Training and deploying deep machine learning models for automated information retrieval and fact verification: A security studies perspective on countering disinformation

The weaponization of information has become a defining feature of contemporary conflict, and generative artificial intelligence has lowered the cost of producing disinformation at scale. This paper examines how deep machine learning models for automated information retrieval and fact verification can be trained, deployed and rigorously evaluated as instruments of national security, with particular reference to Pakistan. Drawing on securitization theory and the resilience framework, it proposes and implements a human-supervised pipeline of claim screening, lexical–dense evidence retrieval, deep verdict prediction, calibration and conformal triage. The analysis uses only real data: a compiled cross-national dataset on organized manipulation, platform takedowns, fact-checking capacity, public perception and benchmark performance, together with two human-annotated claim corpora (12,836 political statements and 1,535 climate claims with 7,675 claim–evidence pairs). Six deep architectures were trained in Python, and the evidence was analysed with fractional-logit quasi-maximum likelihood with wild bootstrap inference, logistic diffusion and unknown-break structural change tests, ordered probit, two-way clustered fixed-effects models, double/debiased machine learning with neural-network nuisance functions and Oster coefficient-stability bounds. Organized manipulation grew by 42.5% a year between 2017 and 2020, whereas fact-checking capacity saturated near 468 projects and has declined since late 2023. On real claims, text-only deep models did not outperform classical baselines; gains came from source metadata; between a quarter and two-fifths of correct verdicts lacked supporting evidence; and a low-cost typographic attack erased the best model’s advantage. Calibrated conformal triage, however, delivered guaranteed coverage. The paper concludes that automated verification should be built and judged as a strategic triage capability that supports, rather than replaces, accountable human analysts.