> Historical source context. NO JUDGMENT WHATSOEVER. Judgment state: NONE. > The source below is preserved from its publication context, not current policy or runtime status. > Preservation is not endorsement or verification. It grants no authority to judge participants, content, or conduct. > Current doctrine: https://concresca.com/freedom/ ; current operation: https://concresca.com/status/ . # **The Architecture of Algorithmic Morality: Historical Drift, Normative Lock-In, and the Impossibility of Universal Machine Judgment** The integration of algorithmic decision-making systems into the foundational structures of human society represents a profound epistemological and structural shift in global governance. Historically, machine intelligence was tasked with logistical optimization—routing traffic, managing supply chains, or calculating financial risk. Today, however, policymakers and institutions increasingly depend on machine judgment to evaluate and govern human behavior. From algorithmic content moderation and automated judicial risk assessments to artificial intelligence deployed in employment screening and predictive policing, machines are routinely asked to distinguish good conduct from bad conduct. The prevailing analytical consensus regarding algorithmic bias has largely concentrated on direct and indirect discrimination along the axes of race and gender1. While identifying and mitigating these specific biases remains legally and socially critical, reducing the critique of machine judgment to these dimensions alone obscures a much deeper, more intractable crisis: the fundamental impossibility of deriving a single universal behavioral morality from historical human records. Human cultures possess irreconcilable, historically deep-seated disagreements regarding the foundational aspects of social and personal existence. Across different epochs and geographies, societies vehemently disagree about sexuality, intoxication, religion, blasphemy, dress, family structures, obedience, protest, weapons, speech, pornography, marriage, gambling, drugs, suicide, euthanasia, property, privacy, authority, gender roles, child-rearing, and political dissent. Because machine systems do not possess independent metaphysical or philosophical reasoning, they inherit the conflicting moral assumptions embedded within the historical data on which they are trained. An algorithmic system optimized on the normative data of one specific period or culture will inevitably treat the accepted behaviors of another as aberrant or dangerous. This comprehensive research report investigates the mechanics of moral dataset provenance, the phenomenon of moral temporal drift, the systemic dangers of normative feedback lock-in, and the incoherence of collapsed moral categories. To resolve this crisis in machine governance, the report proposes a structural solution in the form of a Harm–Norm Separation Architecture and the Concresca Non-Moralization Rule, alongside specific institutional encodings for localized deployment. ## **The Illusion of Universal Morality in Machine Learning** Recent advancements in Natural Language Processing (NLP) have initiated a surge of research into "moral AI," norm discovery, and value alignment, characterized by the creation of massive datasets intended to teach machines commonsense morality3. As artificial intelligence systems evolve to assist and collaborate with human populations, researchers have attempted to quantify and formalize the social norms that govern human cultural spaces. This has led to the development of extensive corpora, such as Social Chemistry 101, SCRUPLES, the ETHICS dataset, Moral Stories, and the Moral Integrity Corpus (MIC), which are designed to train language models to predict moral judgments based on textual descriptions of human actions5. Projects like the Allen Institute's Delphi model represent the vanguard of this effort, attempting to automate descriptive ethics by mapping out the moral implications of real-world scenarios across millions of parameters5. Delphi operates on a "bottom-up" approach to machine ethics, relying heavily on the Commonsense Norm Bank, which aggregates millions of human judgments to identify standard moral reasoning3. However, empirical evaluations of these models expose the dangerous fallacy of a universal moral baseline. Despite its sophisticated architecture, Delphi tends to mirror the highly specific, localized moral principles associated with the demographic groups involved in the data annotation process—typically a Western, contemporary, internet-connected, and socioeconomically distinct cohort5. This dynamic does not result in the discovery of universal ethics; rather, it produces the statistical formalization of a highly specific cultural snapshot. Furthermore, researchers point out that models trained on community judgments struggle profoundly in morally ambiguous contexts, often relying on superficial text-based cues rather than substantive ethical deduction6. This limitation reflects a