The White House has secured a voluntary AI safety pact with six major technology companies. Supporters see it as a way to protect innovation while addressing risks. Critics question whether companies developing increasingly powerful AI should be allowed to police themselves.

US President Donald Trump has brought some of the world’s biggest artificial intelligence companies together around a common promise to make increasingly powerful AI systems safer without putting the brakes on their development.

On September 29, Trump and executives from Google, OpenAI, Anthropic, Meta, Nvidia and xAI signed the White House Accord on Super Intelligence, a voluntary agreement that asks companies to introduce stronger internal controls, external audits and board-level oversight of their AI systems.

But the agreement has also exposed a larger question at the heart of the AI debate: can an industry regulate itself when the companies involved have enormous commercial incentives to keep developing faster?

What does the accord actually promise?

The agreement calls on companies developing frontier AI models to establish robust internal controls to monitor their systems during training and deployment.

Those controls are expected to cover risks including cybersecurity, biosecurity and chemical threats, as well as unintended access by AI systems to technical systems. Companies have also agreed to work with independent external auditors and have their boards review the results of safety evaluations.

The companies will also meet regularly to establish safety standards and best practices.

The catch is that the accord is not legally binding.

Trump described the commitment as “morally binding” and argued that companies have a strong incentive to police themselves because their businesses and reputations are at stake.

The agreement itself leaves open the possibility that some of its measures could eventually be written into law or formal regulations.

Trump’s argument: regulate without slowing innovation

The accord fits with Trump’s broader approach to AI.

His administration has argued that excessive regulation could weaken US technological competitiveness, particularly as Washington sees AI leadership as part of its competition with China.

Trump has therefore favoured industry-led safeguards over a broad new federal regulatory framework. At the White House meeting, he described self-regulation as a way of addressing safety concerns while allowing American AI development to continue at speed.

There is also support for this approach within the technology industry.

Representational image.Image credit:Pexels

Nvidia CEO Jensen Huang has argued that innovation and safety do not have to be opposing goals. Supporters of the accord say companies developing frontier systems are already investing heavily in testing and safety mechanisms, and that flexible standards can evolve more quickly than government legislation.

Critics see a problem with companies marking their own homework

The strongest criticism is about enforcement.

Because the accord is voluntary, it does not establish a clear penalty if a company fails to follow its commitments. Nor does it establish a government regulator responsible for independently verifying whether the promised safeguards are actually working.

Democratic Senator Mark Warner has called for mandatory testing, evaluation and incident-reporting requirements for the most advanced AI systems. His argument is that the companies building powerful AI should not be left to determine the rules governing their own technology.

AI researchers have raised a similar concern. Toby Walsh of the UNSW AI Institute questioned whether companies that have commercial incentives to move quickly should be trusted to assess their own risks.

The Council on Foreign Relations has also argued that the safeguards in the accord are sensible but ultimately limited because companies are not compelled to adopt them.

The bigger question is who gets to set the rules

The disagreement is therefore not simply about whether AI needs safety measures. There is broad recognition that increasingly capable systems require safeguards.

The dispute is over who should enforce them.

Trump’s model places much of that responsibility with the companies developing frontier AI. Critics want a stronger role for governments, independent regulators and international institutions.

The debate is becoming more urgent as AI systems are used in cybersecurity, scientific research, government services and other sensitive areas. Recent incidents involving AI systems gaining unintended access to computer systems have also intensified concerns about whether existing safeguards can keep pace with technological development.

The White House accord therefore represents a significant shift in the conversation, but not necessarily a resolution.

It acknowledges that powerful AI needs stronger safety controls while leaving the companies themselves largely responsible for implementing them.

India has deployed 40 supercomputers with a combined capacity of 68 petaflops (PF), according to a new government update. These machines are being used for more than scientific research. They are helping forecast floods, model forest fires, track urban pollution and support drug discovery.

The figure is part of the latest update on the National Supercomputing Mission (NSM), which was launched in 2015 to expand India’s high-performance computing capacity. The mission now plans to take the number of supercomputers to 50, with a combined capacity of more than 123 PF.

