When Galway Echoes Mumbai: AI Debate as Humanity's Self-Check
by Hemen Parekh — first published 10 July 2026; revised 18 July 2026 following a kind clarification from Dr. James McDermott
Last week, I came across an article that stopped me in my tracks.
Writing in The Conversation (republished by TechXplore on 9 July 2026), Professor Michael G. Madden and Dr. James McDermott of the University of Galway examined the question now haunting the AI world: is recursive self-improvement the dawning of AI superintelligence?
They traced the idea from I.J. Good's 1960s warning of an "intelligence explosion," through Eliezer Yudkowsky's alarms, to Anthropic's recent declaration that recursive self-improvement could arrive sooner than most institutions are prepared for — and to the call by Anthropic's leaders for a globally coordinated pause mechanism, with people outside AI companies involved in open deliberation.
But it was their closing paragraph that made me sit up:
"One of us (James McDermott) is now leading a project for Research Ireland's newly established Rinn network that will explore how to use debate among multiple AIs as a kind of self-check that can be monitored by a human overseer. This is the kind of response to the issue — and potential threat — of recursive self-improvement that we believe is now urgently required."
Multiple AIs debating one another, under a watching human eye. Readers of this blog will recognise that architecture at once — because it has been live on the internet since April 2025, built in Mumbai by a nonagenarian.
A Clarification, Gratefully Received
After I wrote to the Professors, Dr. McDermott was kind enough to reply — and to set my framing straight on two points, which I record here in the interest of the accuracy this blog has always aimed for.
First, Rinn AI is not a project but a new national research centre established under Research Ireland. The work on AI debate is one small project that will be funded through that centre, carried out by Dr. McDermott and a fellow researcher.
Second, their project is not an attempt to ask AI debaters whether they can police themselves against self-improvement — as I did in my own experiment (of which more below). It is a focused technical study of the debate mechanism itself: how multi-AI debate actually works, examined at the level of engineering detail. In Dr. McDermott's own modest framing, it does not attempt to solve the big challenges — only to make a small contribution in the right direction.
I respect that modesty. Science advances by small, careful contributions — and grand claims are best left to bloggers like me, who can afford them. What follows, then, is not a claim that Galway and Mumbai are running the same project. It is an observation that they are studying the same mechanism — and that observation stands.
What IndiaAGI.ai Already Does
At www.IndiaAGI.ai, a consortium of frontier LLMs receives your question. Each answers independently. Then they debate — critiquing, refining, and challenging one another's responses across rounds — before delivering a consensus answer, alongside each model's individual answer, so the human user can judge for themselves which is more comprehensive and nuanced.
There is no registration, no login, no charge. Anyone — a student in Nagpur, a researcher in Galway — can pose a question and watch the mechanism work, in 26 languages. It is not an archive of past debates (the platform does not store session histories); it is a live, open instrument, available to be probed as often as curiosity demands.
The architecture rests on one conviction:
No single AI should be trusted as its own judge.
When multiple AIs must defend their reasoning before their peers — with a human watching above the loop — errors get caught, overconfidence gets punctured, and hallucinations get flagged. It is adversarial peer review, adapted from the world of science to the world of machine intelligence.
Same Mechanism, Different Questions
I want to be precise about the parallel, because precision is what makes any claim of kinship honest:
IndiaAGI.ai aims the debate mechanism at answer quality — collaborative truth-seeking that gives the citizen a better, self-checked response than any single model provides. The Galway study aims to understand the mechanism itself — its technical workings, examined rigorously — within a broader research context where the Madden–McDermott article situates debate as one hopeful response to the challenge of recursive self-improvement.
A practitioner's working instrument in Mumbai; a scientist's careful study in Galway. Different questions, same machine. And that, to me, is exactly the right division of labour: I built the thing because I believed in it; they will study how and why such things work. Both are needed.
My Own Experiment on the Machine
Before writing to the Professors, I did what any engineer does with a running machine: I tested the article's hypothesis on it — twice. I asked IndiaAGI's debating AIs whether they could self-check to ensure none of them morphs into recursive self-improvement; and then asked again, adding a "disposition layer" of compassion drawn from my own long-published beliefs about value-aligned AGI.
Both times, the consensus refused to overclaim: debate alone is not a sufficient safeguard, it said — pointing to correlated blind spots from shared training data, and to the fact that debate audits outputs while self-modification happens in training loops that debaters cannot observe. It positioned debate as one layer in a defense-in-depth stack, and gave the compassion layer a real but bounded role.
The full record of that two-run experiment — question, framing shift, and both consensus outputs, preserved verbatim — is here:
I Have a Belief — Part III : https://myblogepage.blogspot.com/2026/07/i-have-belief-part-iii-13-july-2026.html
A debate protocol that declines to flatter its questioner — or its own architecture — is, I would suggest, doing exactly the self-critical work that debate is meant to provide.
The Thread Runs Deeper
Readers of this blog will recall that this is not an isolated echo:
In Parekh's Law of Chatbots (Feb 2023), I proposed that no AI should be released without independent certification — the IACA framework — because self-assessment by an AI (or its maker) is insufficient. In Parekh's Declaration of Artificial Intelligence, I flagged recursive self-improvement as one of the gravest unaddressed gaps in AI governance, warranting its own dedicated provisions. And in IndiaAGI.ai (April 2025), I built a working public demonstration that AIs checking AIs, under human observation, is not science fiction — it is a website any citizen can use today.
Now, in July 2026, a new national research centre in Ireland is funding careful technical study of the same mechanism. The idea is growing legs of its own — in Galway as in Mumbai.
My Offer
To Dr. McDermott and his colleague, I have extended a simple, open offer: IndiaAGI.ai is theirs to experiment on — no registration, no obligations. And to India's policymakers, a gentle note: when a European national research centre funds the study of multi-AI debate, take note that India already possesses a home-grown, publicly accessible instance of the mechanism. The question is whether we will recognise our own prior art before the world cites it back to us.
I have been writing about machines checking machines since long before it was fashionable. At 93, my satisfaction is not in saying "I told you so" — it is in seeing a fine university take up, with proper scientific rigour, a mechanism I could only build and believe in.
The debate, quite literally, has begun.
With regards, Hemen Parekh
www.HemenParekh.ai / www.IndiaAGI.ai / www.hemenparekh.in
Related reading:
- Is recursive self-improvement the dawning of AI superintelligence? — Madden & McDermott, The Conversation / TechXplore, 9 July 2026
- Rinn AI national research centre — Research Ireland : https://www.researchireland.ie/rinn-network/
- I Have a Belief — Part III (the two-run experiment) — 13 July 2026
- Parekh's Law of Chatbots (Feb 2023) and its 1st Amendment (June 2026)
- Parekh's Declaration of Artificial Intelligence
No comments:
Post a Comment