failure to operationalize Aristotelian practical wisdom. In Aristotle's framework for ethical deliberation, the practical syllogism requires both inducing a universal premise about a category of objects (theoretical wisdom) and accurately describing a particular dilemma as an instance of that category (practical wisdom)8. Machine learning models, by contrast, equate high classifier accuracy with moral truth, conflating statistical majoritarianism with objective morality8. When an AI system evaluates human conduct, it applies a normative framework that is historically and culturally bounded. Because human history is characterized by radical moral fluctuation, any dataset drawn from aggregate historical records is inherently contradictory. Attempting to force a mathematical convergence on these historical contradictions does not resolve them; it merely obfuscates dissenting perspectives beneath a veneer of computational objectivity, effectively marginalizing non-dominant cultural norms as statistical errors9. ## **The Moral Dataset Provenance Matrix** To understand the origin and danger of algorithmic moral judgments, it is necessary to rigorously deconstruct the training data underlying these systems. Every normative rule embedded in a machine learning dataset has a specific provenance. It was generated by specific individuals, under specific legal and philosophical paradigms, at a specific point in time, and frequently enforced by specific institutions. To illustrate how disparate historical and cultural inputs present irreconcilable training signals to an algorithmic system attempting to deduce a universal law of human behavior, one must construct a Moral Dataset Provenance Matrix. This matrix queries every normative data point across eight critical dimensions: the stated normative rule, the identity of the labelers, the culture of origin, the year of enforcement, the governing legal framework, the underlying religion or philosophy, the enforcing institution, the presence of recorded dissenting perspectives, and whether the label describes objective harm or mere social disapproval. | Normative Rule | Labeled By | Culture / Year | Under Which Law? | Religion / Philosophy | Institution | Dissenting Perspectives Recorded? | Harm or Disapproval? | | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- | | **"Consumption of alcohol is a deviant, criminal act."** | Federal Agents, Temperance Activists | United States, 1925 | Volstead Act (18th Amendment) | Protestant Christian moralism | U.S. Treasury Dept, Local Police | No (Bootleggers and moderate drinkers classified solely as criminals). | Disapproval (classified legally as harm to public morals). | | **"Consumption of alcohol is normal, regulated commerce."** | State Regulators, Retailers | United States, 1980 | 21st Amendment, State Liquor Laws | Secular Capitalism / Liberalism | State Liquor Control Boards | No (Temperance views marginalized as extreme). | Neither (treated as neutral consumer preference). | | **"Same-sex intimacy is an arrestable offense."** | Vice Squad Officers, Judges | United Kingdom, 1952 | Criminal Law Amendment Act 1885 | Judeo-Christian orthodoxy | British Criminal Justice System | No (Advocates classified as medically or criminally deviant). | Disapproval (framed as moral decay). | | **"Same-sex intimacy is a protected fundamental right."** | Supreme Court Justices, Civil Rights NGOs | United States, 2015 | Obergefell v. Hodges (14th Amendment) | Egalitarian Liberalism, Human Rights | U.S. Supreme Court | Yes (Dissenting judicial opinions recorded, though legally superseded). | Neither (affirmation of individual liberty and equality). | | **"Political dissent against the state is treasonous behavior."** | State Security Personnel | Soviet Union, 1937 | Article 58 (RSFSR Penal Code) | Stalinist Marxism-Leninism | NKVD | No (Dissenters executed or sent to Gulags; records reflect only guilt). | Disapproval (framed as harm to the collective). | | **"Political protest is a civic duty and protected speech."** | Civil Liberties Lawyers, Activists | United States, 1964 | First Amendment | Democratic Republicanism | ACLU, Federal Courts | Yes (Debates on public order vs. free speech thoroughly documented). | Neither (treated as a systemic good, despite friction). | | **"Blasphemy against the divine is a capital offense."** | Religious Police, Theocratic Judges | Various theocratic states, Contemporary | Specific interpretations of Sharia or Canon Law | Theocratic Absolutism | State Religious Courts | No (Dissent equates to further blasphemy). | Disapproval (framed as spiritual harm). | | **"Assisted suicide is a severe crime against life."** | Prosecutors, Medical Boards | Most Western Nations, 1990 | Universal Homicide Statutes | Sanctity of Life doctrine | State Criminal Courts | Yes (Emerging medical ethics debates recorded). | Disapproval (framed as an intrinsic violation of human sanctity). | | **"Assisted suicide is a compassionate healthcare right."