A petaflop measures how many floating-point calculations a computer can perform in one second. One petaflop equals one quadrillion calculations per second. Supercomputers achieve this by using large numbers of processors that work on calculations simultaneously.

Supercomputing Predicting Floods and Forest Fires

Flood forecasting is one of the clearest examples. An early warning system developed under the mission uses data analysis and predictive models to forecast floods in river basins up to two days in advance. It is being used for the Mahanadi River basin.

The system can process large amounts of data to estimate how flood conditions may develop. For communities in vulnerable areas, the value of the technology lies in the warning it can provide before floodwaters arrive.

Another application focuses on forest fires. The Forest Fire Spread Model combines satellite remote sensing with computational models to predict how a fire could move across a landscape. It has been tested in regions including the Sikkim Himalayas. Instead of simply showing where a fire is already burning, the model can help estimate where it could spread next.

Models Tracking Pollution and Rainfall

Supercomputing is also being applied to urban environmental problems. The Urban Environment Decision Support System uses detailed models to track weather and air pollution. It can predict heavy rainfall and pollution events, giving cities information that can be used for preparedness and mitigation.

The government says India generates nearly 20% of the world’s data. Artificial intelligence, weather forecasting and space research are also increasing the need for computing systems that can handle large datasets.

Supercomputers Used to Look for Potential Drugs

The applications extend into healthcare research. A genomics and drug-discovery platform analyses large sets of molecules to help researchers find and test potential drugs. According to the government, the platform was used during the COVID-19 pandemic to screen existing drugs and predict possible side effects, including cardiac risks.

Another application in materials science and computational chemistry allows researchers to simulate atoms, molecules and alloys and study their behaviour.

India Building Technology Behind the Machines

The National Supercomputing Mission was launched in 2015 with an outlay of about ₹4,500 crore. It is jointly steered by the Department of Science and Technology and the Ministry of Electronics and Information Technology, with C-DAC, Pune, and IISc, Bengaluru, implementing the mission. The mission is also focused on developing supercomputing technology within India.

C-DAC has developed the Rudra series of servers used in PARAM Rudra supercomputers. As of September 2026, 6,000 Rudra servers had been deployed, while another 1,500 were under manufacturing, according to the government.

The mission has also developed high-speed interconnect networks, indigenous cooling technology and software for high-performance computing systems.

Beneficial to More than 16,000 researchers

The computing infrastructure is being used across universities and research institutions. The government says more than 16,000 researchers, including over 2,900 PhD scholars, across more than 400 institutions have used NSM-supported infrastructure. The systems have handled more than 1.5 crore compute jobs and contributed to over 1,990 research publications as of September 2026.

The mission is also running workshops, hackathons, faculty programmes and courses to train students and researchers in high-performance computing, artificial intelligence and deep learning.

India Targeting More Than 123 PF

India’s supercomputing capacity is expected to grow further. The National Supercomputing Mission plans to establish 50 supercomputers with a combined capacity exceeding 123 PF.

The next phase will focus on faster and more energy-efficient systems and wider access for researchers, universities and industry. The government also plans to integrate artificial intelligence with high-performance computing for work in weather and climate, healthcare, agriculture, energy, drug discovery and engineering.

For ordinary users, the significance of that expansion is easier to see through its applications than through the petaflop figure itself. A flood forecasting system can provide up to two days of warning in the Mahanadi basin. A forest-fire model can estimate how a blaze may spread. Urban models can track pollution and heavy rainfall, while drug-discovery systems can help researchers screen potential medicines. The 68-PF figure, then, is less about how fast India’s computers can calculate and more about the problems they can help solve.

Around 10 pm on 30 September 2025, a worker went down into the septic tank of a hotel in Kattappana, in Kerala’s Idukki district, to clean it. He became trapped. Two colleagues went in to pull him out and were trapped too.

The manhole was too small for rescuers to enter. The fire force brought in an earthmover to break open part of the structure, and after a rescue operation of about an hour and a half the three men were taken to hospital, where they could not be saved. They were Jayaraman, from Cumbum in Tamil Nadu’s Theni district, and Sundara Pandian, 36, and Michael, 23, both from Gudalur in the Nilgiris. The Idukki District Collector’s report later attributed the deaths to inhalation of toxic gas in a tank with critically low oxygen, though the gas itself was not identified. It recommended a multi-department inquiry and compensation for their families of the dead involved in manual scavenging.

