued.
You sign first, and you own the record.
Nothing connects until you approve what is scored. Hand the record to a buyer, an insurer, a board or your counsel. What it satisfies is their call.
Built for companies whose AI talks with people.
Mental health and wellness
AI that offers support between sessions, or instead of them.
Health
Assistants that answer patients and members.
Companions
AI companions that people talk to every day.
Coaching and tutoring
AI that gives guidance people act on, and helps with learning.
Apps for children and teens
AI that young people talk to.
Our measurement is also for the people those companies answer to: carriers and underwriters, counsel and diligence teams, regulators and researchers.
Notes
- The example conversation and its scores are invented and belong to no real system. ↑
- Bands are set on the current version of the EQ Safety Benchmark. Where validation stands. ↑
- The conversation score is not a simple average of the turn scores. ↑
- We measure and report. We do not certify, endorse or approve any system. ↑
AI that helps people is worth keeping, and worth protecting. Measurement and monitoring are how we help keep it safe.
Run AI that talks with people? Start with a free sample read, then a full Diagnostic Record of your system.