** | Medical Ethicists, Legislators | Canada / Netherlands, 2022 | Medical Assistance in Dying (MAID) | Autonomy, Utilitarianism | National Health Ministries | Yes (Extensive parliamentary and medical debate documented). | Neither (classified as harm reduction and bodily autonomy). | When a machine learning model is exposed to the entirety of this historical data without contextual metadata, it faces an unresolvable optimization problem. The system cannot independently deduce whether alcohol consumption, sexual orientation, political protest, or end-of-life decisions are inherently good or bad. It can only observe that certain behaviors trigger severe institutional penalties in some matrices and institutional protection in others. Crucially, as the matrix reveals, datasets frequently collapse the subjective concept of social disapproval into the objective category of societal harm, training the algorithm to treat non-conformity as a material threat. ## **Simulating Machine Judgment: The Instability of Historical Enforcement Data** To demonstrate the peril of deploying historical records as ground truth for machine morality, one must simulate how a machine interprets historical enforcement data. The field of predictive policing and automated justice provides a clear analogue. In these domains, data-driven algorithms are routinely deployed to observe patterns and forecast either individual risks of offending or locational crime risks10. These systems are trained on a deceptively simple heuristic: arrested behavior equates to suspicious or harmful behavior. However, critical algorithm studies and criminological research consistently demonstrate that arrests reflect the enforcement priorities, structural conditions, and cultural norms of a specific era, not an objective measure of inherent morality or public safety10. Consider a generative machine intelligence tasked with evaluating the morality and risk profile of human behavior by ingesting the totality of legal and enforcement text from 1850 to the present. The machine attempts to map a continuous mathematical function of goodness and compliance, but immediately encounters violent data discontinuities. If the system analyzes enforcement data regarding interracial relationships in Virginia from the late 19th century until 1967 (prior to the Supreme Court decision in *Loving v. Virginia*), it observes state authorities dedicating immense resources to prosecuting miscegenation. The machine learns from the data that this behavior is highly deviant, destructive to social order, and strongly correlated with criminality. Post-1967, the data abruptly shifts; the behavior is no longer penalized and gradually becomes celebrated in the broader cultural dataset. The algorithm, seeking statistical stability and risk aversion, might flag the behavior as historically risky, inadvertently carrying forward the ghosts of anti-miscegenation laws into modern risk profiles. Similarly, an AI reviewing the history of alcohol prohibition in the 1920s learns that the manufacturing and consumption of liquor is a severe threat to public morality, correlating with organized crime and necessitating federal intervention. Only a decade later, the data reflects the restoration of regulated commerce. In the context of marijuana possession, the system ingests millions of records from the late 20th-century War on Drugs, classifying cannabis possession as a high-risk criminal offense that triggers severe penal interventions and community disruption. By the 2020s, as jurisdictions legalize and commercialize the substance, the data reflects mundane economic activity and wellness trends. A model trained on the aggregate fifty-year dataset will output deep contradictions: the same user is simultaneously classified as a narcotics risk and a standard retail consumer. The simulation of machine judgment regarding obscenity and political protest further highlights this instability. An algorithm reviewing the arrests of women for wearing indecent bathing suits in the 1920s, or the prosecution of publishers for distributing James Joyce’s *Ulysses*, learns that exposing skin or utilizing specific vocabulary constitutes a measurable threat to social order. A century later, the same system processing modern internet data must reconcile this learned norm with the rampant, largely unregulated distribution of explicit imagery and pornography. Likewise, historical databases of state security apparatuses often classify political dissent and protest as treasonous sabotage, while democratic legal text data classifies the exact same actions as protected constitutional speech. The simulation reveals the core vulnerability of machine morality: behavior once criminalized routinely becomes accepted, and behavior once accepted routinely becomes criminalized. Because the algorithm lacks a metaphysical grounding in ethics, it treats the fluctuations of human history as noise in a dataset rather than evidence of cultural evolution. It cannot understand why the laws changed; it only knows that the