TLDR: AI can collect and compress, but it can’t yet replicate the human judgment that HUMINT collection runs on: who knows this, why do they know it, what would make them lie. That draws on human intuition, human skills, and a feel for human behaviour built across years of contact.
Human-machine teams will define the strongest intelligence organizations of the next decade, but only where the human half keeps the capacity to interrogate the machine’s output. An AI agent can surface patterns, but holding an informant accountable over time remains a job for human operatives.
A case officer in Nairobi is working a promising human source that says he runs procurement at a defence contractor.
Over four months, they trade messages, then a voice call, then video. His story holds up: face, voice, and even his colleagues’ posts.
Then a counterintelligence team examines the artifacts: the engineer never existed. An AI system built him frame by frame, run by a hostile service to feed a rival a poisoned stream of sensitive information.
That sketch points at the finding on AI HUMINT. AI won’t retire the human intelligence officer, but it’s corroding the trust conditions that human source work runs on.
There are three forces at play:
- Generative AI lets adversaries spin up synthetic personas at industrial scale.
- Agentic AI runs automated deception end to end for almost nothing.
- Deepfakes have crossed the point where ordinary people, and some institutions, can’t tell a clone from a person.
Each force attacks the ability to determine whether a human source is who they claim to be.
The deception economy is massive, with INTERPOL putting global financial fraud losses at $442 billion in 2025 (citing GASA) and rating AI-enhanced fraud as 4.5 times more profitable than older methods, driven by cheap AI technology.
The techniques behind that fraud—cloning voices, fabricating identities, grooming patiently—are the stack a foreign service might use to run a false HUMINT source.
So the advantage in intelligence work moves from holding more data toward judging which sources deserve belief—the heart of intelligence gathering.
In other words, human intelligence grows more valuable as automated collection and other technological advancements grow cheaper.
Three forces poisoning HUMINT today
Start with synthetic personas. Researchers tracking cyber units link Russia’s Unit 74455, known as Sandworm, to AI-generated personas built to win the trust of analysts, journalists, and civil society figures, with realistic photos and natural-sounding language.
None of this is new, of course—Cold War services ran false identities for decades. What’s changed is the cost and scale.
At Infosecurity Europe 2026 in London, former CIA chief of disguise Jonna Mendez framed deepfakes and voice cloning as an old craft aimed at the assumption of trust.
Then there’s automated deception. INTERPOL describes agentic AI systems that run whole fraud campaigns on their own, and has tracked such cyber operations spreading from Southeast Asia into the Middle East and West Africa.
UK Finance’s Annual Fraud Report calculated a 19% rise in authorised push payment fraud to £576.4 million in 2025, calling fraud a national security threat.
The same machinery serves threat actors and attackers running influence operations, source recruitment, and impersonation attacks.
Deepfakes push the problem further. A peer-reviewed meta-analysis of 56 studies covering 86,155 participants found that humans identify high-quality deepfakes at roughly chance level (55.54%)—accuracy little better than a coin toss.
The damage already reaches elections. In Ireland’s 2025 presidential race, a fabricated video showed the eventual winner withdrawing days before the vote, and the Netherlands saw roughly 400 synthetic images aimed at political figures.
If voters can’t authenticate a candidate, a HUMINT operative may not be able to easily authenticate a source.
How the rest of the world is responding
China appears to treat the AI HUMINT question as an integration target, not a dilemma. Georgetown’s analysis of Beijing’s military-civil fusion program describes civilian AI work flowing automatically into military and security use.
In October 2024, a Chinese ordnance research academy filed a patent proposing to fuse OSINT, HUMINT, signals intelligence, GEOINT, and TECHINT into one training set for an AI model meant to support every phase of the intelligence cycle.
The UK and the EU face an oversight puzzle. European rules on AI (such as the EU AI Act) demand transparency and accountability that pull against the secrecy that covert action and HUMINT operations require.
Intelligence services operate under national security carve-outs, yet the question of how democracies supervise AI-assisted intelligence operations remains open. Besides armies and military personnel, the persona campaigns we’ve outlined above target European civil society too.
South Asia shows the cost of moving late, with analysts describing how the human-intelligence-centred models that sustained Pakistan’s Inter-Services Intelligence (ISI) and India’s Research and Analysis Wing (RAW) now strain against AI, big-data surveillance, and automated analysis.
Without fusion between human source networks and AI tools, agencies may misread fast-moving crises—potentially dangerous in a nuclearised neighbourhood where one misread standoff can escalate. The shortfall there is integration, not effort.
Africa faces pressure from both sides. Extremist groups such as ISWAP already use AI for content production and communications, while the African Union’s continental AI strategy carries no binding obligation on member states and offers little guidance on AI in counterterrorism.
Meanwhile, African AI capacity concentrates in a few economies: 83% of the region’s AI startup funding in early 2025 went to four countries (Kenya, Nigeria, South Africa, Egypt).
This leaves most national security organizations on the continent dependent on human intelligence while adversaries pick up cheaper AI tools.
| Intelligence cycle phase | What AI changes |
| Direction | Machine-built profiles of threats can misdirect collection toward the wrong targets |
| Collection | Synthetic sources and fake informants contaminate the human reporting stream |
| Processing | Models surface patterns that reflect training bias as much as ground truth |
| Analysis | A confident AI summary can hide a rotten chain of evidence beneath it |
| Dissemination | Deepfaked channels carry disinformation into trusted briefings |
| Feedback | AI-equipped adversaries learn from how their deception performed and adapt |
What rises in value: human judgment
AI also strengthens the collector’s hand. A CSIS task force describes AI algorithms that help HUMINT officers spot and assess potential agents by combing open data, then build a digital pattern of life from a target’s personal information.
But that capability cuts both ways. If the system maps a target who is themselves a synthetic persona, it merely produces a careful map of fiction.
The CIA’s OSIRIS platform hints at a workable version: a model that digests vast amounts of open-source material and hands summaries to a human analyst. The model compresses, but the human analyst still judges.