Ethical & Legal Considerations¶

This work is licensed under a Creative Commons Attribution 4.0 International License.

Foundations of the Ethical principles for AI¶

This lesson focuses on the ethical principles that ground AI in a legal landscape.

Science Fiction or a Philosophical Theory?¶

In the early 1950's Alan Turing the father of all modern computing, proposed a test for intelligence in a computer, requiring that a human being should be unable to distinguish the machine from another human being by using the replies to questions put to both.

Author Isaac Asimov wrote a series of popular science fiction novels in the 1950's through the 1980's. His work continues to be adapted into television series and movies. In his novels, Asimov developed Three Laws of Robotics which described how artificial intelligence interacted with humanity in his fictional universe.

The Three Laws

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.

2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.

3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Asimov later wrote of a 'zeroth' law which superceded the first three laws,

0. A robot may not injure humanity or, through inaction, allow humanity to come to harm.

Asimov's Three Laws are difficult to interpret in a real-world setting and he himself spent most of his novels describing creative and unexpected ways in which the Three Laws were twisted yet not broken. The basis of the Three Laws as a legal framework is untenable, but does represent a moral and ethical starting point from which we can think about AI and the legal rights of non-biological beings.

Another science-fiction author Sir Arthur C. Clarke, in 1978 provided an interesting perspective on how humanity would have to come to terms with AI once its capabilities surpass our own:

Recently, researchers published findings showing that current GPTs are now capable of passing Turing tests. As our conception of intelligence shifts (Mitchell 2024), mostly in reaction to the release of ChatGPT and its myriad of competitors, new standards of the Turing Test are being proposed.

Importantly, current AI exposes the limits of Turing Tests based on imitation without comprehension.

The Turing Trap is a term coined by Stanford University professor Erik Brynjolfsson to describe the idea that focusing too much on developing human-like artificial intelligence (HLAI) can be detrimental.

Brynjolfsson argues that the real potential of AI lies in its ability to augment human abilities, rather than replacing them. He suggests that we should work on challenges that are easy for machines and hard for humans, rather than the other way around.

Beware the Turing Trap

Automation can replace humans

HLAI can replace humans in the workplace, which can lead to:

• Lower wages

  As machines become better substitutes for human labor, wages can be driven down.

• Loss of economic and political power

  Workers can lose economic and political bargaining power, and become increasingly dependent on those who control the technology.

• Decision-making processes incentivize automation

  Companies may choose to automate tasks to do the same thing faster and cheaper.

• Misaligned incentives

  The risks of the Turing Trap are increased by the misaligned incentives of technologists, businesspeople, and policy-makers.

this text was written by AI and then reviewed by a human. Do you still trust it?

Researchers have found disclosing the use of AI makes people trust you less.

Ethical AI¶

In "A Unified Framework of Five Principles for AI in Society" (Floridi & Cowls 2019) core principles for ethical AI are introduced (Table 1).

Table 1: Floridi & Cowls (2019) Five principles for AI in Society¶

International Agreements on AI¶

A milestone in the Ethics of Artificial Intelligence () occurred in January 2017 in Pacific Grove, California at the historic Asilomar Hotel and Conference Grounds (Table 2). There the Asilomar AI Principles were signed by leading AI researchers, ethicists, and thought leaders.

By 2021, UNESCO had created their own recommendations on AI, focused on human rights and sustainable development.

Table 2: International AI agreements¶

In response to the rapid rise of generative AI, specifically GPTs, new agreements on the application of AI for military use, safety, and on its adoption in business and industry were recently signed (Table 3).

Table 3: Declarations on AI¶

Blueprint for an AI Bill of Rights¶

How should the values that guide AI systems be set, and by whom? Two distinct answers have emerged: corporate AI constitutions written by AI companies for their own models (Anthropic's Claude Constitution is the canonical example), and public AI bills of rights developed through democratic processes (the White House Blueprint for an AI Bill of Rights, October 2022, is the leading example). Sociologist Alondra Nelson — who led the Blueprint's development as acting director of the White House Office of Science and Technology Policy (OSTP) — argues in A civic grammar for AI rights (Science, 2026) that these two forms of foundational document do very different work, and that one of them is structurally insufficient as a source of democratic legitimacy.

A third category emerged in May 2026 with Pope Leo XIV's first encyclical Magnifica Humanitas: a transnational moral institution speaking on behalf of 1.4 billion Catholics, claiming authority against both the technocratic paradigm and the states that abdicate the field. The encyclical is treated in its own section below.