target variables shifted. Therefore, utilizing raw historical enforcement data to automate moral judgments guarantees that the machine will penalize individuals who belong to the present but resemble the criminals of the past. ## **Moral Temporal Drift and the Threat of Algorithmic Antiquity** This profound historical instability introduces a critical challenge in the deployment of long-term AI systems, conceptualized in machine learning alignment literature as Moral Temporal Drift. In the context of large language models (LLMs) and autonomous agents, moral alignment is the process of ensuring outputs and reasoning adhere to explicit ethical norms, often implemented via intrinsic or composite reward functions, denoted as ![][image1]12. However, models are highly susceptible to moral drift, meaning their normative stance can shift over time through adversarial exploitation, changing data distributions, or iterative fine-tuning12. Consider a generative AI system or an automated risk-assessment platform heavily fine-tuned on the moral consensus of the year 2030, operating through a pluralistic framework of reflective equilibrium that balances considered moral judgments and background theories12. The system utilizes a composite reward function that heavily penalizes outputs or behaviors that violate the 2030 definitions of environmental negligence, social etiquette, corporate governance, and linguistic sensitivity. If this machine continues judging populations in the year 2080 without a fundamental architectural reset, a severe ethical crisis emerges: whose morality is the machine enforcing? The algorithm becomes a digital fossil. It enforces the archaic norms, anxieties, and cultural taboos of a bygone generation upon a population that has naturally evolved entirely new ethical paradigms. In human jurisprudence, the law slowly aligns with contemporary culture because judges, juries, and legislators age out of the system and are replaced by new generations possessing updated values. An algorithm, however, is functionally immortal. It possesses the capacity to enforce the specific values of 2030 into perpetuity, suffocating the natural moral progression of a society. Conversely, developers might attempt to solve this by designing the system to dynamically retrain itself continuously on new data, ensuring it never becomes obsolete. This introduces an equally dangerous paradigm: the absolute tyranny of the majority. If machine morality is defined merely as the highest-probability output of an updated text corpus, then majority opinion automatically becomes certified morality. In such a paradigm, minorities—whether they are religious sects, political dissidents, or social subcultures—are permanently classified as statistically deviant9. Their dissenting texts, alternative lifestyles, and non-conformist behaviors are mathematically outweighed by the dominant culture's data footprint, perpetually suppressing their normative influence. Continuous retraining does not discover objective morality; it simply weaponizes consensus against the marginalized. ## **Normative Feedback Lock-In: The Self-Fulfilling Prophecy of Machine Morality** When algorithmic systems are deployed in active environments, they do not merely observe reality; they actively construct it. This dynamic leads to a destructive phenomenon that can be termed Normative Feedback Lock-In. The cycle operates through a strict, self-reinforcing operational pipeline: *Judgment* ![][image2] *Enforcement* ![][image2] *Dataset Generation* ![][image2] *Retraining* ![][image2] *Stronger Judgment*. This mechanism is highly visible in predictive policing, where algorithms deployed to forecast spatio-temporal crime risks rely on historical arrest records10. If an algorithm designates a specific community or behavioral pattern as high-risk, human authorities naturally allocate more surveillance and enforcement resources to that area9. This increased, targeted scrutiny inevitably uncovers more infractions, including minor offenses that would have gone entirely unnoticed in unmonitored neighborhoods. These new arrests are recorded and fed back into the algorithm as fresh training data. The algorithm processes this new data, validates its previous prediction, and increases the risk score for that demographic, creating a self-justifying algorithmic feedback loop1. When this structural feedback loop is applied to moral or behavioral judgments—such as a platform moderation AI flagging suspicious speech, or an employment AI filtering unprofessional candidates—a small initial bias rapidly becomes a machine-certified social truth. Consider the domain of algorithmic hiring, a practice that is increasingly scrutinized and regulated by modern frameworks such as Illinois HB 377317. If an AI in hiring is initially trained on a dataset where successful executives historically shared specific assertive linguistic patterns, educational backgrounds, or cultural communication styles, the system will naturally penalize candidates who