Introduction
AI ethics loves to cosplay as a brand-new problem. New model, new panic, new "unprecedented" headline. But the shape of the debate is ancient: humans build things that imitate life, then immediately worry about (1) control, (2) responsibility, and (3) whether we've accidentally outsourced our conscience to a tool that doesn't have one. This essay sketches that long arc, from myth and moral philosophy, through cybernetics and early algorithmic discrimination, to the very fractured governance landscape we find ourselves in as of February 2026.
Myth and philosophy: artificial agency before microchips
Long before the silicon revolution, cultures built moral thought-experiments out of folk tales and mythologies. These stories weren't pre-scientific entertainment; they were early simulators for questions we still haven't finished answering: What counts as agency? Who is responsible when a powerful thing follows orders? What happens when human desire builds a being that looks obedient, alive, or both?
In Greek mythology, Talos, a giant bronze automaton from Greek mythology designed by Hephaestus (the Greek god of fire, metallurgy, forges, and craftsmanship) to protect Crete, executes its mandate with lethal literalism. Talos belongs to a wider Greek habit of telling cautionary tales about power without wisdom: hubris is the engine of many tragedies, and the gods (or fate) have a nasty tendency to punish overconfidence with precision. The ethics is however baked into the warning label. Talos doesn't decide, but executes. That is exactly why the story still maps cleanly onto modern fears about lethal autonomous systems and "compliance without compassion"1–2.
In Ovid's Pygmalion myth is often taught as romance. The artist makes a statue, the statue becomes beloved, reminiscent of the 80s classic Mannequin. But the ethical reading is more complicated. Ovid is writing in a Roman world saturated with questions about art, power, and the moral risks of turning living beings into objects (and objects into living beings). The statue's "agency" is a gift from the gods, but it's also a product of Pygmalion's desire. The story is basically a prototype for critiques of systems engineered to be agreeable, flattering, and socially compliant, especially when those systems mimic personhood3–5.
The Golem of Prague shifts the moral burden onto the creator. This legend sits in the broader Jewish tradition of debating what it means to create, to name, and to wield sacred power responsibly. In many tellings, the golem is animated by words, sometimes the divine name, sometimes an inscription like emet ("truth"), and deactivated by altering that inscription. That detail matters: the entity has force, but does not have an internal moral compass or capacity for ethical judgment. The danger isn't that it becomes evil, but that it's humans treat the act of creation as a permission slip to abdicate responsibility ("the system did it!")6–8. This is the alignment debate in folkloric attire. Power without an internal moral framework still creates moral debt, owed by the builder.
Chinese moral traditions, especially Confucian ethics, tend to evaluate moral life as relational and role-based, rather than adversarial rights claimed between isolated individuals. Historically, Confucian thought develops in response to social disorder. It asks how communities become stable, humane, and worth living in, not by perfect rules alone, but by cultivating character and fitting conduct to context—family roles, civic duties, ritual practice, and reciprocal obligation. That framing naturally pushes "ethical AI" away from the fantasy of a solitary, neutral tool and toward sociotechnical systems, i.e. humans and machines co-shape norms, obligations, and cooperative practice9–11.
Cybernetics and early machine ethics: the 20th-century hinge
In the 1940s, Norbert Wiener formalised cybernetics, the science of control and communication in animals and machines, and warned that automated decision systems would collide with labour, power, and social stability12. The big move here is inevitably conceptual: treat humans and machines as comparable information-processing systems, then ask what happens when one starts managing the other.
It was Alan Turing's Imitation Game (more widely known as "the Turing Test") that gave a pragmatic benchmark for machine intelligence, but it also set a trap: the risk of equating linguistic mimicry with understanding — a trap we find ourselves in at the moment. That confusion becomes ethically expensive once systems can speak fluently without grounding 13–14.
In the 60s Joseph Weizenbaum built ELIZA and was alarmed at how quickly people projected mind and care onto pattern matching. His thesis in Computer Power and Human Reason is an ethical boundary argument that states computation can simulate conversation, but it cannot supply wisdom, compassion, or moral judgment, and society should not hand over "human" tasks to systems that only look human. He forcefully contended that society must never delegate tasks requiring wisdom to computational systems, warning against the human psychological vulnerability to ascribe sentience to unfeeling algorithms 15–17.
The 1980s proved the point. Bias doesn’t disappear in mathematics.
The cautionary tales turned empirical in the UK. At St George’s Hospital Medical School, an admissions algorithm trained on historical decisions encoded discrimination by systematically penalising women and applicants with non-European sounding names. The Commission for Racial Equality investigation (1988) made the lesson explicit by revealing that a computer can become a laundering machine for human prejudice by producing biased outcomes with under the false guise of objectivity 18–20.
