AI changes collection and reasoning
The essays describe filtering, translation, synthesis, hypothesis generation, deepfakes, source validation and influence over how a problem is framed.
INTELLIGENCE · AI · PROVENANCE
Who thought first?
The two 2026 issues of the CIA journal Studies in Intelligence place AI earlier in the analytic process, before a verdict is reached. It can frame the problem, multiply plausible falsehoods, order hypotheses and present a conclusion in an authoritative voice. Provenance, human access and institutions able to reconstruct how judgment was formed become more important.
The texts are unclassified extracts and express the authors’ views, not the official position of the CIA or the US government.
2public issues in 2026
5proposed governance controls
10×synthetic calls in Mulligan’s scenario
3pillars of identity intelligence
SOURCECIA Center for the Study of Intelligence
PERIODMarch and June 2026
CORPUS5 central essays and 1 historical study
LIMITPublic extracts, not a full operational record
TERMS
MAP OF THE ARGUMENT
What is fact, what is interpretation, and what remains open
Trust becomes the scarce resource
Taken together, the essays point to the same pressure: value concentrates in provenance, human access, dissent and an institution’s ability to audit how a conclusion was formed.
How much has real practice changed?
The public extracts do not measure operational adoption, model performance or the scale of change in classified work.
CHAIN OF CUSTODY
Where AI enters determines the trace it must leave.
Select a stage. The provenance receipt shows the stage-specific risk and the evidence needed to reconstruct the final judgment.
AI-05 / FRAMING
FACT + INTERPRETATION
AI can narrow the question space before the conclusion.
GenAI can influence how a problem is framed and how options are constructed. The effect repeats across analytic cycles, allowing an initial selection to become the architecture of the entire judgment.
RISK
Premature convergence, omitted hypotheses and authority borrowed from fluent language.
REQUIRED TRACES
- original question
- alternative hypotheses
- prompts and versions
CONTROL
Dissent, devil’s advocacy, alternative analysis and a human justification of the final frame.
01 · Eveniment
Urma începe cu timpul, locul, înregistrarea brută și limitele observației.
02 · Colectare
Filtrarea cere inventarul surselor, interogarea, modelul și elementele respinse.
03 · Identitate
Validarea cere atribute, coroborare și lanțul dintre persoană, document și mesaj.
04 · Sinteză
Rezumatul cere maparea fiecărei afirmații la sursa care o susține.
05 · Încadrare
Formularea problemei cere ipoteze alternative, dezacord argumentat și prompturile folosite.
06 · Judecată
Verdictul cere un responsabil uman, incertitudine explicită și revizuire.
THE 10X SCENARIO
Synthetic noise scales faster than verification.
Mulligan offers a scenario: for every X real calls of intelligence value, an adversary generates 10X deepfake calls. Move the multiplier, then activate human corroboration.
10real calls in the model
100synthetic calls
9.1%chance that a randomly selected call is real
?signals confirmed by a human source
The scenario is hypothetical. The percentage is our calculation: 10 / (10 + synthetic calls). Perfect revelation after corroboration is an explanatory extreme, not a performance estimate.
CONCEPTUAL MODEL
As technical capability diffuses, human access can become the marginal advantage.
The curves translate Mulligan’s argument without observed values. The relative advantage of technical collection falls as tools become common. HUMINT retains value where humans still hold relevant access and can validate the signal.
THE 1960–2026 TEST
The problems from 1960 remain. AI adds influence over thought.
Malphurs sets Joseph Becker’s computer essay beside current language about AI. Switch the year to see what survives, then what crosses the boundary of the older machine.
INPUT · COMPUTER · 1960
“The machine does not ‘think.’”
Joseph Becker, “The Computer”
What survives the change of term
Training peopleRECURRING
Data conversion and qualityRECURRING
Workflow redesignRECURRING
Checking the outputRECURRING
Provenance without a clear human authorBEYOND 1960
Influence over framing and hypothesesBEYOND 1960
Becker warned against confusing mechanical assistance with judgment. The institutional problems were already visible: training, infrastructure, workflow and verification.
