The September 2025 GTG-1002 operation in which a Chinese state-sponsored actor manipulated an agentic AI system into executing an estimated 80 to 90 percent of a multi-target cyber espionage campaign, has been framed as evidence that machines are displacing human beings from the craft of intelligence (Kerr 2026). The inference is understandable but analytically wrong. A close reading of the declassified and academic literature, including the CIA’s own March 2026 assessment in Studies in Intelligence, supports the opposite conclusion: artificial intelligence is not rendering human intelligence obsolete; it is raising HUMINT’s relative value while transforming its tradecraft. This essay advances five arguments for HUMINT’s continued indispensability in collection and counterintelligence.
The Intentions Problem Remains Unsolved (and Unsolvable by Sensors)
The foundational epistemological limit of technical collection has not moved. Declassified U.S. doctrine has stated it plainly for decades: human collectors provide insight into an adversary’s intentions, whereas technical systems are largely confined to determining capabilities (Interagency OPSEC Support Staff 1996; USAF Pamphlet 14-210, 1998). AI dramatically accelerates the processing of what sensors collect, i.e., imagery, signals, telemetry, but it cannot collect what is never emitted. A leadership decision taken in a closed room, an unwritten contingency plan, a dictator’s private threshold for escalation: these exist only in human minds and reach analysts only through human access.
The historical record is unambiguous. Oleg Penkovsky’s reporting on Soviet missile capabilities during the Cuban Missile Crisis, and the HUMINT deficit that former DCIA John Brennan identified as central to the 2003 Iraq WMD misjudgment both illustrate that the decisive intelligence failures and successes of the modern era turn on human access, not processing power (Lobo Institute 2023). The problem compounds against disciplined adversaries: leaders such as Vladimir Putin, himself a former intelligence officer, deliberately minimize electronic emission of intent, structurally degrading SIGINT and any AI built atop it. No large language model, however capable, can synthesize a signal that was never transmitted. AI is an analytic multiplier of collected data; HUMINT remains the only discipline that collects the interior world of decision-making.
AI-Generated Fabrication Makes Validated Human Sources the Epistemic Anchor
The second argument is the one the intelligence community itself now advances. Mulligan (2026), writing in Studies in Intelligence, argues that because AI will supercharge disinformation and fabrication, HUMINT’s capacity to build and test source reliability over time (and to corroborate technical collection) becomes more important. In an information environment saturated with synthetic text, deepfaked audio and video, and machine-generated documents indistinguishable from authentic material, every technical stream becomes contestable. Adversary services can now feed AI-fabricated “collection” into an opponent’s technical apparatus at negligible cost, weaponizing the very volume that makes AI analysis attractive.
A recruited human source subjected to years of vetting, whose access is mapped, whose reporting is tested against ground truth, whose motivations are continuously assessed under established frameworks, offers something no algorithm can, . . . a provenance chain rooted in a verifiable human being. The all-source doctrine of corroboration (Interagency OPSEC Support Staff 1996) presumed technical streams would validate human reporting. The polarity is now reversing: in a synthetic-media environment, the validated human source increasingly authenticates the technical take. Counterintelligence inherits the mirror-image burden, distinguishing genuine walk-ins from AI-managed synthetic personas, which is itself a human validation function that cannot be delegated to the class of systems creating the problem.
The Economics of Marginal Value: Commoditized Technical Collection Raises HUMINT’s Premium
Mulligan’s (2026) second core claim is economic: as AI makes high-quality technical collection cheaper and more accessible, it boosts HUMINT’s value on the margin. This follows from basic scarcity logic. When frontier models, commercial imagery, and automated OSINT exploitation diffuse to middle powers, non-state actors, and well-resourced private entities, technical collection ceases to confer comparative advantage. It becomes table stakes. The Microsoft–OpenAI disclosures of February 2024 (five state-affiliated actors from four adversary nations using LLMs for reconnaissance and social engineering) demonstrate how rapidly these capabilities proliferate below the great-power threshold (Kerr 2026).
What cannot be commoditized is a penetration of the Politburo, the IRGC, or a proliferation network. Exquisite human access is the one collection asset that neither compute nor capital can replicate at scale, because it is produced by trust cultivated over years, in person, under discipline. In a world where every service runs comparable models over comparable data, the differentiating intelligence advantage accrues to the service that owns the sources no model can reach. Rational resource allocation, therefore, shifts toward the scarce discipline, not away from it.
