'The AI did it' is no excuse, says Sanjeev Sanyal

The EAC-PM member says no country has yet found a good way to regulate AI. He argues that partitions, audits and "skin in the game" must be in place before AI spreads across the economy.

ET AI

Sanjeev Sanyal, membe​r of the Economic Advisory Council to the Prime Minister

No country in the world has yet found a good way to regulate artificial intelligence (AI), and India should borrow from the way it regulates its financial system to fill that gap, believes Sanjeev Sanyal, member of the Economic Advisory Council to the Prime Minister (EAC-PM).

“AI is obviously an important tool, but it has certain emergent behaviours. It evolves in different ways, sometimes difficult to predict or explain how it is behaving, and handing over large systems to such a technology has obvious risks,” Sanyal told The Economic Times Digital at the Bengaluru Tech Summit. “If we create an integrated world run by such a system, it creates a single point of failure”

He pointed to the recent incidents in the US, where the entire civil aviation sector was suddenly stalled by a single point of failure, followed by a major problem with the financial and banking system.


“You will get butterfly effects as well. Something you think is low risk, but has then infected all points in the system and suddenly goes rogue in some fashion. You lose control of it. So this has to be regulated in some fashion. And unfortunately, no country in the world yet has created a good way of doing it,” he said.

Lessons from the financial system

Sanyal said the financial system is also a complex adaptive system. It evolves in ways even its regulators cannot predict, yet over the last 200 to 300 years humans have built tools to keep it working despite occasional breakdowns.
ADVERTISEMENT

Regulators contain shocks by building “walls and partitions” into the system, barring banks from certain activities and using circuit breakers to halt trading when things go wrong. They also force transparency. “Every company that is listed has to have audited financial records. You cannot just say that, you know, my CA came up with this piece of paper. The board has to explain how those numbers came,” Sanyal said, adding that AI needs similar explainability audits.

Accountability, however, is what AI lacks most. When an AI system fails today, it is unclear whether the user, the company that deployed it, the developer or whoever supplied faulty training data should answer for it, and because AI keeps evolving on its own, “nobody’s in charge”, he said.

“Everybody’s blaming, ‘No, no, AI did it because I didn’t programme it. It created this on its own.’ This should not be an excuse. We have to ex ante state who has skin in the game, who is to be blamed. Then you will have respect for the law, and you will also have an incentive to keep track of what’s going on,” Sanyal said.

Why Gen AI is harder to regulate
ADVERTISEMENT

Sanyal questioned the global race to build general AI highlighting the challenges of regulated an interconnected system built using AI.

“Everybody’s in pursuit of general AI intelligence. This is inefficient in multiple ways. One, not the least, is that you're going to run the whole system in order to do particular tasks, when in fact it's efficient to have localised systems. If you have an interconnected system, AI of all things, then it's impossible for people to regulate it. You'll need some great regulator who knows everything,” he explained.
ADVERTISEMENT

He would rather have existing regulators oversee AI in their own sectors. “In the financial system, you have SEBI, you have RBI. If you deploy it in telecommunications, you have TRAI. I have come to the view over time that a horizontal regulator that knows everything may be too difficult,” he said.

“Human in the loop is absolutely critical. So you need choke points where credentialised humans have to basically take the decision for the next step," Sanyal added. He expects the financial sector to be safer than most because these safeguards already exist, but sees electrical grids, for instance, as vulnerable.

“People who run electrical grids think in engineering terms. They simply do not think in complex adaptive system terms,” he said.

DEPA: the algorithm goes to the data

Sanyal warned that data shared in good faith can be manipulated, citing open-sourcedp platform Wikipedia.

Elaborating the need to work with unbiased data, he said: “It is actually the general public who, in good faith, shared information. But then a system of editors was deliberately imposed as a layer on it, and now Wikipedia, and therefore the training of AI, a lot of which happens on Wikipedia, is being done on data that has already been tilted in a particular way. So data should not be shared without caution. But at the same time, you need it to train things.

He pointed to DEPA, the Data Empowerment and Protection Architecture, as one answer.

“The idea which India is doing is creating a sandbox where the data does not go to the algorithm developer. Instead, the algorithm developer has to come to the data. It could be GST, AIIMS, medical data, whatever it is. You create a box, you put the data in it, and then those who want to train the algo come to that sandbox, train it, and leave. They don't own the data,” he said.

DEPA was first proposed by NITI Aayog in 2020 for consent-based data sharing. On October 30, the Ministry of Electronics and IT released a draft handbook in Bengaluru to extend it to AI training, under which developers receive only the trained model and never the raw data.

“This is actually quite an interesting thought, an experiment we are doing. I think this is an example of how this can be done,” Sanyal said.
Download
The Economic Times Business News App
for the Latest News in Business, Sensex, Stock Market Updates & More.
READ MORE
ADVERTISEMENT

READ MORE:

LOGIN & CLAIM

50 TIMESPOINTS

More from our Partners

Loading next story
Business News › AI › AI Insights › 'The AI did it' is no excuse, says Sanjeev Sanyal
Text Size:AAA
Success
This article has been saved

*

+