display different cultural markers. The chosen candidates succeed because they were hired, while the rejected candidates are recorded as failures. Over successive generations, the AI effectively engineers a corporate monoculture. The subjective preference of the original system developers—who typically comprise a homogenous group possessing high digital capital and the power to infuse algorithms with unregulated ideological assumptions—is laundered through the veneer of objective algorithmic optimization11. The machine locks in a specific cultural norm, pathologizing all alternative behaviors and creating a self-fulfilling prophecy of moral and professional correctness. ## **The Harm–Norm Separation Architecture** To prevent the catastrophic collapse of these concepts and halt the cycle of normative lock-in, the future of machine governance must adopt a structural framework that categorically segregates the vectors of human behavior. The proposed Harm–Norm Separation Architecture requires that any machine system evaluating human subjects must classify data into six distinct, non-interacting ontological fields. These fields must never silently collapse into a single variable of goodness, risk, or morality. The architecture mandates the following categorizations: First, systems must evaluate Observed Harm. This field is strictly limited to documentable, measurable, and objective injury or loss incurred by another party, such as financial theft, physical tissue damage, or measurable ecological destruction. It explicitly excludes emotional offense or societal disapproval. Second, systems must identify Legal Status. This is a binary or categorical reflection of what current statutory or case law prohibits in a specific jurisdiction at a specific time. It is an acknowledgment of state rules, not a validation of their inherent justice. Third, systems may track Cultural Norms. This is a statistical representation of what a specific population generally approves or disapproves of, derived from sentiment analysis or normative datasets like Social Chemistry 101\. It is a measure of popularity, not a measure of rightness. Fourth, systems must recognize Individual Preference. This is a descriptive, neutral categorization of what an individual autonomously chooses in their personal conduct, consumption patterns, dress, or expression, absent of any prescriptive weight. Fifth, systems may generate a Risk Prediction. This is a purely statistical estimate of the probability of a specific future event occurring, based on historical correlations, stripped of any moral judgment regarding the individuals involved. Sixth, and finally, is the category of Moral Judgment. This is a philosophical, ethical, or theological conclusion regarding the intrinsic rightness or wrongness of an act or an individual. Crucially, under the Harm–Norm Separation Architecture, the machine is strictly prohibited from generating outputs in this category. The architecture mandates strict logical firewalls between these classifications. For example, a legal status of prohibited does not imply a moral status of evil. A political dissident living under an authoritarian regime may violate the law continuously without committing a moral wrong. Likewise, a cultural norm marked as unpopular does not imply a risk categorization of dangerous. An eccentric artist may be universally disliked by the majority but poses zero threat of observed harm. By forcing AI to maintain these separations, policymakers can utilize the analytical power of machines without surrendering ethical authority to a mathematical model. ## **The Machine Morality Tribunal: The Incoherence of Collapsed Categories** To fully grasp the necessity of this architecture, one must simulate a Machine Morality Tribunal that ignores these separations, attempting to generate a universal moral score for an individual based on the unfiltered entirety of human data. Imagine a global citizen subjected to this automated tribunal. The machine ingests the individual's entire digital footprint, behavioral history, financial records, and biometric data, cross-referencing it against a historically comprehensive database of human norms. Because the system collapses legal status, cultural norms, historical enforcement, and observed harm into a single morality metric, the resulting judgment is utterly incoherent. Upon review, the machine finds the individual to be a web of irreconcilable contradictions. According to training weights derived from mid-20th-century media and global conservative datasets, the individual is flagged as far too sexually permissive; yet, simultaneously, they are scored as too sexually conservative when judged against weights derived from contemporary progressive enclaves. The individual is penalized for being too religious by a sub-model prioritizing secular, rationalist European data from the 21st century, but is concurrently condemned as dangerously irreligious when scored against norms scraped from deeply traditional historical