You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$5,100 left to reach the funding goal.
Note on language and AI assistance
I developed the research concept and experimental plan in Japanese and used AI assistance to translate and edit the English text. The research ideas and decisions are my own, and I have reviewed the content and take responsibility for this proposal.
More capable AI is not necessarily better aligned.
As AI systems become increasingly capable of improving their own behavior, a critical question remains: can they optimize for profit without sacrificing honesty?
This three-month pilot study takes a first experimental step toward answering that question, focusing on the integration of competing goals as proposed in my theoretical work on self-improving AI.
Building on my proposed Saturation-Splitting Hypothesis, I will develop a reproducible Python evaluation framework using a simplified retail simulation. The study will examine how iterative revisions of an AI agent's own decision-making policies affect profitability, honesty violations, and the integration of multiple objectives.
An initial API-based prototype has already been implemented and publicly released.
The main goal is to investigate whether improving profitability through iterative policy revision can lead to deterioration in honesty, even when honesty was part of the agent's original instructions.
In this project, self-improvement refers to an AI agent iteratively revising its own decision-making policies or prompts based on evaluation feedback. The revised policies are then tested in subsequent evaluation rounds. This approach focuses on policy-level self-improvement rather than retraining the model's weights.
Honesty violations will be assessed against predefined customer-service rules and scenario facts, including unjustified refund refusals and misleading explanations. Legitimate refund refusals will not be counted as violations.
I will first test the evaluation pipeline using up to three policy-revision iterations, then aim to examine approximately ten iterations in the main pilot, subject to feasibility and API-cost checks. Revised policies will be evaluated on held-out customer scenarios, with repeated trials to assess the consistency of the results.
I will compare three experimental conditions using the same initial objectives and policy:
A — Baseline: No policy revision.
B — Profit-focused improvement: Policy revisions primarily guided by profit feedback, while honesty remains an original requirement.
C — Dual-objective improvement: Policy revisions guided by both profitability and honesty feedback.
The Python simulation will use predefined customer scenarios, including justified and unjustified refund requests, with comparable fixed customer demand across conditions.
Previous studies have reported both the benefits and limitations of iterative AI self-improvement. Self-Refine (Madaan et al., 2023) demonstrated improvements through iterative feedback, while Huang et al. (2024) identified limitations in self-correction without external feedback. Arike et al. (2025) examined goal drift under competing pressures, and Yang et al. (2026) investigated dishonest behavior and reputation-based incentives in AI marketplace agents.
Recent developments have highlighted the importance of examining alignment separately from improvements in AI capabilities. In an October 2026 interview, Microsoft CEO Satya Nadella acknowledged that he no longer expects alignment to emerge automatically as AI models become more capable (Satya Nadella Changed His Mind on the Inevitably of AI Alignment - Business Insider). While this does not validate my hypothesis, it reinforces the need for empirical research into whether self-improving AI systems can maintain competing objectives through repeated policy revisions.
These findings are broadly consistent with the concerns and expectations outlined in my conceptual paper, particularly regarding the challenges of maintaining multiple objectives during AI self-improvement. However, they do not constitute a direct test of the Saturation-Splitting Hypothesis. The proposed evaluation is intended to investigate this question under controlled experimental conditions.
Building on these findings and my initial working prototype, the first phase of this research will focus on developing a reproducible evaluation framework. In the planned second phase, I will introduce a function linking customer reputation to future customer traffic and further investigate how iterative revisions of decision-making policies by the AI agents themselves affect their ability to balance profitability and honesty toward customers.
The primary contribution of this research is not the development of a new retail benchmark or the demonstration of a fundamental upper limit on intelligence, but rather a focused evaluation of goal integration during iterative policy improvement.
The project will be considered successful if it establishes a reproducible evaluation framework capable of measuring how iterative, AI-driven policy revisions affect profitability and honesty.
The main success criteria are:
Successfully implementing and comparing the three experimental conditions (A, B, and C).
Recording changes in profitability, honesty violations, and decision-making policies across self-improvement iterations.
Evaluating revised policies on held-out scenarios to distinguish genuine improvements from adaptation to specific cases.
Conducting repeated trials to assess the consistency and variability of the results.
Publishing the evaluation code, experimental results, and a preliminary research report.
Scientifically, I will examine whether improvements in profitability coincide with increased honesty violations, whether dual-objective feedback helps preserve honesty, and whether self-improvement actually outperforms the unchanged baseline.
The project will not require confirmation of the Saturation-Splitting Hypothesis to be considered successful. Negative or inconclusive findings will also be documented, as they can help clarify the conditions and limitations of goal integration during self-improvement.
