Two supply shocks are hitting a small analytic workforce at once — open source becoming first-class collection, and AI that is already operational in some missions and arriving fast in others. For a small community like Australia's National Intelligence Community the shift is leverage, provided it invests in data engineering, evaluation and analytic standards rather than demo-ware.
Open source as first-class collection
Commercial satellite constellations, AIS and flight tracking, leaked and stolen datasets, social media, procurement records, corporate filings, and data brokers have shifted the balance between classified collection and what a good analyst can do unclassified. For a country of Australia's size this is leverage: OSINT is cheaper, faster and shareable with ministers, ADF operators, and regional partners who cannot sit on Five Eyes systems. AGO now lives in a market where Maxar, Planet and others sell imagery that was national-technical-means territory a decade ago. The corresponding risks are manipulation (including state-seeded OSINT), provenance, and the temptation to treat "it was on the internet" as corroboration — sourcing and deception are now daily, not specialist, collection tradecraft.
AI: where it is already real and where it is not
- Already operational in analogue form: computer vision for object detection in GEOINT, speech-to-text and machine translation for SIGINT, anomaly detection over large transactional or network datasets. Volume in ASD-type missions exceeded human reading capacity long ago; models are not optional.
- Arriving fast: LLMs for first-pass summarisation, reporting and discovery across cables and traffic. The catastrophic failure mode is hallucination inside an assessment that still reads as authoritative. ONI's product is endorsed through the National Assessments Board; a black box that cannot be sourced or confidence-rated does not survive that process unless humans remain accountable for every analytic line (Estimative Language).
- Security: commercial models are generally unusable on classified holdings, implying accredited, air-gapped or appropriately partitioned systems — which Australia will often buy or adapt with partners rather than build alone. AUKUS Pillar II and Five Eyes R&D are the practical path; they also deepen dependence.
- Adversarial use: deepfakes and synthetic media as denial and deception; AI-accelerated phishing and influence; PRC and other services using machine analysis against Australian targets. Analysts now need a working literacy in how the other side's models fail, not just their own.
Institutional consequences
Junior-analyst work (first drafts, compilation, watch-keeping) is the obvious automation target; senior judgement, challenge and contextualisation are not. A small community can in principle gain more from this shift than a large one — if it invests in data engineering, evaluation and analytic standards. ONI's enterprise-manager role (post-2017) is the logical place to set community standards for AI-assisted assessment; whether it has the authority and technical depth to do so is part of what periodic intelligence reviews exist to test. DSTG, universities and industry will do much of the tool-building; the NIC has to decide what it is willing to trust.
Related
- National Intelligence Community — the analytic workforce being rewritten
- Structured Analytic Techniques — the tradecraft baseline AI must integrate with
- Estimative Language — confidence-rating discipline under AI-assisted drafting
- ASD Project Redspice — the volume-driven missions where models are not optional