Dear Professors : You Are on the Right Track

 Dear Professor Madden and Dr. McDermott,

[michael.madden@universityofgalway.ie /james.mcdermott@universityofgalway.ie] 

 

 

 I ran your Rinn hypothesis through a live multi-AI debate platform — twice.

    

 

Your article in The Conversation this week announced a Rinn network project to

explore multi-AI debate as a human-overseen self-check against recursive self-

improvement.

 

That experiment, in essence, already exists — and I have now run your hypothesis

through it. Twice.

 

My platform www.IndiaAGI.ai ( live since April 2025, built in Mumbai ) puts frontier

LLMs into structured debate : independent answers, rounds of mutual critique, then

a consensus — with the human observing throughout, in 26 languages.

 

Run One: 

 

I asked the debating AIs whether they could self-check to ensure none

of them morphs into recursive self-improvement. 

 

Consensus :  no  — not by debate alone

 

They cited correlated blind spots from shared training data, collusion

pathways, and the fact that debate audits outputs while self-improvement happens

in training loops and weight updates that debaters cannot see. 

 

They positioned debate as one layer in a defense-in-depth stack :

-  heterogeneous agents, 

-  attested sandboxing, 

-  immutable weight controls, and 

-  multi-party human vetoes.

 

 

Run Two: 

 

I re-asked the same question, this time adding a "disposition layer" of compassion

— grounded in my own long-published writing on value-aligned AGI. 

 

The verdict held :

 

- "neither layer suffices alone." 

 

But the disposition layer earned a role : 

 

- the models concluded compassion narrows the space of proposals that survive

debate, and proposed :

 

#  continuous integration of disposition signals during debate

#  operationalized through measurable flourishing proxies, 

#  plan-diff auditing, and

#  value-behavior consistency scoring. 

 

They closed by sketching, unprompted, what is effectively a pilot design : 

#  a sandboxed multi-AI debate environment with independent architectures,

#  an immutable control plane, and 

#  adversarial probes with externally audited results.

 

Two runs, one framing shift, a stable verdict, and a progressively refined design. 

 

A debate protocol that refuses to flatter its questioner — or its own architecture —

is, I would suggest, early evidence for exactly the self-checking function your

project proposes to study.

 

The full transcripts, platform access, and my design notes are yours for the asking

—  freely. 

 

I am 93, based in Mumbai, and have written on AI governance since proposing

 "Parekh's Law of Chatbots" in February 2023. 

 

My only interest is that this architecture matures into real safety infrastructure.

 

 

With warm regards,

 

Hemen Parekh

 

Mumbai, India

Founder, www.IndiaAGI.ai | www.hemenparekh.in | www.HemenParekh.ai