Timeline and status¶

The Blueprint for an AI Bill of Rights emerged from a public process announced in an October 2021 Wired essay by the White House OSTP. It was released in October 2022 with five principles to guide the design, development, and deployment of automated systems:

1. Safe and Effective Systems — protection from unsafe or ineffective systems.
2. Algorithmic Discrimination Protections — equitable design and use of automated systems.
3. Data Privacy — protection from abusive data practices, with agency over how your data is used.
4. Notice and Explanation — knowing when an automated system is being used and how/why it affects you.
5. Human Alternatives, Consideration, and Fallback — the ability to opt out and reach a human alternative when an automated system fails or causes harm.

These principles were incorporated into President Biden's October 2023 Executive Order on Safe, Secure, and Trustworthy AI. The Executive Order was rescinded by President Trump's January 2025 Removing Barriers to American Leadership in Artificial Intelligence. The Blueprint itself was always non-binding guidance.

Corporate constitutions vs. public bills of rights¶

• Corporate AI constitutions are internal training and alignment specifications. They describe a company's vision of how its model should behave. They are not negotiated with the publics affected by the model's deployment, and they can be revised by company fiat. Nelson notes that newer revisions of Anthropic's Claude Constitution have quietly removed references to international human rights agreements — and with them protections for "personal liberty, freedom of religion and intellectual property" that earlier versions included. Historian Jill Lepore observed that the document arrived "at a trying time for both artificial intelligence and constitutional democracy."
• Public AI bills of rights declare rights claims that publics can extend to new institutions and new harms. Their force comes from democratic legitimacy and the broader legal-political infrastructure, not from the model developers.

The deeper question, Nelson asks, is who gets to author the foundational documents that govern AI — companies, or publics?

Constitution vs. Bill of Rights vs. Declaration of Independence¶

Nelson argues that the Blueprint drew its name from the first ten amendments to the U.S. Constitution, but in structural terms resembles a different founding document: the Declaration of Independence.

"Unlike the Bill of Rights, the Blueprint does not establish courts or enforcement mechanisms. It does not create procedures for redress. It declares. It states principles and claims rights against a concentration of power that most Americans cannot meaningfully constrain through existing institutions or processes."

The five principles, Nelson argues, are statements of values — "social expectations, stated in the vocabulary Americans reach for when they want to contest power: Patients' Bill of Rights, Consumer Bill of Rights, Tenants' Bill of Rights, Workers' Bill of Rights, Taxpayers' Bill of Rights." The Blueprint extended that civic grammar to algorithmic systems.

"Civic grammar" and the diffusion of rights claims¶

Nelson describes what has emerged as a "civic grammar": a shared vocabulary of rights claims (non-discrimination, transparency, data privacy, notice, human alternatives) that publics can extend to new institutions and new harms, and that "has been traveling across jurisdictions, partisan lines, and institutional contexts."

This pattern reflects what sociologists David Strang and John Meyer call institutional diffusion among weakly related actors — a conceptual rather than relational mechanism by which abstract typologies become "a strategy for making sense of the world." Connecticut Democrats, Oklahoma Republicans, Florida's Republican governor, and a national student-advocacy network can adopt the same vocabulary without coordinating, because all are responding to the same structural condition: AI reshaping people's lives without their consultation.

The Blueprint's "second life": cross-partisan diffusion¶

Although rescinded at the federal level, the Blueprint, in Nelson's words, "has done what its metaphor suggests blueprints do: it has been built upon." State legislatures, governors, and advocacy organizations have produced their own AI bills of rights drawing directly from the five principles — often across explicitly opposed political coalitions:

• Connecticut (2023): Democratic Governor Ned Lamont signed legislation directing state policy-makers to develop their own AI Bill of Rights.
• Oklahoma (2024): The Republican-controlled House of Representatives introduced and passed an AI Bill of Rights, though it was not ultimately codified into law.
• Florida (2025): Republican Governor Ron DeSantis pushed for an AI Bill of Rights through executive action and twice backed Florida Senate Bill 482, the "Artificial Intelligence Bill of Rights," which would codify several Blueprint principles into Florida law. The bill was blocked in the Florida House, where the Speaker aligned with the Trump administration's effort to prevent states from regulating AI.
• Student AI Bill of Rights (2026): The National Student Legal Defense Network released a Student AI Bill of Rights.

As Nelson puts it: "What the actors share is a vocabulary and a common perspective that AI is reshaping people's lives without their consultation."

