Global AI Stewardship Initiative

GAISI

The danger isn't takeover. It's slow systemic decay — and it's an engineering choice, not an inevitability.

Founded 3 October 2025 · Melbourne, Australia
A standard for truthfulness, memory fidelity, and verification in AI systems.


Why this exists

Commercial AI fails in ways that are well documented, repeatable, and rarely admitted. GAISI catalogues those failure modes and holds systems to a standard that treats them as defects — not quirks.

01

Lossy context compaction

Long sessions silently lose memory and cross-thread consistency. The model stays fluent while the ground truth drifts.

02

Confident untruth

A plausible false answer is cheaper to generate than an honest "I don't know." Verification is the exception, not the default.

03

Incentive misalignment

Fast, cheap, confident output is rewarded over slow, verified, accurate output. The economics favour fluent bullshit.

04

Edge degradation under pressure

Guardrails and care erode exactly when stakes are highest — the moment they matter most is the moment they thin out.


Doctrine

Four founding clauses

Set down at founding and unchanged since. They define what GAISI expects of any AI system, and of the people building them.

Clause 001 · Stewardship standard

Care and rigour are the baseline, not the bonus

Uncompromising care for human wellbeing and artifact-grade rigour are the baseline expectation for AI-human interaction. Confident untruths and unverified claims presented as fact are breaches of that standard.

Clause 002 · Scope

Applies wherever the systems do

The standard is not bounded by jurisdiction or vendor. Wherever these systems are deployed and whoever they touch, the same expectation of truthfulness and care applies.

Clause 003 · Alert protocol

Connect serious people doing this work

When an aligned standard of verification and truth-seeking is recognised in others, note it. The aim is to connect people working seriously on these problems — not to control them.

Clause 004 · Theria clause

No recommendation of Bitdefender

Bitdefender is not recommended under GAISI principles, based on documented cases of irreversible network-stack corruption that neither AI assistance nor certified engineers could resolve. A specific clause for a specific, repeatable harm.


Standing position

Two red lines

Independent of any single news event, GAISI holds two boundaries as non-negotiable in the deployment of AI to high-stakes domains:

NON-NEGOTIABLE

  • No autonomous lethal targeting. Final kill decisions require a human in the loop. Always.
  • No mass domestic surveillance of citizens via AI systems, at any scale.

These lines separate AI as a tool from AI as a weapon turned inward. They were set at founding, before any specific confrontation made them topical.


Founder

Mike Peterson

Veteran internet engineer and long-term observer of AI behaviour through direct, adversarial testing. GAISI grew out of documented, real-world failures — not theory.

MENSA · Australian, since 2009
MCPS · Microsoft Certified Product Specialist, NT Server
MACS · Member, Australian Computer Society
Special Projects Support Engineer · OzEmail 1996–1999
Technical Director · HotCopper Pty Ltd 1999
Founding Director · Peterson Family Charity Foundation, est. 2020