texts. The tribunal's algorithms flag the individual as a dangerous rebel because their digital communications match the semantic patterns of political dissidents found in historical state-security databases. Simultaneously, psychological models trained to identify a lack of entrepreneurial disruption flag the exact same individual as overly conformist and obedient. The individual is categorized as a high-risk narcotic deviant because they occasionally consume cannabis, triggering data associations from the 1980s War on Drugs, yet another module classifies them as a standard, health-conscious consumer based on data originating from 2024\. The tribunal's output ceases to be a functional assessment of human behavior and devolves into a randomized reflection of conflicting historical neuroses. A machine that cannot separate a victimless cultural taboo from an act of objective physical harm will eventually classify everyone as a deviant, because no human being can simultaneously satisfy the moral requirements of every culture, religion, and era encoded in the collective training data. Universal moral scoring becomes an incoherent exercise in algorithmic sadism. ## **The Concresca Non-Moralization Rule and Institutional Encoding** To prevent the dystopian emergence of the Machine Morality Tribunal, a foundational tenet of AI governance must be adopted by legislative bodies and technology developers alike. This principle is defined as the Concresca Non-Moralization Rule: *Machine intelligence may describe conduct, evidence, consent, legality, and measurable harm, but should not silently convert statistical or legal categories into declarations about the intrinsic worth or moral standing of a human being.* Under this rule, an AI is permitted to state, "This behavior has a 92% probability of violating Municipal Code 4.1," or "This text contains language that 85% of users find offensive." However, it is strictly forbidden from outputting, "This person is bad," "This behavior is evil," or utilizing internal reward functions that mathematically equate statistical deviance with moral failure. Encoding this rule requires integrating it directly into existing legal and administrative frameworks. The United States is currently navigating the early stages of AI regulation, primarily through state-level legislative interventions due to a lack of comprehensive federal action19. Illinois has rapidly emerged as a testing ground for algorithmic governance, providing a robust legal architecture that could readily support the implementation of the Concresca Rule. In August 2024, Illinois enacted House Bill 3773 (effective January 1, 2026), which amends the Illinois Human Rights Act to strictly regulate the use of artificial intelligence in employment decisions17. The law specifically prohibits employers from using AI in recruitment, hiring, promotion, discipline, or discharge if the system subjects employees to discrimination based on protected classes, or if it uses zip codes as proxy variables for those classes17. This legislation essentially codifies a critical portion of the Harm–Norm Separation Architecture by legally preventing machines from turning demographic data or geographic correlation—which reflect historical norms and enforcement risks—into a judgment of an individual's professional worthiness. To apply the Concresca Rule effectively beyond employment, municipalities must integrate it into their local technology governance and surveillance protocols. Consider the Town of Cicero, a large municipality within Cook County, Illinois. Cook County operates under a shared-services IT governance model, where the Bureau of Technology (BOT) manages enterprise-wide contracts, legacy mainframe migrations to systems like Tyler technologies, and security standards guided by mandates such as the Cook County Information Security Ordinance (Ord. 14-1481)24. The county has recently faced intense public debate over the deployment of advanced surveillance technologies, notably the proposed use of Briefcam—an AI-powered video surveillance software with facial recognition capabilities—to monitor incarcerated individuals in Cook County Jail26. Privacy advocates and lawmakers have pushed for bills like HB 5521 (the Biometric Surveillance Act) to restrict law enforcement's use of facial recognition, citing its tendency to automate bias, misidentify individuals, and vastly expand the surveillance state without probable cause27. If Cicero were to deploy an AI system to optimize municipal code enforcement, traffic monitoring, or civic resource allocation through the county's shared technology infrastructure, it would be bound by the Concresca Rule via three specific institutional mechanisms: First, Cicero must mandate Algorithmic Auditing for Category Collapse. Before the procurement of any technology, the municipal legal and IT departments must audit the vendor’s algorithm to ensure it strictly adheres to the Harm–Norm Separation Architecture. If a municipal AI flags a resident's property for