This is a foundational ethical result that automation is not moral sterilisation. It’s moral reproduction at scale!
Deep learning and the rise (and limits) of “FATE”
The post-AI-winter renaissance replaced symbolic rules with statistical learning from large datasets. That shift multiplied ethical risk because the "rules" became opaque correlations in data rather than explicit logic 21–22.
To answer this, the field gravitated toward FATE: Fairness, Accountability, Transparency, and Ethics, along with institutional guidance (notably from UK research bodies like the Alan Turing Institute). FATE is useful, but it struggles when models become too large, too general, and too socially entangled 23–25.
Stochastic Parrots: the 2021 turning point
The Stochastic Parrots paper argued that scaling language models does not automatically produce understanding; it produces fluent statistical behaviour with serious externalities like environmental costs, dataset-baked hegemonic bias, and the ability to generate persuasive nonsense 26–27. The controversy around the paper (and its authors' treatment) made "near-term harms" impossible to dismiss as mere hypotheticals 27.
2025–2026: alignment, interpretability, and a wobbling safety culture
By early 2026, the debate has forked:
- Near-term harms: bias, surveillance, labour exploitation, misinformation, and power concentration.
- Long-term/existential risk: highly capable systems pursuing goals misaligned with human intent.
Alignment now increasingly treats values as non-stationary (through shifts in human preferences) and safety as a moving target. Work presented around major ML venues continues exploring alignment under competition and misalignment, suggesting decentralised checks may sometimes reduce catastrophic failure modes 28–29.
A recurring ethical demand is not just "explainability theatre" but verifiable tracing by means of a transparent audit trail. The "black box" problem is not just a marketing problem; it’s a governance problem. Regulators and the public want to know not just what a system did, but the internal causal pathway that produced the output. A push for certification-style interpretability, similar to financial auditing more than marketing is growing.30
In February 2026, reporting confirmed that OpenAI disbanded its Mission Alignment team and reassigned staff, after the team’s formation in late 2024 31–32. In contrast, Anthropic expanded external safety pipelines, opening applications for fellows beginning May and July 2026 33. Whatever you believe about AGI timelines, this divergence matters since governance is being shaped as much by organisational incentives as by technical arguments.
The fractured geopolitics of AI governance (February 2026)
No unified "global AI regime" exists. Instead we have blocs, each encoding different values and power structures. I will briefly elaborate on the major ones, but the key point is that the ethical landscape is deeply fractured and contested, with no guarantee of convergence, yet.
European Union: risk-based regulation with a ticking timeline
The EU’s AI Act is the most comprehensive horizontal framework currently on the board.
- Prohibited practices under Article 5 became effective 2 February 2025 and enforceable 2 August 2025 34–35.
- The broader Act phases in, with major applicability milestones around 2 August 2026 36.
- Implementation depends heavily on harmonised standards (most notably via CEN-CENELEC), and delays in standard-setting have been flagged publicly 37–38.
The ethical risk is that the Act’s risk-based approach may be undermined if the standards that define compliance are not ready in time, leading to a regulatory gap or uneven enforcement.
United Kingdom: “pro-innovation” meets rising public demand
The UK’s approach has leaned principles-based and regulator-led. But public pressure is moving the equilibrium; 72% of the British public say laws and regulation would make them more comfortable with AI39, but the UK is visibly drifting toward more statutory intervention and institutional capacity-building 39–40. The ethical risk is that a "pro-innovation" stance may be perceived as regulatory capture or a failure to protect citizens, especially if high-profile incidents occur.
United States (California): state-level frontier rules become de facto standards
With federal consensus still elusive, California continues to legislate aggressively on AI, and its market power means these laws often become de facto standards for the US and beyond:
- SB 53 (Transparency in Frontier Artificial Intelligence Act) was signed in 2025 and ties multiple disclosure/testing obligations to January 2026 in state law materials 41–42.
- AB 316 took effect on 1 January 2026 and aims to prevent defendants from using "the AI acted autonomously" as a liability escape hatch 43–45. Good.
The philosophical move here is basically Golem-ethics made statutory by imposing that accountability sticks to builders and deployers. See my blog on the role of Machine Learning Engineers standing at the precipice of that ethical responsibility here.
China: state-security alignment with escalating penalties
China's approach remains iterative, application-specific, and tightly bound to content and data control. Reporting on amendments effective 1 January 2026 highlights dramatically increased penalties, with fines reaching up to RMB 10 million in serious cases 46–47. The ethical risk is that the focus on state security and content control may stifle innovation and suppress dissent, while the severity of penalties could create a chilling effect on legitimate AI development and use.
But did you see China's new robots?! I need proof that these reports aren't AI-generated...
United Nations and the Global South: capacity and equity as the meta-problem
The most under-discussed risk is the AI divide: unequal access to compute, data, and governance capacity. Something we know all too well in Africa.