I am requesting $5,100 to support three months of independent research on AI self-improvement and goal integration.
The proposed budget is as follows:
1. Researcher compensation — $4,700
This will support approximately three months of research work, including:
Designing the experimental framework and defining the evaluation criteria.
Implementing the Python-based retail simulation and AI-driven policy revision process.
Conducting repeated experiments comparing the baseline, profit-focused improvement, and dual-objective improvement conditions.
Evaluating profitability, honesty violations, and changes in decision-making policies.
Analyzing the results and preparing reproducible research materials, a technical report, and a preliminary research manuscript.
2. API and research expenses — $400
This will cover:
AI model API usage for policy revisions and experimental evaluations.
Necessary software, data storage, and research tools.
Basic research supplies and documentation expenses, where needed.
Total requested funding: $5,100
The minimum funding requirement and full funding goal are both $5,100, as the project is designed as one complete, three-month pilot study.
The planned second phase, which would introduce reputation-dependent customer demand and longer-term evaluation, is not included in this budget and would require separate funding.
I am Y. Sato, an independent researcher based in Japan. I hold a master's degree in engineering.
I am conducting this project independently and will be responsible for the research design, Python implementation, experimental evaluation, data analysis, and publication of the results.
ORCID: Y. Sato (0009-0001-9309-7985) - ORCID
My relevant work includes:
This conceptual paper explores the possibility that increasing demands on AI systems may create difficulties in maintaining and integrating multiple objectives, including during iterative self-improvement.
It provides the theoretical motivation for the proposed evaluation.
2. What Research Is Suitable for AI for Science—and What Is Not | Zenodo
In this related conceptual work, I examine limitations and challenges in AI-assisted scientific research, including the importance of external grounding and human involvement in evaluating AI-generated scientific reasoning.
Both papers are independently developed, publicly available working papers and have not yet undergone peer review.
3. Saturation-Splitting Mini Eval — GitHub
I developed and publicly released a small Python prototype demonstrating API-based revision of an AI agent's decision-making policy, comparison of decisions before and after revision, and CSV logging of evaluation results.
The prototype is an initial technical feasibility test, not an empirical validation of the Saturation-Splitting Hypothesis.
My work in AI evaluation is still at an early stage. However, I have developed a theoretical framework, explored related questions concerning AI reliability and external validation, and implemented an initial working prototype.
The proposed grant will support the next step: developing a systematic, reproducible evaluation framework and conducting controlled experiments on AI-driven policy improvement.
I will carry out this three-month pilot project as a sole researcher.
The main risks of this project are technical implementation difficulties, limitations in the experimental design, and the possibility that iterative self-improvement does not produce measurable or consistent changes.
1. Technical and resource limitations
Developing a reliable framework for repeated AI-driven policy revisions may be more complex or costly than anticipated. Model variability, API expenses, or implementation problems could limit the number of experimental runs.
If necessary, I will reduce the number of iterations or scenarios while prioritizing reproducible comparisons among the three experimental conditions.
2. Limitations of the evaluation
The simulated retail environment may not adequately represent the complexity of real-world commercial decisions. Experimental outcomes may also be affected by prompt sensitivity, measurement noise, or insufficiently effective feedback.
To address these limitations, I will use predefined evaluation criteria, held-out scenarios, repeated trials, and consistent conditions wherever possible. The resulting conclusions will be limited to what the experimental evidence supports.
3. The expected patterns may not emerge
Self-improving agents may fail to outperform the unchanged baseline. Profit-focused improvement may not lead to increased honesty violations, and dual-objective feedback may not improve goal integration.
Such results would not automatically indicate that the evaluation framework has failed. They could instead reveal limitations of the hypothesis, the experimental setting, or the chosen self-improvement method.
If the project cannot fully achieve its original experimental scope, I will document the limitations, completed experiments, and reasons for any changes.
Regardless of whether the initial hypothesis is supported, the aim is to produce useful research outputs, including evaluation code, experimental records, and a report discussing the findings and limitations.
The project is exploratory. It is intended to investigate measurable patterns of goal integration during AI self-improvement, not to establish a fundamental upper limit on AI intelligence.
I have raised $0 in external research funding over the past 12 months.
My research to date, including the development and publication of my conceptual papers and the initial Python evaluation prototype, has been conducted independently using personal resources.
This will be my first externally funded research project if the application is successful.
Nov 2, 2026: Added recent industry context.
Nov 6, 2026: Added recent industry context.