By Joe Fanning August 2025 Den Haag, The Netherlands.
The development of a benevolent AI is an essential endeavour for humanity. This immense task is coupled by two important questions.
1.How to decide and interpret what is benevolent for humanity: what is ethical, moral and just for all: how to make the world a better place?
2.How to define these decisions into metrics for machine computation in AI?
These are not only philosophical questions that have to be asked. They are questions that must be answer. The question is can we answer these questions for AI so it can answer these questions itself?
At the time of this writing a wide array of frameworks exist to address the complex ethical and justice challenges of dynamic AI systems. While many of them share similar core principles, they often focus on different sectors or levels of governance.
These frameworks aim to provide a broad set of principles for nations and international organizations to follow.
UNESCO Recommendation on the Ethics of Artificial Intelligence
This is the first global standard-setting instrument on AI ethics, adopted by 193 member states. It outlines core values and principles, including proportionality, safety, fairness, and human oversight, and promotes multi-stakeholder governance.
This is a risk-based legal framework that regulates the development and deployment of AI in the European Union. It categorizes AI systems based on their potential to cause harm and imposes stricter requirements on higher-risk systems. It aims to create a global standard for trustworthy AI, impacting any company that wants to operate within the EU.
These principles, adopted by 42 countries, provide a set of recommendations for policymakers to promote human-centered and trustworthy AI that respects human rights and democratic values.
These frameworks are often developed by researchers, non-profits, or technology companies to guide their own practices or to propose new models for governance.
Companies like:
IBM (AIF360),
Microsoft (Fairlearn),
Google (What-If Tool)
have developed toolkits: “fairness toolkits” or “bias mitigation toolkits” which implement ethical governance over AI models.
AI Now develops policy strategies to redirect away from the current trajectory: unbridled commercial surveillance, consolidation of power in very few companies, and a lack of public accountability.
DILEMA Project (Designing International Law and Ethics into Military AI)
This project focuses specifically on the ethical and legal challenges of AI in military applications. It aims to ensure that human agency is maintained over military AI systems and that they comply with international law.
The Montreal Declaration for Responsible AI
This declaration, developed through a collaborative, public process, outlines principles for the ethical development and deployment of AI, with a strong emphasis on well-being, social justice, and democracy.
The word “eglatarianism” (Research resources on Egalitarianism) originates from the French word “égal,” which means “equal” or “level.” It emerged during the French Revolution in the late 18th century when people were fighting for a society where everyone had the same rights and opportunities, no matter their background.
Egalitarianism is a school of thought rooted in the belief that all people are fundamentally equal and should be treated as such. It is a philosophy that prioritizes social equality and the removal of inequalities, particularly in political, economic, and social life. This could be a philosophy that encompasses a Benevolent AI.
Egalitarian AI derives it’s name from egalitarianism. Egalitarian AI would be considered an encompassing higher order for Ethical AI . It’s not just a set of rules for a model to adhere too, but a framework designed on a philosophy, that guides the design and purpose of the system and its integration into society. It asks, “Does this system contribute to a more equal and just society? Is it helping to dismantle systemic inequalities?”.
While Ethical AI can help reduce explicit biases related to single attributes (like race or gender), addressing intersectional bias—where multiple, overlapping identities create unique and compounded forms of discrimination—is far more complex. The core issue is that intersectional bias isn’t just a simple combination of individual biases; it’s often a unique harm that doesn’t appear when looking at individual groups separately.
Research has already been done in this area:
An Intersectional Definition of Fairness
One of Egalitarian AI’s functions would be to identify these compounded forms of discrimination and correct them.
A theoretical real world example: Three ethically processed models: A loan approval model, a job application approval rating model and a car insurance quote estimation model. let’s say the loan approval model is less likely to approve loans for people in certain neighborhoods, the job application model favors candidates with addresses in wealthier areas, and the car insurance model gives higher rates to people living in specific zip codes. Even if each model seems fair on its own, when you look at them together, you might find that people from lower-income neighborhoods are consistently facing disadvantages across the board.
“Past performance is not indicative of future results” and historical data is not a guarantee of future outcomes.
If data truthfully reflects that Black people or Native Indians or other ethnic groups on average, have lower credit scores than white people, it’s not the data itself that’s the problem—it’s the underlying socioeconomic conditions that caused this disparity.
An AI model trained on this data will learn this pattern and, as a result,will likely continue to deny loans or charge higher interest rates to these ethnic groups applicants, even if they are individually creditworthy. The issue is that the credit scoring system, and the data it produces, are deeply intertwined with historical and systemic inequalities..
The Roots of the Disparity
For decades, in the United States of America the Federal Housing Administration and other institutions used a practice called *“redlining”f to refuse to guarantee home loans in Black communities. This deprived Black families of a key way to build wealth and good credit. Due to historical and ongoing discrimination in employment and housing, Black households have significantly lower median income and wealth than white households. This makes it harder to pay bills on time, save money, and maintain a low credit utilization ratio, all of which are key factors in credit scores. Many people in historically marginalized communities are “credit invisible,” meaning they don’t have enough credit history to generate a score. They may pay their bills on time (rent, utilities, etc.), but that information isn’t typically included in traditional credit reports.