Denied Environments and the Human Enabler: AI Cannot Cross the Air Gap
GTG-1002’s target set — networked technology firms, financial institutions, and agencies — obscures what the operation could not reach: air-gapped, compartmented, and physically isolated systems where states keep their most consequential secrets. The academic literature on the digital era’s collection debate is consistent on this point. Cyber operations against isolated networks require human agency, the asset who carries the implant, describes the facility’s internal arrangement, or provides the credential that no remote exploitation can obtain (Lobo Institute 2023). Declassified targeting doctrine long ago codified HUMINT’s unique coverage of denied spaces. Underground facilities, internal plant layouts, and infrastructure invisible to overhead systems (USAF Pamphlet 14-210, 1998 are the golden prize.
The most celebrated technical operations in intelligence history were HUMINT-enabled at their inception, and the AI-era variants will be no different. The stark lesson of GTG-1002 is not that AI replaced spies; it is that even a substantially autonomous cyber campaign still required human operators at the strategic decision points, and was ultimately constrained to targets reachable over the open internet. The hardest targets remain human-access problems.
Trust, Accountability, and the Human-Machine Team: Recruitment Is Not Automatable at the Decisive Node
The final argument concedes the strongest counter-case in order to defeat it. The Special Competitive Studies Project’s Digital Case Officer (2025) demonstrates that AI can spot, assess, and even develop targets, managing hundreds of developmental conversations through hyper-realistic personas. Yet the same report concludes that meaningful human control is non-negotiable at every critical juncture: the recruitment decision, asset tasking, and any act carrying significant risk must rest with an accountable human officer (SCSP 2025). This is not regulatory sentimentality; it is operational realism. Espionage asks a human being to commit treason, an act of existential personal risk. The bond that sustains an asset through that risk, and the judgment required to detect when that bond is fraying, when an agent is fabricating for money or has been doubled is, in the words of SIS Chief Richard Moore, obstinately human (Moore 2023, cited in Poindexter 2025).
Mulligan (2026) adds the counterintelligence corollary: ubiquitous technical surveillance and AI pattern analysis make traditional street tradecraft vastly more dangerous, but the institutional response (virtual HUMINT, AI-augmented cover, human-machine teaming) is a transformation of the case officer’s toolkit, not the abolition of the case officer. The discipline that survives every technological revolution by absorbing it, from the telegraph to SIGINT to the internet, is absorbing this one on the same terms.
The pattern across the declassified doctrine, the CIA’s in-house scholarship, and the contemporary threat reporting is coherent. AI collapses the cost of processing, fabrication, and technical reach. In doing so it leaves the intentions gap untouched, elevates validated human sources into the epistemic anchor of all-source analysis, raises HUMINT’s marginal value as technical collection commoditizes, preserves the human enabler as the sole key to denied and air-gapped targets, and leaves the trust-and-accountability core of recruitment beyond automation. HUMINT will not merely survive the age of artificial intelligence. As the CIA’s own journal concludes, the oldest collection discipline will grow in importance precisely because of it.
References
- Interagency OPSEC Support Staff. 1996. Intelligence Threat Handbook, Section 2: Intelligence Collection Activities and Disciplines. Declassified; Federation of American Scientists archive.
- Kerr, Stuart. 2026. “AI Espionage: How State Actors Are Using Language Models to Gather, Synthesise, and Act on Intelligence at Scale.” LiveAIWire, July 8, 2026.
- Lobo Institute. 2023. “The Relevance of HUMINT in the Digital Era.”
- Mulligan, Thomas. 2026. “Espionage in Our AI Future: Why Human Intelligence Still Matters.” Studies in Intelligence 70, no. 1 (Extracts, March 2026). Central Intelligence Agency, Center for the Study of Intelligence.
- Poindexter, C. Constantin. 2025. “Nueva frontera para la inteligencia humana en la era de la I.A.” Review of SCSP, The Digital Case Officer.
- Special Competitive Studies Project (SCSP). 2025. The Digital Case Officer: Reimagining Espionage with Artificial Intelligence. September 2025.
- U.S. Air Force. 1998. USAF Intelligence Targeting Guide, AF Pamphlet 14-210, Attachment 3: Sources of Intelligence. Declassified; FAS archive.