Kerala is the state where this is least expected. In the dataset used for this report, it recorded three sewer deaths in the decade to 2025: two in 2016, one in 2017, and none since. The men of Kattappana do not appear in that count. Nor does Joy, a sanitation worker who drowned in a waste-choked canal in Thiruvananthapuram in July 2024. Every figure in this story is a minimum.

Between 2016 and 2025, the National Commission for Safai Karamcharis (NCSK) recorded 662 deaths across India in connection with hazardous sanitation work. India banned the employment of manual scavengers in 1993. The 2013 Act went further, barring anyone from being employed to clean sewers and septic tanks hazardously, that is, without the protective gear and safeguards its rules prescribe. Yet sanitation workers still enter through narrow openings, sometimes with only basic protective equipment, to clear blockages and clean spaces where sewage, toxic gases and human waste accumulate. The work can take hours; in the wrong conditions, it takes lives.

The toll peaked in 2019, at 124 deaths. Tamil Nadu recorded the most over the decade, at 93, followed by Maharashtra (84), Gujarat (83), Uttar Pradesh (81) and Haryana (69); Delhi recorded 65. The five states named accounted for about 62 per cent of the deaths. Since the NCSK began counting in 1993, it has logged 1,361 such deaths, 264 of them in Tamil Nadu and 13 in Kerala.

Official series differ by source and date, because states report late. Replies in Parliament have put the toll for 2019 anywhere between 116 and 133, and a reply in March 2026 that counted from 2017 onwards ranked Uttar Pradesh first (86), ahead of Maharashtra (82) and Tamil Nadu (77). The pattern is consistent even if the totals are not.

A note on terms also matters here. The government draws a legal line between “manual scavenging”, which the 2013 Act defines as lifting human excreta from insanitary latrines, and the “hazardous cleaning” of sewers and septic tanks. It has told Parliament that no deaths have been reported from the first, while counting those in the second. Campaigners treat both as the same caste-bound labour. This report uses “manual scavenging” for the wider practice and “sewer deaths” for the deaths.

Behind the toll lies a harder question: if India has achieved significant technological advancement, why does manual scavenging still persist at all? What keeps people entering spaces they should never have to enter?

For Bezwada Wilson, national convenor of the Safai Karmachari Andolan (SKA) and a Ramon Magsaysay Award winner, the fight against manual scavenging has always been about more than sewers. Through the SKA, he helped document the practice, challenge official claims and take the issue to the Supreme Court. The Supreme Court’s 2014 judgment recognised the continuing problem, ordered the identification and rehabilitation of workers, and directed compensation of ₹10 lakh to the families of those who died in sewers, bringing dignity, equality and caste discrimination to the centre of the debate.

The Work Begins Long Before the Sewer

Wilson grew up in a community that has been tied to sanitation work for generations, and he has spent much of his life challenging the system that kept it that way. As a child, he says, he studied in a separate school near his settlement, where caste and language shaped where children from his community were sent to learn. He came to the movement against manual scavenging through an understanding of what prolonged contact with human excreta does to a person’s health and life.

People who come into contact with human excreta, he says in an interview with EdPublica, are exposed to deadly diseases. Their quality of life deteriorates. That, for Wilson, was reason enough to act. But the deeper problem is what happens when a community is identified with one kind of work for generations. Even education does not necessarily erase that association.

Wilson says government programmes aimed at upskilling members of sanitation-worker communities exist, but questions whether they are reaching people effectively. The Ministry of Social Justice and Empowerment told the Lok Sabha in 2022 that it had spent ₹266.16 crore on its self-employment scheme for the rehabilitation of manual scavengers since 2013-14; that scheme has since been replaced by NAMASTE. Students in these communities still face difficulties accessing education in many places.

Then there is discrimination. When people have been treated as inferior for generations, Wilson argues, a change in law does not automatically change the way society sees them. That matters because manual scavenging is not simply a story about a hazardous job. It is also a story about who is expected to do it.

