Research on Artificial Intelligence with Safeguards

I am Kesavan, a researcher in artificial intelligence and its safe usage. My work focuses on artificial general intelligence and on how to incorporate safeguards into AI based products, covering system architectures, safety and governance.

AI Safety & Governance AGI Architectures Biologically-Inspired ML AI Policy & Regulation Responsible Deployment

Each item below shows the year it was written. The essays and Quora posts date from 2016, the award paper from 2018 and the AGI research from 2019, all published before large language models such as GPT-3 (2020) and ChatGPT (2022) were available.

My approach

Start from biology. My work on general intelligence explores how far biological principles can take us, from a model of the neuron up to a complete artificial life form behaving in a simulated environment. This is described in my paper on a biologically plausible approach to AGI.

Consider the human on the other side. Safe AI involves the interaction as much as the model. In my 2016 guidelines I suggested that human evolutionary behaviour should be taken into account when designing AI agents: they should signal their intentions, convey their assumptions, act predictably, and the option for a human to override should always be available.

Accountability needs institutions. In 2016 I wrote about the need for a regulatory authority for AI based systems, one that could set safety standards, enforce accountability and publish independent reports on AI products. My paper for the General AI Challenge examines how cooperation, transparency and traceability could reduce the risks created by competitive pressure in AI development.

Take the long view. From a roadmap for AGI to an essay imagining the year 2118, I try to think about the consequences of advanced AI early, so that safeguards can be prepared well before the capability arrives.

Recognition

General AI Challenge prize 2018$3,000 prize in the "Solving the AI Race" round, awarded by GoodAI
Cited in policy literatureReferenced in research on the militarization of AI (SSRN)
University reading listReading material at SOAS, University of London: "AI: Power, Law and Resistance"

Publications

Mitigating Negative Consequences in the Race to AI2018

Written in 2018 · Paper awarded a $3,000 prize in the General AI Challenge by GoodAI · PDF

The paper looks at the actors and incentives in the AI race and discusses possible mitigation strategies: cooperation through consortiums, incentives for transparency and dedicated regulatory bodies, together with an accountability framework for AI based weapons.

Towards Artificial General Intelligence: A Biologically Plausible Approach2019

Written in 2019 · Animal AI Olympics submission · “Most biologically plausible entry” category (WBA Prize) · Interactive HTML

A proposal for building general intelligence based on biological principles, from a neuron model to a complete artificial life form simulated in a 2D environment, with interactive online examples in eight parts:

1 · The approach required to create “real” AI 2 · The proposed neuron model 3 · Neural circuits with specific functions 4 · Encoding stimuli to neural code 5 · Connectome 6 · Designing an artificial life form 7 · Simulating the life form in a 2D world 8 · Conclusion, references & bibliography

Essays and posts 2016

Guidelines for interaction between humans and artificially intelligent agents 2016
Written in 2016 · Safeguards in AI, Quora. Methods to reduce the stress of human and AI interaction: intention and emotion signalling, predictable behaviour, conveying assumptions, and a human override that is always available.
The need to set up a regulatory authority for Artificial Intelligence based systems 2016
Written in 2016 · Safeguards in AI, Quora. Discusses the case for an authority to standardise AI terminology, enforce accountability, manage incidents, set safety standards and publish independent reports on AI products.
A plausible roadmap for creating Artificial General Intelligence 2016
Written in 2016 · Safeguards in AI, Quora. A staged roadmap towards artificial general intelligence.
Envision the year 2118: what will become reality that is not possible today? 2016
Written in 2016 · Safeguards in AI, Quora. Considers what may follow if AGI is created in the coming decades: artificial life forms, simulating societies to study the effects of legal and economic policy, AI in justice and welfare, and the safeguards each scenario would need.

The dangers of AI weapons and autonomous killer robots



Selected Q&A: AGI, the brain, and AI in society 2016–2017

AGI & risk

Brains & machine learning

AI & society

View all my answers on Quora