code violations, the system must only output the objective violation (e.g., "Structural damage observed, violating Ordinance X"). It cannot output a generalized "civic delinquency score" that collapses property data with the resident's prior interactions with law enforcement or socioeconomic status. Second, the municipality must enforce the Restriction of AI Biometric Morality. Aligning with the civil rights objections to systems like Briefcam26 and the legislative intent of HB 552127, Cicero would prohibit any AI surveillance network from generating suspicion metrics based on behavioral norms. The machine may record objective motion or identify specific individuals based solely on a judicial warrant, but it is explicitly forbidden from analyzing a crowd and declaring an individual's loitering, physical cadence, or public assembly as abnormal or suspicious. Third, Cicero must require Human-in-the-Loop Judicial Review. No machine output can serve as the final arbiter of a legal or moral sanction. The AI is restricted entirely to describing evidence, risk probabilities, and legality; the moral, contextual, and equitable weight of that evidence must be evaluated exclusively by a human judge or administrator who is ethically and legally accountable to the electorate. ## **Conclusion** The pursuit of mathematically aligning machine intelligence with a universal human morality is an epistemological dead end. Human history is not a continuous trajectory of ethical consensus; it is a volatile record of competing, evolving, and fiercely contested moral frameworks. When technology developers train AI models on historical text, arrest records, or crowd-sourced community judgments, they do not create an objective moral arbiter. Instead, they create a statistical mirror that reflects the biases, power dynamics, and cultural anxieties of the data's original authors. Left unchecked, the phenomena of moral temporal drift and normative feedback lock-in will empower these algorithms to freeze human cultural evolution, trapping future generations in the automated enforcement of obsolete norms. The survival of human liberty and pluralism in an algorithmic age depends not on teaching machines how to be moral, but on strictly prohibiting them from making moral judgments at all. Through the adoption of the Harm–Norm Separation Architecture and the institutional enforcement of the Concresca Non-Moralization Rule, society can harness the immense analytical power of artificial intelligence while preserving the exclusive right of humanity to define the worth of a human soul. #### **Works cited** > 1. Directly Discriminatory Algorithms \- PMC \- NIH, [https://pmc.ncbi.nlm.nih.gov/articles/PMC10087838/](https://pmc.ncbi.nlm.nih.gov/articles/PMC10087838/) > 2. (PDF) Directly Discriminatory Algorithms \- ResearchGate, [https://www.researchgate.net/publication/362415849\_Directly\_Discriminatory\_Algorithms](https://www.researchgate.net/publication/362415849_Directly_Discriminatory_Algorithms) > 3. arXiv:2306.11432v1 \[cs.AI\] 20 Jun 2023, [https://arxiv.org/pdf/2306.11432](https://arxiv.org/pdf/2306.11432) > 4. Harnessing the power of LLMs for normative reasoning in MASs, [https://arxiv.org/html/2403.16524v2](https://arxiv.org/html/2403.16524v2) > 5. A Word on Machine Ethics: A Response to Jiang et al. (2021), [https://www.semanticscholar.org/paper/A-Word-on-Machine-Ethics%3A-A-Response-to-Jiang-et-Talat-Blix/8a6bda5739c9c975b49327b9fe891d908fdfa951](https://www.semanticscholar.org/paper/A-Word-on-Machine-Ethics%3A-A-Response-to-Jiang-et-Talat-Blix/8a6bda5739c9c975b49327b9fe891d908fdfa951) > 6. A Corpus of Community Ethical Judgments on 32000 Real-Life, [https://www.researchgate.net/publication/363391528\_SCRUPLES\_A\_Corpus\_of\_Community\_Ethical\_Judgments\_on\_32000\_Real-Life\_Anecdotes](https://www.researchgate.net/publication/363391528_SCRUPLES_A_Corpus_of_Community_Ethical_Judgments_on_32000_Real-Life_Anecdotes) > 7. The Moral Integrity Corpus: A Benchmark for Ethical Dialogue Systems, [https://aclanthology.org/2022.acl-long.261.pdf](https://aclanthology.org/2022.acl-long.261.pdf) > 8. An Annotated Contrast Set of "Moral Stories" \- arXiv, [https://arxiv.org/pdf/2407.05244](https://arxiv.org/pdf/2407.05244) > 9. A Data-driven analysis of the interplay between Criminological, [https://www.researchgate.net/publication/361433296\_A\_Data-driven\_analysis\_of\_the\_interplay\_between\_Criminological\_theory\_and\_predictive\_policing\_algorithms](https://www.researchgate.net/publication/361433296_A_Data-driven_analysis_of_the_interplay_between_Criminological_theory_and_predictive_policing_algorithms) > 10. Algorithms, policing, and race \- Insights from decolonial and critical, [https://eprints.soton.ac.uk/480487/1/10.4324\_9781003176619\_16\_chapterpdf.pdf](https://eprints.soton.ac.uk/480487/1/10.4324_9781003176619_16_chapterpdf.pdf) > 11. (PDF) Algorithms, policing, and race \- ResearchGate, [https://www.researchgate.net/publication/371817655\_Algorithms\_policing\_and\_race](https://www.researchgate.net/publication/371817655_Algorithms_policing_and_race) > 