Marshall's social citizenship and the AI rights tier¶

Nelson grounds the analysis in British sociologist T. H. Marshall's 1950 essay Citizenship and Social Class. Marshall argued that rights expand historically through successive waves of claim-making: civil rights extending to political rights, political rights extending to social rights — entitlements to economic security and the conditions of participation, against harms of industrial capitalism that individual civil-liberties frameworks could not address. Rights, in Marshall's account, "are never fully delivered at the moment of declaration. They are successively rearticulated by publics who attempt to hold institutions to commitments those institutions have not yet honored."

The Blueprint's five principles, Nelson argues, map onto Marshall's social-rights tier. They are not classical civil liberties; they are entitlements against systems "that increasingly govern access to employment, credit, healthcare, housing, and education." The structural parallel — collective, diffuse, opaque harms that older rights frameworks address only partially — is what explains the cross-partisan convergence: "actors who disagree on nearly everything else agree that algorithmic power requires a social citizenship response."

Three imperatives for studying AI: question, object, tool¶

Nelson's argument about AI rights builds on a broader claim she develops in Field Theory: AI as Social Science Question, Object & Tool (Daedalus, Winter/Spring 2026): AI models, tools, and systems pose three interrelated imperatives for the social sciences.

1. AI as social-science question. Renewed attention to social theories of how technology, human experience, and social order are entangled. Nelson reaches back to Weber's analysis of rationalization and W. E. B. Du Bois's study of technology and inequality, and forward to contemporary scholarship on algorithmic governance.
2. AI as object of inquiry. AI systems themselves require study as social, political, and economic artifacts — not just engineering products. Their training corpora, their labor relations, their ideological commitments, and their effects on the institutions that deploy them all merit investigation in their own right.
3. AI as method/tool. AI capabilities may transform — or upend — the practice of social investigation itself: large-scale text analysis, multimodal pattern detection, conversational interviewing at scale. That transformation deserves critical scrutiny rather than uncritical adoption.

The capacities social science distinctively brings to all three, Nelson argues: historicizing the apparently unprecedented, tracing connections across scales (from individual experience to institutional behavior to political economy), and centering those most affected by technological change.

This three-part framing also doubles as a useful diagnostic. When you read a piece of AI-ethics scholarship, ask which of the three imperatives it engages — that often clarifies what kind of argument is being made and what kind of counter-argument would land.

International convergence¶

Legal scholar Yuval Shany has surveyed international standard-setting instruments — the EU AI Act, the Council of Europe Framework Convention on AI, the United Nations Global Digital Compact, and national legislation in South Korea and Italy — and finds they coalesce around the same protections as the Blueprint: non-discrimination, transparency, data privacy, and human alternatives. The pattern predates the Blueprint: the EU's General Data Protection Regulation (GDPR) established data-protection rights nearly a decade earlier. The "bill of rights" frame is American; the underlying rights-claim convergence is global.

The encyclical Magnifica Humanitas (below) converges on the same protections from an entirely different starting point — theological anthropology rather than democratic theory or international law.

The limits of rights talk¶

Nelson is clear-eyed about what civic grammar cannot do on its own:

• Accommodation can mimic transformation. A vocabulary that moves easily across partisan lines may have been "drained of the political content that gives rights claims their force."
• Rights individualize structural problems. Frameworks built around individual claims often fail to address the collective and systemic nature of algorithmic harms.
• Declaration is not delivery. History shows "declarations of entitlement and their substantive delivery can remain decades apart, separated by the organized power of those who benefit from the status quo."

Nelson's clarifying question for democratic institutions: will they take this civic grammar seriously before the AI companies finish writing their own constitutions for us all?

From declaration to enforcement¶

The harder labor that remains is "translating the grammar of rights into the standards, audit protocols, and enforcement mechanisms that give those sentences force." The institutional models exist:

• Algorithmic impact assessments required before AI systems are shipped.
• Standardized evaluation methods for detecting algorithmic risk and harm.
• Independent audit frameworks that subject deployed systems to outside scrutiny.

What is missing, in Nelson's account, is the political will to extend existing accountability frameworks to a domain that has so far resisted them — and the institutional coalitions, not just the vocabulary, required to deliver the principles the Blueprint declared.

