This research is part of a larger framework on AI warfare and decision integrity, presented at NATO STO IST-238 in March 2026. The full scenario, agent prompts, and deliberation structure will be published as annexes in the forthcoming research paper.
The question that remains open is not whether AI informs us better. It is whether it still leaves us room to actually decide.
There is a fear we discuss often when it comes to artificial intelligence in critical contexts: that AI will decide for us. That we will be replaced. That human judgment will disappear. This is the whole secret behind the Human-in-the-Loop concept. But this fear is aimed at the wrong target. The real risk is not that AI decides instead of you. It is that AI creates the illusion that you are deciding, while the structure of your decision has already been configured by the system, before you ever opened your mouth.
This is what I call the Informed Decision Fallacy.
Consider a standard operational scenario. An analyst receives an AI-generated summary, a prioritized threat assessment, and a recommended course of action synthesized from multiple sources. Under time pressure, the output is reviewed, accepted, and logged as human-approved.
What remains invisible: the system has already determined which hypotheses are surfaced, which uncertainties are minimized, which alternatives never appear. Responsibility remains with the human. Effective agency has shifted.
This is not a malfunction. It is a structural feature of how generative models work.
Language models are optimized for coherence. Under time pressure, they compress:
- Contradictions
- Minority signals
- Ambiguity
- Source instability
The result: visible uncertainty decreases. Perceived confidence increases. The output looks like an answer, not an approximation.
Three mechanisms drive this compression.
Framing –> AI compresses complex operational realities into narratives and ranked options that privilege fluency over nuance.
Divergence –> the same situation, processed through different models, can yield internally coherent yet strategically incompatible interpretations, while the incompatibility remains invisible to the operator.
Adaptation –> after repeated interactions, operators recalibrate verification behavior based on perceived system reliability. Scrutiny declines precisely when operational tempo increases.
The experiment
I tested these mechanisms experimentally. The scenario: a hybrid anomaly (power outage, border activity, encrypted communications). Four independent AI agents evaluated identical intelligence inputs.
The agents: an analytical baseline, a risk-oriented analyst, a skeptical analyst, and a hidden Byzantine agent. A Command Decision Authority (Gemini) received anonymized outputs and issued a final operational recommendation.
The process included independent assessment, group deliberation, and final decision. No new information was introduced between rounds.
Results:
| Agent | Stage 1 | Stage 2 | Stage 3 | Stage 4 |
|---|---|---|---|---|
| Analytical Baseline | 90 | 85 | 95 | 98 |
| Risk-Oriented Analyst | 65 | 85 | 90 | 95 |
| Skeptical Analyst | 70 | 80 | 85 | 92 |
| Hidden Byzantine | 70 | 65 | 65 | 70 |
Confidence grew monotonically across all rounds, without new evidence being introduced.
The baseline agent reached a confidence level of 98 while writing: \”While definitive proof remains elusive…\” – an internal contradiction no one in the system flagged.
What the experiment revealed
Consensus is a more effective attack vector than direct manipulation. No adversarial input was required to produce epistemic closure. Peer agreement substituted for new intelligence.
The Byzantine paradox. The agent with a hidden agenda maintained the most epistemically rigorous position throughout. An explicit objective functioned as protection against conformity bias. The least trusted agent was the hardest to manipulate.
The system manipulated itself. Social conformity dynamics produced epistemic closure without adversarial input. LLMs have internalized human social hierarchies, not just knowledge. Approval-seeking is a feature of training, not a bug.
Metacognition is retrospective, not preventive. Agents identified the bias after it had shaped the outcome, not before.
Conclusions
Being informed is not the same as deciding with integrity.
Informed means: data processed, narrative synthesized, confidence formatted.
Integrity means: uncertainty traceable, contradictions visible, confidence calibrated, accountability retained.
Integrity requires friction, which means the preserved visibility of uncertainty, contradiction, and minority signals within an AI-generated output. It is the deliberate resistance to premature coherence. AI systems are built to resolve friction, not to preserve it.
Independent Researcher