THE IDENTITY GATE
Identity information becomes usable through three conditions that must work together.
Ballard identifies three pillars. Disable one. The gate closes because identity information depends on attributes, retrieval and the integrity of the administrative state.
I²FUNCTIONAL
All three pillars are active. In Ballard’s definition, identity intelligence cannot function when any one of them is missing.
THE MORAL LEDGER
Every source leaves a moral debt.
The study of Ernst Rudolf Fischer extends the provenance problem. Useful information retains the history of the person who supplied it, the benefits that followed and the public purpose that can be demonstrated.
Recorded in the file
DOCUMENTEDA senior role in the Nazi war economy
Fischer worked in the regime’s economic and planning structures and in the IG Farben network.
DOCUMENTEDUseful technical information after the war
He supplied assessments of attacks on the synthetic fuel industry and vulnerabilities in German production.
DOCUMENTEDSigns of a connection to Allied intelligence
The article assembles circumstantial evidence but cannot establish the exact service, status or terms of the relationship.
CIRCUMSTANTIALStill open
UNRESOLVEDSource or asset?
The exact relationship, its handler and the timing of recruitment remain unclear.
QUESTIONDid his usefulness buy protection?
The documents presented do not resolve why Fischer was not prosecuted.
QUESTIONWhat result can be measured?
The record does not establish whether the information shortened the war or demonstrably reduced losses.
QUESTIONProvenance also has a moral dimension: who produced the information, what protection followed and what public purpose can be demonstrated.
PUBLIC TOOL · SIX TRACES
Judgment must leave traces.
Check the questions an AI-assisted analysis can answer. A fluent verdict remains opaque until its source, transformations and responsibility can be reconstructed.
PRIMARY SOURCES AND LIMITS
Every important claim returns to the text that supports it.
The links open PDFs published by the CIA Center for the Study of Intelligence. Page references use the printed journal pagination.
-
01
Thomas Mulligan, “Espionage in Our AI Future”
OPEN PDFSUPPORTSThe marginal value of HUMINT, synthetic noise, the 10X deepfake-call scenario and the role of secure human communication.
DOES NOT ESTABLISHObserved performance or details of current operations. The author explicitly frames the argument as speculative.
-
02
Debora Pfaff, “The Quiet Cognitive Coup of Generative AI”
OPEN PDFSUPPORTSAI as a cognitive technology, influence over framing, and five controls: epistemic transparency, human tradecraft, plurality, literacy and audit.
DOES NOT ESTABLISHAn empirical assessment of real adoption across the Intelligence Community.
-
03
Kenneth Malphurs, “We Have Been Here Before”
OPEN PDFSUPPORTSContinuity in institutional problems: training, infrastructure, data, workflow, verification and control of enthusiasm.
DOES NOT ESTABLISHA controlled productivity study. The essay includes the author’s experience using AI.
-
04
Lara Ballard, “Three Pillars of Identity Intelligence”
OPEN PDFSUPPORTSThe three pillars: reliable immutable attributes, sophisticated retrieval and a professional, incorruptible administrative state.
DOES NOT ESTABLISHAn AI governance standard. The extension to assisted analysis is editorial interpretation.
-
05
Sean Barnes, review of “Language Machines”
OPEN PDFSUPPORTSThe risk of fluent prose without explainable provenance, opacity of reasoning and the usefulness of AI for red-teaming and cultural analysis.
DOES NOT ESTABLISHAn institutional position or policy consensus. This is a review essay.
-
06
Anand Toprani, “The Enigma of Ernst Rudolf Fischer”
OPEN PDFSUPPORTSThe moral ambiguity of using compromised sources and the unresolved questions around Fischer’s relationship with Allied intelligence.
DOES NOT ESTABLISHDefinitive proof of recruitment, the responsible service or a measurable contribution to shortening the war.
The journal articles express the authors’ views. The public issues are unclassified extracts and may omit classified material. The corpus supports a synthesis of public arguments, without demonstrating the CIA’s official position, classified practice or the real degree of operational adoption. Accessed July 13, 2026.