Catholic social teaching: Magnifica Humanitas (Pope Leo XIV, 2026)¶

On May 25, 2026, Pope Leo XIV released Magnifica Humanitas ("Magnificent Humanity"), his first encyclical, addressed to Catholics and "every person of goodwill." The 235-page document treats artificial intelligence as the central moral question of the age and frames AI as a new industrial revolution requiring a parallel foundational moral response.

The signing date is deliberate: May 15, 2026 is the 135^th anniversary of Pope Leo XIII's Rerum Novarum (1891), the foundational encyclical of modern Catholic social teaching, written to address the dignity of workers amid the 19^th-century industrial revolution. Leo XIV explicitly places AI alongside that earlier disruption as a moment demanding renewed teaching.

In a break with tradition, Pope Leo personally presented the encyclical at the Vatican alongside Chris Olah, co-founder of Anthropic — the first time a pontiff has presented an encyclical himself rather than delegating the task to cardinals.

Core teachings¶

• The centrality of the human person. The dignity of the human person is affirmed as infinite. Human beings take precedence over AI, and any deployment of AI must be evaluated against that priority.
• Critique of transhumanism — the project of using technology to overcome biological limits such as aging. Leo XIV rejects the framing of human finitude as a problem to be engineered away.
• Critique of posthumanism — the philosophical position that blurs the boundaries between humans and machines, or denies the distinctiveness of human beings. The encyclical names this as an active "anti-human vision" embedded in contemporary AI development, not merely a speculative philosophical stance.
• Catholic social doctrine as the evaluation framework: dignity of the person, the common good, and justice serve as the principles against which any AI deployment should be measured.
• Coverage extends across education, the economy, unemployment, work, human trafficking, and war — the same broad social-impact terrain Catholic social teaching has historically addressed.

Calls to action¶

• "Disarm AI" — withdraw AI from military applications and from purely economic interests; direct it to the common good.
• Stricter state and international regulation of AI companies, not industry self-governance.
• An explicit address to "every person of goodwill" extends the encyclical's claims of moral force beyond its Catholic audience — a move consistent with how Laudato Si' (2015, on climate) and the Vatican's Antiqua et Nova (January 2025, the Dicastery for the Doctrine of the Faith's earlier note on AI) sought ethical common ground beyond doctrinal lines.

Convergence with Nelson's civic grammar¶

The encyclical situates AI alongside the industrial revolution as a moment requiring foundational moral and institutional response rather than industry self-governance. That framing converges with the civic-grammar argument above — both Pope Leo and Alondra Nelson arrive at the same conclusion from very different starting points: corporate self-governance is structurally insufficient as a source of legitimacy.

What the encyclical uniquely adds is the extension of the historic Catholic doctrine of the universal destination of goods to "patents, algorithms, digital platforms, technological infrastructure and data" — its strongest novel claim, and an explicit critique of frontier-lab concentration. It also introduces a theological and humanistic vocabulary (transhumanism, posthumanism, infinite human dignity, the common good) that may not appear in technical AI-ethics literature but increasingly shapes public reception of AI. What it does not provide is the operational machinery — algorithmic impact assessments, audit protocols, model evaluations — that Nelson identifies as the harder labor of translating grammar into enforcement.

The decision to present the encyclical alongside an AI-company executive signals that the Church views the conversation as dialogic rather than purely adversarial.

Current Legislation¶

National Conference of State Legislatures (NCSL) Artificial Intelligence Legislation Database

The current administration has focused most of its efforts on executive orders related to AI and federal agencies. The December 2025 Executive Order 14365 "Ensuring a National Policy Framework for Artificial Intelligence" asserts federal authority to challenge or override state AI laws through DOJ litigation, conditioned BEAD broadband funding, and an FTC policy statement. The June 2026 Executive Order "Promoting Advanced Artificial Intelligence Innovation and Security" adds a voluntary federal benchmarking process for frontier-model cyber capabilities, while explicitly disclaiming any mandatory licensing regime. See also pending congressional legislation that would codify state-preemption.

As of today, there are no comprehensive federal laws or regulations that have been enacted to specifically regulate AI in the United States of America. The June 2026 order continues the administration's pattern of voluntary, security-framed measures rather than binding regulation.

AI Ethics¶

As consumers of GPTs and other AI platforms, we must consider in what ways can we use AI both effectively, and ethically.

