AI could cut a submarine’s survival chance to 5%: Chinese defence scientists
Era of ‘invisible’ submarines ending with next-gen tech that could prevent one in 20 from escaping attack
A new defence industry study from China suggests that artificial intelligence (AI) could soon make it extremely hard – even nearly impossible – for submarines to survive in a future naval conflict.
The research, published in the peer-reviewed journal Electronics Optics & Control and led by senior engineer Meng Hao with the China Helicopter Research and Development Institute in August, unveiled for the first time an advanced AI-driven anti-submarine warfare (ASW) system capable of hunting even the quietest submarines through intelligent, real-time decision-making.
According to the research, the new ASW system could reduce a submarine’s chance of escape to just 5 per cent, meaning only one out of every 20 submarines would likely escape detection and attack.
As global powers intensify their race to put AI into military use, the study suggests the era of the “invisible” submarine – long a cornerstone of naval deterrence – may be coming to an end.
Instead of relying on old search patterns, the AI system acts like a smart commander in the ocean.
It uses data from sonar buoys dropped by helicopters, underwater sensors, radar and even ocean temperature and salt levels to build a live picture of what is happening under the sea.
Then, it quickly decides where to look, how to adjust its equipment and how to respond when a submarine tries to escape by zigzagging, going silent or releasing fake signals to throw off the hunters.
In computer simulations, this AI system was able to find and track enemy submarines about 95 per cent of the time, no matter how hard they tried to hide.
Even when submarines used hi-tech decoys or drones to distract the searchers, the AI kept up and stayed on their trail.
Submarines have long been viewed as the ultimate asymmetric weapon: able to deliver nuclear strikes, gather intelligence or sink carrier groups while remaining nearly undetectable.
The US Navy, for instance, maintains that large fleets of nuclear submarines could be a deterrent to the rapidly growing PLA Navy in a future conflict.
The US nuclear submarine fleet consists of about 70 nuclear-powered submarines as of mid-2025, according to openly available information.
These submarines could blend into the background noise of the ocean and they could carry cutting-edge drones to distract the hunters.
In traditional anti-submarine warfare, a quiet submarine equipped with advanced decoys has a survival chance as high as 85 per cent, making them “one of the biggest threats” to China’s surface fleets, according to Meng’s team.
But AI could make the strategy outdated.
“The ultimate success rate keeps stable at around 95 per cent,” Meng and his colleagues said.
The system operates on a three-layer architecture: perception, decision making, and human-machine interaction.
First, the AI fuses real-time data from sonar, radar, magnetic anomaly detectors and oceanographic sensors to build a dynamic picture of the undersea environment. It accounts for shifting variables like water temperature, salinity and background noise – conditions that traditionally hinder sonar effectiveness.
Then, in the decision layer, a multi-agent reinforcement learning model pits AI “hunter” agents such as helicopters and sonobuoys against simulated “prey” agents, including submarines and unmanned underwater vehicles.
Through thousands of simulated engagements, the AI learns optimal tactics, such as forming sonar barriers, executing coordinated sweeps or concentrating sensors in high-probability zones.
Critically, the AI does not just detect submarines, it anticipates their behaviour. In simulations, it recognised evasive tactics like “silent running” or zigzag manoeuvres and adjusted search patterns accordingly.
Even when submarines deployed decoys or operated in complex acoustic environments, the AI maintained a high detection rate, according to the researchers.
The project team also built large language model-based assistant interfaces to coordinate different AI agents’ interaction with human operators, translating complex sensor data and AI-generated strategies into plain-language recommendations.
It dramatically reduced cognitive load during high-pressure missions, they said.
The researchers say this technology could still get even better.
Future versions could team up with drones in the air, ships on the surface and unmanned vehicles underwater, creating a complete “three-dimensional” hunting network.
The AI could also keep learning during real missions, getting smarter with every deployment.
And lighter, faster versions could be installed on smaller combat robots, allowing more platforms to make quick decisions without needing to send data back to a central base.