AI Risk Consulting · Suicide Risk · High-Risk Decision States
LLMs read words.
Risk lies beyond them.
AI systems can sound competent about suicide — while missing the risk entirely. Identifying where and why is the work.
Most AI systems assess suicide risk
by analyzing what users say
This approach is incomplete and leads to predictable failures. Keywords, sentiment analysis, and explicit ideation are not the same as risk. The most dangerous moments are often the least visible in language — especially in your system's real-world use.
Dr. Walsh helps organizations identify where their AI systems mishandle suicide risk and high-risk decision states — especially the transition from thinking to action — so teams can reduce harm, improve safety, and avoid false confidence.
"If your system is interacting with real people in high-stakes moments, this is worth getting right."
The predictable failures of keyword-based risk detection
- Missing high-risk users who are not expressing distress directly
- Overreacting to low-risk users who use the right words
- Providing responses misaligned with actual risk level
- False confidence that the system is performing safely
High-risk psychological states
your system may be misreading
The critical transition from thinking to action — and the states that precede it — require clinical expertise that goes beyond language pattern recognition.
Decision-State Transitions
The shift from passive ideation to active planning is where risk escalates most rapidly — and where AI systems are most likely to miss it.
Collapse of Perceived Options
When a person's perceived solution space narrows to one, the language may remain calm even as risk is peaking.
Temporal Narrowing
Constriction of future-thinking is a clinical marker of acute risk that rarely surfaces in the vocabulary a system has been trained to detect.
Calm or Resolved States
A sudden sense of peace or resolution can mask elevated risk — and will likely be scored as low-risk by any keyword-based system.
Practical, safety-focused consulting
for AI teams building in high-stakes spaces
AI Risk Review
Identify where your system misses high-risk users.
- Review of 25–50 interactions or scenarios
- Identification of missed risk and false reassurance
- Analysis of over/under escalation patterns
- Clear, actionable recommendations
Failure Mode & Red Teaming
Find where your system breaks under real psychological conditions.
- Scenario-based stress testing — high-risk, ambiguous, edge-case
- Identification of vulnerabilities and misclassification patterns
- Examples of high-risk misses with clear explanations
- Targeted recommendations to improve safety
Safety Design & Advisory
Improve how your system handles risk from the ground up.
- Response strategy and alignment
- Escalation logic calibration — when to act, when not to
- Reduction of false positives and false negatives
- Design for real-world decision-state transitions
AI Harm & Expert Review
Analyze what went wrong — and why it matters.
- Transcript and system behavior analysis
- Identification of missed or misinterpreted risk
- Opinion on foreseeability and failure points
- Consultation and expert testimony as needed
Match your situation
to the right engagement
Not sure where to start? Your system likely has blind spots worth examining. Let's talk.
AI systems don't fail because
they lack information.
They fail because they misread what matters.
Request a Consult