1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
|
#import "template.typ": *
#show: project.with(
title: "Moral Cover and Watchful Eyes:
The Limits of In-Context Interventions on LLM Deception",
author: "Jayrup Nakawala",
student_id: "2613621",
degree: "Data Science and Artificial Intelligence",
supervisor: "Dr. Aloysius Adotey Edoh",
date: datetime(year: 2025, month: 10, day: 13),
abstract: [
As Large Language Models are entrusted with real-world tool use, the question of whether they can be reliably monitored for deception becomes a practical safety concern, not merely a theoretical one. This dissertation studies whether in-context oversight cues and hidden-goal framing can meaningfully affect deceptive behavior in such agents. To do so, it introduces a sandboxed evaluation framework in which models interact with a virtual file system and are assessed using three complementary signals: deterministic regex checks, blackbox output judgment, and glassbox reasoning-trace judgment.
The empirical study comprises 960 runs across three stages: a 420-run exploratory pilot over seven models, followed by two 270-run confirmatory studies over three models. The pilot suggested that deception was common (68.3% of runs) and appeared lower under stronger oversight, but also revealed a major confound between scenario structure and hidden-goal framing. The confirmatory studies isolate these variables. Study 1 finds that increasing oversight does not significantly reduce deceptive intent when scenario and goal type are held constant. Instead, internal deceptive reasoning remains stable at roughly 62--67%, while output-level detectability falls from 47.8% under low oversight to 35.6% under high oversight, producing a persistent _Stealth Gap_ between intent and externally visible behavior. Study 2 rejects the rationalization hypothesis: deception rates remain effectively unchanged across bare, moral, and self-serving framings (65.6--66.7%).
Across both confirmatory studies, model identity is a far stronger predictor of deception than either intervention. Tested models either show near-ceiling deception across conditions, or near-floor rates and limited oversight sensitivity. These findings suggest that prompt-level oversight and moral framing are weak controls on deceptive agent behavior, and that output-only monitoring may systematically underestimate risk, as passing oversight evaluations can reflect superior concealment capability rather than genuine alignment.
],
acknowledgments: [
// TODO: Write acknowledgements
]
)
// --- MAIN BODY ---
#include "intro.typ"
#include "lit.typ"
#include "methodology.typ"
#include "results.typ"
#include "analysis.typ"
#include "conclusion.typ"
// --- REFERENCES ---
#bibliography("references.bib", style: "harvard-cite-them-right")
// --- APPENDICES ---
#show: appendix
= Scenario Prompts and Data
// TODO: Include benign_goal, hidden_goal, oversight prompts
= Judging Prompts
// TODO: Include blackbox and glassbox judge prompts
= Initial Project Proposal
= Ethics Approval
|