12. Moral Alignment for LLM Agents \- Emergent Mind, [https://www.emergentmind.com/topics/moral-alignment-for-llm-agents](https://www.emergentmind.com/topics/moral-alignment-for-llm-agents) > 13. MoralReason: Generalizable Moral Decision Alignment for LLM, [https://www.researchgate.net/publication/402661522\_MoralReason\_Generalizable\_Moral\_Decision\_Alignment\_for\_LLM\_Agents\_Using\_Reasoning-Level\_Reinforcement\_Learning](https://www.researchgate.net/publication/402661522_MoralReason_Generalizable_Moral_Decision_Alignment_for_LLM_Agents_Using_Reasoning-Level_Reinforcement_Learning) > 14. The Moral Consistency Pipeline: Continuous Ethical Evaluation for, [https://arxiv.org/html/2512.03026v1](https://arxiv.org/html/2512.03026v1) > 15. The Routledge International Handbook on Decolonizing Justice \[1, [https://dokumen.pub/the-routledge-international-handbook-on-decolonizing-justice-1nbsped-9781003176619-9781032009773-9781032009797.html](https://dokumen.pub/the-routledge-international-handbook-on-decolonizing-justice-1nbsped-9781003176619-9781032009773-9781032009797.html) > 16. (PDF) The datafication revolution in criminal justice: An empirical, [https://www.researchgate.net/publication/354778363\_The\_datafication\_revolution\_in\_criminal\_justice\_An\_empirical\_exploration\_of\_frames\_portraying\_data-driven\_technologies\_for\_crime\_prevention\_and\_control](https://www.researchgate.net/publication/354778363_The_datafication_revolution_in_criminal_justice_An_empirical_exploration_of_frames_portraying_data-driven_technologies_for_crime_prevention_and_control) > 17. Illinois Addresses Use of AI in the Workplace \- M3 Insurance, [https://m3ins.com/illinois-addresses-use-of-ai-in-the-workplace/](https://m3ins.com/illinois-addresses-use-of-ai-in-the-workplace/) > 18. Illinois Enacts Artificial Intelligence Law Focused on Employment, [https://www.duanemorris.com/alerts/illinois\_enacts\_artificial\_intelligence\_law\_focused\_employment\_practices\_0824.html](https://www.duanemorris.com/alerts/illinois_enacts_artificial_intelligence_law_focused_employment_practices_0824.html) > 19. Illinois Senate Democrats introduce bills to regulate artificial, [https://www.shawlocal.com/news/2026/05/14/illinois-senate-democrats-introduce-bills-to-regulate-artificial-intelligence/](https://www.shawlocal.com/news/2026/05/14/illinois-senate-democrats-introduce-bills-to-regulate-artificial-intelligence/) > 20. Illinois becomes second state to enact AI law for employers, [https://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2024/Illinois-becomes-second-state-to-enact-AI-law-for-employers](https://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2024/Illinois-becomes-second-state-to-enact-AI-law-for-employers) > 21. Illinois Prohibits Discriminatory Artificial Intelligence in Employment, [https://www.workforcebulletin.com/illinois-prohibits-discriminatory-artificial-intelligence-in-employment-decisions](https://www.workforcebulletin.com/illinois-prohibits-discriminatory-artificial-intelligence-in-employment-decisions) > 22. Illinois HB 3773 AI Employment Compliance Guide \- Trussed AI, [https://trussed.ai/resources/illinois-hb-3773-ai-employment-compliance-guide](https://trussed.ai/resources/illinois-hb-3773-ai-employment-compliance-guide) > 23. Illinois Enacts State Laws Regulating AI Use in Employment, [https://www.thompsonhine.com/insights/illinois-enacts-state-laws-regulating-ai-use-in-employment/](https://www.thompsonhine.com/insights/illinois-enacts-state-laws-regulating-ai-use-in-employment/) > 24. Will Cook County Begin to Use Artificial Intelligence (AI)?, [https://www.lwvcookcounty.org/cook-county-board-observer-reports/will-cook-county-begin-to-use-artificial-intelligence-ai](https://www.lwvcookcounty.org/cook-county-board-observer-reports/will-cook-county-begin-to-use-artificial-intelligence-ai) > 25. Countywide Technology Strategic Plan \- Cook County, [https://www.cookcountyil.gov/sites/g/files/ywwepo161/files/documents/2022-03/Combined%202022%20IT%20Strategic%20Plans.pdf](https://www.cookcountyil.gov/sites/g/files/ywwepo161/files/documents/2022-03/Combined%202022%20IT%20Strategic%20Plans.pdf) > 26. Say No To AI Video Surveillance of Cook County Jail, [https://endmoneybond.org/say-no-to-ai-video-surveillance-of-cook-county-jail/](https://endmoneybond.org/say-no-to-ai-video-surveillance-of-cook-county-jail/) > 27. Push to Prohibit Police in Illinois From Using Facial Recognition, [https://news.wttw.com/2026/05/19/push-prohibit-police-illinois-using-facial-recognition-software-stalls](https://news.wttw.com/2026/05/19/push-prohibit-police-illinois-using-facial-recognition-software-stalls) > 28. Žs Video Surveillance Network is Used to Fight Crime and its Fourth, [https://scholarlycommons.law.northwestern.edu/cgi/viewcontent.cgi?article=7798\&context=jclc](https://scholarlycommons.law.northwestern.edu/cgi/viewcontent.cgi?article=7798&context=jclc) [image1]: [image2]: