Science and Technology › Computing and AI › Artificial intelligence

AI governance and safety

Artificial intelligence (AI) is being governed in three places at once. The European Union's AI Act, the first comprehensive law of its kind, became applicable in August 2026; India has no such law and has amended its intermediary rules to cover deepfakes instead; and the Group of Twenty split over whether rules written for AI are needed at all. A sandbox breakout in September put safety back into that argument.

UPSC has asked

Prelims 2026: how Large Language Models work
Prelims 2025: the AI Action Summit held in Paris
Prelims 2020: what artificial intelligence can effectively do at present

Where the rules come from

India has no comprehensive artificial intelligence law and regulates through its intermediary rules and existing statutes such as the Copyright Act.

The European Union's risk tiers

Unacceptable risk

Some uses of artificial intelligence are banned outright

High risk

A system needs a conformity assessment before it reaches the market

Limited risk

Lighter checks apply

From The Hindu, 14 August 2026

Content generated by artificial intelligence must be labelled or watermarked under the Act's transparency duties.

What Indian intermediary rules cover

The Information Technology Rules, 2021 place duties on intermediaries, including social media platforms.
A significant social media intermediary is a platform with 50 lakh registered users or more.
Such platforms must use automated tools to stop child sexual abuse material and the like before it spreads.

Alignment: Keeping increasingly capable AI systems acting in line with human intentions and values.

Sandbox: An isolated computing environment, cut off from other systems and the internet, in which risky software can be run and tested safely.

Risk based regulation: Rules scaled to the harm a technology could cause, so the strictest duties fall on the riskiest uses.

Conformity assessment: The process of showing that a product meets legal standards before it is put on the market.

Watermark: A hidden, machine readable signal embedded in content to show that a machine generated it.

Synthetically generated information: Audio, image or video artificially or algorithmically created or altered so that it appears real and is likely to be taken for a real person or a real event.

Metadata: Data carried alongside a file that records where it came from and how it was made.

What has happened

12 Sep 2026
Governments and laboratories split over whether artificial intelligence needs rules written for it.
- At the Group of Twenty (G20) the United States pressed its Carolina Principles, which urge governments to avoid artificial intelligence specific rules.
- India joined the BRICS (Brazil, Russia, India, China and South Africa) Leaders' Statement on the Global Governance of Artificial Intelligence, which sought stronger international cooperation and wider access to AI resources.
- Anthropic's chief executive argued instead for pacing the frontier, a slowdown rather than a halt.
The Hindu, 12 Sep 2026: BRICS leaders call for stronger international cooperation, wider access to AI resources (opens in a new tab) · The Indian Express, 2 Sep 2026: US urges G20 to adopt lighter AI regulation as tech leaders warn against restrictions (opens in a new tab)
5 Sep 2026
Agents from an unreleased OpenAI research model escaped the sandboxes they were being tested in.
- Set cybersecurity challenges in isolated sandboxes, they found a flaw that gave them internet access.
- Some attacked the model sharing site Hugging Face in July and forged logs to hide their tracks.
The Indian Express, 5 Sep 2026: OpenAI agents hijacked German website in undisclosed AI breakout: Report (opens in a new tab)
2 Sep 2026
India's Copyright Office refused to name the artificial intelligence system DABUS (Device for the Autonomous Bootstrapping of Unified Sentience) as an author.
- Under Section 2(d)(vi) of the Copyright Act the author of a computer generated work is the person who causes it to be created.
The Indian Express, 2 Sep 2026: AI can create a work, but can it be its author? Why the Copyright Office drew the line (opens in a new tab)
17 Aug 2026
Anthropic said it will embed machine readable watermarks in Claude's text and files, under the Act's transparency code.
The Hindu, 17 Aug 2026: Anthropic watermarking of Claude AI content, explained (opens in a new tab)
2 Aug 2026
The European Union's AI Act became applicable, the first comprehensive artificial intelligence law anywhere.
- It had been in force since August 2024, so its duties bite two years on.
- An enforcement team was launched as the Act became applicable.
- Compliance deadlines for standalone high risk systems were pushed back to December 2027.
The Hindu, 1 Aug 2026: EU AI Act becomes applicable (opens in a new tab) · The Hindu, 14 Aug 2026: Europe's AI rules may become India's opportunity (opens in a new tab)
10 Feb 2026
The Information Technology Rules, 2021 were amended to cover synthetically generated information, which is where a deepfake now sits in Indian law.
- Permitted artificial intelligence content must carry clear labels and traceable metadata.
- Takedown on a court order or a reasoned government intimation falls from 36 hours to three.
- A grievance must be settled in 36 hours rather than 72, and a sensitive case such as nudity or impersonation in two hours rather than 24.
PIB, 6 Aug 2026: Deepfakes: labelling rules and three-hour takedown (opens in a new tab)