On January 29, the world’s largest religious gathering, Maha Kumbh in Prayagraj, was struck by a horrifyingly familiar tragedy — a stampede that claimed 48 lives.
This despite decades of experience in organising similar congregrations, improved technology, and advanced crowd-monitoring systems. A stampede during the Kumbh in 1954 claimed no less than 700 lives.
A study titled ‘At the Mahakumbh, Faith Met Tragedy: Computational analysis of stampede patterns using machine learning and NLP’, by Abhinav Pratap of Amity University, examines the recurring failures in crowd management at India’s grandest pilgrimage. Applying machine learning, historical data analysis, and natural language processing (NLP) to seven decades of administrative records, the research uncovers a disturbing reality: stampedes at the Kumbh Mela are not accidents but predictable failures, resulting from infrastructural limitations, governance inertia, and a tendency to normalise disaster.
Integrating computational modelling and sociological theories, the researchers designed a three-phase analysis to derive insights from past tragedies, including patterns of administrative failure and recurring risk factors. They then placed the findings within the framework of the Emergent Norm Theory (which explains how collective behaviour in large crowds overrides individual rationality) and Institutional Amnesia Theory (which describes how organisations, institutions, or even societies gradually lose their collective memory of past events due to staff turnover and lack of proper documentation).
A recurrent crisis
The analysis of stampede incidents between 1954 and 2025 throws up a striking pattern — critical crowd density thresholds (equal to or more than seven persons per sq m) consistently lead to deadly outcomes. Once this limit is breached, individual control is lost, and panic spreads through the crowd like a chain reaction.
VIP movement
One of the most alarming findings is that 92 per cent of past stampedes occurred near infrastructural choke points — narrow pathways, riverbank access routes, or restricted entry points that become dangerously overcrowded.
The study also highlights the role of restrictions due to VIP movement, which redirect police forces and create dangerous bottlenecks. Even in 2025, drone footage revealed barricaded exits left unmanned during a ministerial visit, exacerbating the deadly crush.
The NLP-driven analysis of official stampede inquiry reports from 1954 to 2025 repeatedly show administrative narratives that deflect blame from authorities:
• 1954: “Unforeseen surge”
• 1986: “Crowd became unruly”
• 2003: “Poor coordination”
• 2013: “Railway station mismanagement”
• 2025: “Barricade collapse”
Instead of acknowledging systemic failures, these reports frame stampedes as random and unavoidable incidents.
Improved governance reduces fatalities, but its impact remains limited due to systemic failures.
In this case, there were three failures: Delayed emergency response, as reflected by the inquiry reports from 1954, 1986, 2003, and 2025; VIP route prioritisation disrupts normal crowd flow, creating bottlenecks that intensify stampedes; and AI-based risk predictions were ignored, leading to reactionary, rather than preventive measures.
The 2025 stampede occurred at a barricade breach where drone footage showed unmanned exits — a repeat of the failure from 1954.
