Why Probabilistic AI Is Negligent and Uninsurable

Defining the new standard of care for the autonomous enterprise.

Study2026Dustin Allen and Hearsch Jariwala

18Chapters
4Field manuals
300+Pages

The core thesis

Chaos vs. order.

The central conflict of this paper is between the chaos of AI, the actor, and the order of a predictable guard, the governor.

Before: chance
A fuzzy, shifting cloud. It changes shape when you touch it. A black box where inputs go in and unpredictable outputs come out.
After: rules
A rigid, crystal-clear structure. A glass box where every piece of data can be traced.

The physics of failure

Safety that depends on server load is not safety. It is luck.

Today’s AI safety tools are flawed at the root. These are not bugs to fix. They are physics to replace.

The drift problem
On a GPU, the math changes with server load. 1 + 1 does not always equal 2. Safety checks drift by up to 21.4% under real production load.
Compound failure
A 99% safe model is fine for chat. It is fatal for an agent doing 50 tasks. Each step leaks safety. By step 50, you are at coin-flip odds.
The rare-fact problem
AI must make things up about rare facts. It is a statistical law. Your company’s data is rare, so the AI will make things up about it.

Field manuals

Four teams. One standard.

Each appendix turns the physics into the language of one role, so every team can see its risk and its way out.

For general counsel

  1. The liability shift: from publisher to operator
  2. Why built-in safety fails the foreseeability test
  3. The glass box defense: chain of custody
  4. From hidden liability to insurable assets
  5. The end of the black box defense

For actuaries and risk officers

  1. The black box pricing crisis
  2. Why standard deviation fails for AI
  3. The Net Insurable Token framework
  4. Risk decay curves and reserve release
  5. AI’s steam boiler moment

For engineering leads and CISOs

  1. Floating-point math and safety drift
  2. Load-proof guard architecture
  3. Hot-swappable protection
  4. Training small guards from big ones
  5. Replaying any decision exactly

For auditors

  1. Why chance-based logs are a weakness
  2. Checking every decision, not a sample
  3. Replaying the past
  4. Mapping to SOC 2, ISO, SOX, and GDPR
  5. The control function exemption

The verdict

The standard has shifted.

  • The risks are known.
  • Hiding how AI decides is negligence.
  • A predictable guard is the new baseline.

The ways AI fails are now known, so not knowing is no longer a defense. Running a black box when a glass box is available invites real damages. Chance-based guardrails are not enough.

Get your own AI on your side.

Everything we learn goes into Emu.

Free for Mac and Windows. Opens October 12.

Get early access to Emu.

Emu opens October 12. Leave your name and email and we will let you in as soon as a spot opens.

Free. We only email you about Emu.