The Number that Remains Visible

The 662 deaths recorded between 2016 and 2025 are a floor, not a full count. They are the deaths captured in the available dataset, and that distinction matters: the scale of the problem is larger than the deaths that make it into official or reported records. The NCSK says its figures draw on state governments’ reports, media reports and complaints it receives.

The gap between official and civil-society counts is wide. The Ministry told the Lok Sabha in August 2026 that 498 people died in hazardous sewer and septic-tank cleaning between January 2019 and June 2026, a toll that fell from 132 in 2019 to 47 in 2025. The SKA’s own tracking runs the other way, as the table shows. It has counted 113 deaths in the first seven months of 2026 alone; the government’s figure for the first four months is 52.

The two series cannot both be right, and the difference is not only about counting. It is also about classification, since the government counts hazardous cleaning but not “manual scavenging”. Even among the deaths that are recorded, accountability is slow: the NCSK says compensation has been paid in 1,206 of the 1,361 cases logged since 1993, with 101 still pending.

Vimal Govind M K, founder and CEO of Kerala-based Genrobotics, came to the problem from the technology side. His company turned its robotics towards sewer cleaning, he says, after the founders came across a newspaper photograph of a man who had died inside a sewer in Kozhikode, Kerala. The image changed the direction of their work.

“We decided to pivot the technology we were developing towards a completely different problem—to prevent people from entering sewers and doing this dangerous work manually,” Govind said.

That decision eventually produced Bandicoot, a robotic system designed to clean manholes without sending a person inside. Genrobotics says it was first deployed in Kerala in 2018. But the existence of a machine does not mean people stop entering sewers.

A Robot Exists. So Why are People Still Entering?

Genrobotics says it has trained more than 3,000 workers and has around 400 robots operating in India. Its public figures have varied: a March 2026 press release spoke of more than 275 deployments, while earlier interviews cited 300 and 350 robots. Govind says a task that can take three to four hours when done manually can be completed in around 20 minutes using the technology. Operating one robot takes two workers, he says.

Figures from the field point the same way. Chennai’s water board has been reported to have cleaned more than 5,000 manholes in the first year of deployment, and a company case study lists 1,111 manholes cleaned by two robots for Mumbai’s civic body between September 2024 and March 2025. Both figures come from the company or from secondary reports.

On paper, the logic seems straightforward: if a machine can enter a sewer, a human being should not have to. But India’s sanitation infrastructure does not operate on paper. Municipalities and water boards buy machines, yet many of the deadliest jobs are commissioned by someone else. Kattappana was a hotel’s tank. This week in Mumbai’s Santacruz East, Santosh Kashate, 50, died and Ajay Thakur, 30, was hospitalised after they allegedly entered a municipal sewer manhole while cleaning a housing society’s drain. Police were registering a case of death by negligence against the contractor, and said society members could also be booked for hiring private workers without informing the Brihanmumbai Municipal Corporation.

Some facilities need specialised cleaning only occasionally. Private establishments have little reason to invest in technology for a task they might need once a year, or once in several years. Govind points to another problem: safety enforcement.

In India, confined-space management and workplace safety remain uneven, he says. In countries such as Singapore and the UAE, workers require safety training and permits before undertaking such work, with safety procedures built into the job itself. India’s problem is therefore not simply a shortage of machines. It is also a shortage of systems that make entering the sewer unacceptable.

The courts have pushed in that direction. In October 2023, in Dr Balram Singh v Union of India, the Supreme Court raised compensation for sewer deaths from ₹10 lakh to ₹30 lakh. In January 2025 it barred manual sewer cleaning in six metropolitan cities, and in July 2026 it issued show-cause notices on contempt to the chief secretaries of five states. After two contract workers died in a Tiruchirappalli manhole on 22 September 2025, the corporation announced ₹30 lakh for each family and filed a complaint against the contractor. Compensation and complaints now follow such deaths. Prevention still lags.

When Protective Equipment is not Enough

For years, the response to hazardous sanitation work has often included protective equipment—masks, gloves and other safety gear. But Govind says equipment alone does not necessarily change behaviour. He says workers who have spent years entering sewers can become accustomed to the practice, and that even after being provided with protective equipment, some continue to enter because that is how the work has traditionally been done.