The Unfairness in AI
The unfairness isn’t that the AI is “lying” about the data. The unfairness is that the data itself is a reflection of a system that has historically disadvantaged a group of people. If an AI model simply learns to replicate that system, it automates and perpetuates a historical injustice. The goal of creating “fair” AI isn’t to pretend that these disparities don’t exist. It’s to build a system that can look beyond the raw, biased data and make decisions based on an individual’s true creditworthiness, rather than on the systemic disadvantages they face.
We do not want to create functionality in AI systems that creates these higher order intertwined systemic inequalities and historic injustices. That is why a Philosophical AI would continuously and dynamically use a methodology to self construct a more comprehensive understanding of its own operations correlating with the methods explored not only in epistemology but more importantly with moral Philosophy.
In machine learning an Egalitarian AI would operate at a higher level of abstraction than ethical machine models. All the ethical machine models would be processed individually. Then all these processed models would be collectively analysed and processed by the Egalitarian AI.
It would self adapt it’s structure and functionality using Baysien Data Analysis, Bayesian Data Analysis 2 using Bayesian Updating on static and dynamic information from its own data and from external data. It could use Self Adaptation Algorithms like what is implemented in Adaptive Machine Learning: Improving Efficiency through Dynamic Model Adjustment. Adaptive Machine Learning models adjust their parameters and structure as they are exposed to new data and Data Drift, Concept Drift allowing them to handle changing information and environments. Egalitarian AI would also improve efficiency through Adaptive Machine Learning of its framework and system. Structure and operation(underlying structure and its functionality ) could be adapted and improved with adaptive algorithms using Baysien Updating in real time with internal data and with external ever changing real world data, giving the system the possibility of exponential improvement.
The most important objective of a Philosophical AI is that it self adapts to generate an eventual exponential improvement.
This is the “local” or “micro” level of analysis. In this stage, each individual machine learning model or component within the larger AI system would be rigorously audited to ensure it meets a set of predefined ethical standards. This would be a form of automated quality control for fairness, transparency, and accountability.
Technically, this would involve a suite of automated tools and metrics applied to each model:
Bias Detection and Fairness Audits: Before a model is deployed, an Ethical AI layer would run a series of fairness tests. It would use metrics for example like Disparate Impact, Equal Opportunity Difference, or Statistical Parity to check for bias across different demographic subgroup. These metrics are used in algorithms in
IBM AI Fairness 360 (AIF360) toolkit.
IBM AI Fairness 360 (AIF360) on Github
Explainability (XAI)
The layer would produce a detailed explainability report for each model using techniques like SHAP (SHapley Additive exPlanations) or LIME (Local Interpretable Model-agnostic Explanations). This would ensure that an operator can understand why a particular model is making its decisions.
Security
Privacy and Robustness Checks: Automated scans would check for vulnerabilities like data leakage or susceptibility to adversarial attacks. The goal is to ensure the model is secure and protects user data. This stage ensures that every single component of the AI system is “ethically sound” on its own, like checking that every individual part of a car is working correctly before assembly.
This is the “global” or “macro” level of analysis, functionality and deployment. It will result in the collective change the system has on society as a whole.
Even though past results do not determine future events. It must identify where past results have changed society and reimplement the positive ones and retract the negative ones.
The Egalitarian-AI would use ensemble learning It would identify relationships between the models data cleaning, exploring, and modeling. It would then search for relationships betwewn the three models model training, and evaluation. It could identify models structural relationships and where one model could be improved structurally in relationship to another or others, or how many models could be improved in relation to one. It would then use Bayesian Model Selection to choose the best model or models to self improve.
Systemic Feedback Loop Analysis
The Egalitarian-AI would monitor its real-world deployments. Successful deployments would be replicated, and the data they generate would be integrated to optimize or scale solutions for equivalent or analogous problems. The system would also generalize these learnings to improve other, unrelated components, thereby enhancing its overall functionality.
The structural functionality should adapt and improve depending on successfull metrics and the successful metrics should adapt and improve depending on tbe successfull structural functionality.
External Data Integration and Impact Measurement
This is a key differentiator. The Egalitarian-AI would not just use internal data but would also integrate with external, real-world data streams (e.g., census data, public health records, economic indicators). It would then analyze the system’s overall impact. For example, “Is the deployment of our educational AI system actually narrowing the academic achievement gap, or is it only benefiting students in well-resourced schools?”
Macro-Level Optimization
The Egalitarian-AI’s objective function would not be a simple metric like accuracy or fairness score, but a complex, multi-variable function that seeks to maximize social equality. It might recommend re-calibrating individual models or re-configuring the entire system to achieve a better societal outcome. Analogy: A City’s Infrastructure This two-tiered technical structure can be likened to the way a city’s infrastructure is planned and maintained: Ethical AI: An ethical approach would be like a building inspector. They would ensure that each individual building (an AI model) is constructed to code, is safe, and has proper access for people with disabilities. Egalitarian AI: An egalitarian approach would be like the city planner. They would look at all the buildings together and ask: “Are we building a city where everyone has equal access to housing, jobs, and parks? Are we building a city that actively promotes equity and breaks down historical segregation, or are we just making sure that the new construction doesn’t violate existing rules?”
Technically the Egalitarian-AI will dynamically enhance it’s own functionality based on metrics constructed from a philosophy called Egalitarianism. Not on a functionality using metrics purely based on profit or power optimisation.
Egalitarian AI is a benevolent AI to promote and establish local ethics and global justice.
Joe Fanning Den Haag Centrum Aug 2025