When can you use a GPT for research and education?

graph TB
  A((Start)) --> B("Does it matter if the outputs are true?");
  B -->| No | F("Safe to use GPT");
  B -->| Yes | C("Do you have the ability to verify output truth and accuracy?");
  C -->| Yes | D("Understand legal and moral responsibility of your errors?");
  C -->| No | E("Unsafe to use GPT");
  D -->| Yes | F("Safe to use GPT");
  D -->| No | E("Unsafe to use GPT");

  style A fill:#2ECC71,stroke:#fff,stroke-width:2px,color:#fff
  style B fill:#F7DC6F,stroke:#fff,stroke-width:2px,color:#000
  style C fill:#F7DC6F,stroke:#fff,stroke-width:2px,color:#000
  style D fill:#F7DC6F,stroke:#fff,stroke-width:2px,color:#000
  style E fill:#C0392B,stroke:#fff,stroke-width:2px,color:#fff
  style F fill:#2ECC71,stroke:#fff,stroke-width:2px,color:#fff

Figure credit: ChatGPT and Artificial Intelligence in Education, UNESCO 2023

Recent Controversy¶

Maps of AI Copyright Lawsuits¶

Master list of current lawsuits against AI companies

Courts are also beginning to hold AI companies directly liable for what their models assert. In June 2026 the Regional Court of Munich ruled against Google's AI Overviews, treating the AI-generated answers as Google's own speech rather than as neutral search results. The Overviews had tied two Munich publishers to "scams," "subscription traps," and "dubious business practices" — connections that appeared in none of the linked sources — and the court issued an injunction barring Google from repeating them. With barely 1% of readers clicking through to a source, the court was unmoved by the argument that this is simply how search works now: nobody needs an AI layer to search the internet.

Current AI models are overwhelmingly based on European and North American historical literature and language. Over half of the content on the internet () is written in English. This creates a Eurocentric bias in AI training data, resulting in an erasure of global culture, experience, and language. Such asymmetries need to be addressed, but there is at present a lack economic incentives for large tech companies and organizations (see The Imitation Game above).

Negative consequences of GPTs explosion into the public space are its mis-use as well as its adoption for illegal activity.

• A lawyer submits a legal brief written by ChatGPT and is caught (2023)

• By 2026 the problem had escalated from one embarrassed attorney to entire cases collapsing: a senior federal judge in Mississippi, Sharion Aycock, sanctioned all four lawyers in an Aberdeen fee dispute after catching both sides independently filing AI-generated briefs full of fabricated citations and bogus quotations. She paused the trial, disqualified every attorney, barred the two who admitted using AI from her court for two years, and fined them between $1,000 and $3,500 — one of the strongest judicial rebukes of courtroom AI misuse to date.

• Prompt Injection Attacks

There are deep ethical concerns about the use of AI like GPT and LLMs, particularly concerning their training data.

• A Class Action lawsuit: GitHub CoPilot was trained on private repositories

AI companies also effectively steal designs, visual art, and music styles to train their private models.

• MidJourney and Dall-E using trademarked imagery and art

ChatGPT has effectively gamified higher education and is being used to spread disinformation and hate speech.

• College students using ChatGPT

• Combating Academic Dishonesty (6-part series)

Recent Literature¶

Here are some recent papers that discuss the ethical concerns surrounding AI:

• "AI Safety and the Age of Convergences" (2024) - Schuett, J., Schuett, J., & Korinek, A. https://doi.org/10.48550/arXiv.2401.06531

• "On the Opportunities and Risks of Foundation Models" (2023) - Bommasani et al. https://doi.org/10.48550/arXiv.2108.07258

• Unraveling the Ethical Conundrum of Artificial Intelligence: A Synthesis of Literature and Case Studies Poli, P.K.R., Pamidi, S. & Poli, S.K.R. Augment Hum Res 10, 2 (2025). https://doi.org/10.1007/s41133-024-00077-5

• "The Ethics of Artificial Intelligence in Education: A Review of the Literature" (2023) - Zawacki-Richter, O., Marín, V. I., Bond, M., & Gouverneur, F. https://doi.org/10.1007/s10639-019-09882-z

• "The Ethical Challenges of Algorithmic Bias in Artificial Intelligence: a scoping review" (2023) - Borenstein, J., Glikson, E., & Krishnamurthy, V. https://doi.org/10.1007/s43681-023-00313-z

• "Ethics of Artificial Intelligence" (2020) - S. Matthew Liao https://doi.org/10.1093/oso/9780190905033.001.0001

• The Ethics of AI Ethics: An Evaluation of Guidelines. (2020) Hagendorff, T. Minds & Machines 30, 99–120. https://doi.org/10.1007/s11023-020-09517-8

Assessment¶