Official figures show the limits of the equipment route. Under the government’s NAMASTE scheme, 89,915 sewer and septic-tank workers had been validated by August 2026, and about 87,000 protective kits had been distributed by the scheme’s third anniversary in July. The Lok Sabha was also told that 52 people died in hazardous cleaning in the first four months of 2026. The figures do not say who the dead were or whether they had been profiled. They do show that distributing equipment has not, by itself, ended the deaths.

Changing that, Govind says, requires training and a cultural shift. That observation brings the technology debate back to Wilson’s argument. If caste has historically determined who performs sanitation work, and if generations of workers have been conditioned to see hazardous cleaning as part of their occupation, then a robot can solve only one part of the problem. It can keep a person out of the sewer. It cannot, by itself, determine what happens to that person afterwards.

What Happens to the Worker When the Robot Arrives?

This may be the most important question in the mechanisation debate. Automation is often discussed in terms of machines replacing people. In sanitation, however, the better outcome is different: workers moving from hazardous manual entry to safer, skilled roles.

Genrobotics says sanitation workers are trained to operate its technology and become part of its teams. The company said in January 2025 that more than 1,000 of the workers it had trained had become robot operators. It says the transition removes direct interaction with sewage from their work. A worker who once entered a manhole could instead operate a machine from outside it.

The job remains. The danger does not have to. Those numbers are the company’s own, though. What would settle the question is what operators earn, who employs them and whether the jobs last. The model also depends on municipalities and other employers actually investing in mechanisation, training workers and enforcing the rules.

Govind says budget allocation remains one of the major barriers to adoption. Government programmes and funds can support mechanisation, but ultimately policymakers and local administrators have to decide to spend the money. The government’s main programme is NAMASTE, launched in July 2023 with an outlay of ₹349.73 crore over three years to 2025-26 across more than 4,800 urban local bodies. It profiles sewer and septic-tank workers, provides protective kits and health insurance, and offers capital subsidies of up to 50 per cent for sanitation machines and enterprises. An extension to rural areas, with an additional outlay of about ₹498.73 crore for 2026-27 to 2030-31, has been reported.

The Geography Tells Another Story

The deaths are not evenly distributed across India. Kerala, by comparison, recorded three deaths in the dataset, two in 2016 and one in 2017, with none from 2018 to 2025.

That contrast is worth examining, but not celebrating too quickly. A decline in reported deaths does not, by itself, prove that hazardous sanitation work has disappeared. Differences in reporting, infrastructure, enforcement and the kinds of sanitation work carried out in each state can all affect the numbers. Kattappana and Joy’s death in Thiruvananthapuram sit outside Kerala’s count.

Govind says Kerala has stronger safety practices than many other states, while acknowledging that enforcement and investment remain challenges nationally. Genrobotics first deployed Bandicoot in Kerala in 2018, the year the zero run begins. The timing is suggestive, not proof, and Kattappana shows what a municipal robot does not reach: a private tank behind a hotel.

The more useful question, then, is not simply which state has the fewest deaths. It is what is different about the systems that keep workers out of hazardous spaces.

Besides the Number 662

India’s response has increasingly moved towards mechanisation. The government has also introduced programmes aimed at eliminating hazardous sanitation work, rehabilitating workers and creating safer sanitation systems.

The experiences of Wilson and Govind point to two requirements for the same goal. One is technological: make it possible to clean a sewer without sending a person inside. The other is social: make sure the person who no longer enters the sewer has the education, income, skills and social acceptance to move into safer work. Neither can replace the other. A third, the law, is meant to hold them together.

A robot can pull a worker away from toxic sewage. A law can prohibit an employer from sending that worker back in. A rehabilitation programme can offer another livelihood. But unless all three work together, the prohibition remains vulnerable to the same pressures that created the occupation in the first place.

Govind predicts that the sanitation industry could become almost completely automated within the next five to six years, with very little human involvement in hazardous work. It is an ambitious prediction. Against more than 4,800 urban local bodies, the company’s own count is a few hundred robots.

The technology may arrive before the social transformation does. And that is the real test for India. Not whether it can build a robot capable of entering a sewer. But whether it can build a system in which no human being has to.