At the AI Impact Summit in New Delhi (February 2026), reporting described major investment ambitions and a strong equity narrative. Importantly, UN Secretary-General António Guterres called for a $3 billion fund to help ensure AI benefits all rather than entrenching inequality 48–49. Whether that fund materialises is an open political question, but the ethical diagnosis is clear. The AI era is also a resource-distribution era, which eerily echoes the Golem’s moral debt. If we build powerful tools, we owe it to the world to ensure they don’t just serve the privileged few, or we end up in an era of neo-colonial AI, where the Global South is left to deal with the fallout of systems designed and deployed by the Global North.
Closing thought: the ethical through-line is power, not novelty
The myths were never about “robots” in the modern sense. They were about humans. They were about our appetite for control, our talent for self-deception, and our habit of building tools that magnify whatever values we smuggle into them.
The St George's case proved that computation won't cleanse discrimination. The LLM era proved that fluency is not truth. And the regulatory reality of February 2026 proves that governance will be plural, contested, and deeply shaped by geopolitics.
So the real ethical imperative is boring in a powerful way. We need to keep accountability attached to human institutions, keep systems auditable, and keep the benefits of AI from becoming just another mechanism for concentrating wealth, surveillance, and control. I think we need to reconsider building AI systems for the sake of having AI systems, and instead ask: what are we building, who is it for, and what power dynamics are we reinforcing or disrupting? It's a classic case of a typical arms race that we can only win by refusing to play. The ethical through-line is not novelty; it's power. Is this a bubble? And if it is a bubble, capitalism is the air...
References
1. Lessons in Greek Mythology: Hubris — MHS Headlight. https://mhsheadlight.com/2022/05/16/lessons-in-greek-mythology-hubris/ ↩
2. Charles Handy, The deadly dangers of hubris — The Idler. https://www.idler.co.uk/article/charles-handy-the-deadly-dangers-of-hubris/ ↩
3. Pygmalion (mythology) — Wikipedia. https://en.wikipedia.org/wiki/Pygmalion_(mythology) ↩
4. What is the Symbolism Of Pygmalion and Galatea? — Rest and Trust. https://restandtrust.org/what-is-the-symbolism-of-pygmalion-and-galatea/ ↩
5. Ian Lockey, Not a “charming little tale”: Teaching the Pygmalion Myth Ethically — Medium. https://medium.com/ad-meliora/not-a-charming-little-tale-teaching-the-pygmalion-myth-ethically-7e918b00384b ↩
6. An Epoch of Golem-Making': Artificial Intelligence and the Jewish Imaginary — LCFI. https://www.lcfi.ac.uk/news-events/blog/post/an-epoch-of-golem-making-artificial-intelligence-and-the-jewish-imaginary ↩
7. The Curious Case of the Golem & Artificial Intelligence — Itzikr’s Blog. https://itzikr.wordpress.com/2023/06/09/the-curious-case-of-the-golem-artificial-intelligence/ ↩
8. From the Tree of Knowledge & the Golem of Prague… — SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3675987 ↩
9. Applying Ancient Chinese Philosophy To Artificial Intelligence — Noema. https://www.noemamag.com/applying-ancient-chinese-philosophy-to-artificial-intelligence/ ↩
10. Chinese Ethics — Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/ethics-chinese/ ↩
11. Confucian Robot Ethics — ODU Digital Commons. https://digitalcommons.odu.edu/cgi/viewcontent.cgi?article=1016&context=cepe_proceedings ↩
12. Norbert Wiener and cybernetics (overview context) — “Cybernetics” episode, The Engines of Our Ingenuity. https://engines.egr.uh.edu/episode/2901 ↩
13. Turing, Cybernetics and the Forgotten Histories of AI — Royal Signals. https://royalsignals.org/royal-signals-institution/editorial/turing-cybernetics-and-the-forgotten-histories-of-ai ↩
14. History of artificial intelligence — Wikipedia. https://en.wikipedia.org/wiki/History_of_artificial_intelligence ↩
15. Joseph Weizenbaum, Computer Power and Human Reason (PDF). http://blogs.evergreen.edu/cpat/files/2013/05/Computer-Power-and-Human-Reason.pdf ↩
16. Computer Power and Human Reason — Wikipedia. https://en.wikipedia.org/wiki/Computer_Power_and_Human_Reason ↩
17. Why the Computer Scientist Behind the World's First Chatbot… — Smithsonian Magazine. https://www.smithsonianmag.com/history/why-the-computer-scientist-behind-the-worlds-first-chatbot-dedicated-his-life-to-publicizing-the-threat-posed-by-ai-180987971/ ↩
18. Medical School Admissions: Report of a formal investigation into St. George’s Hospital Medical School (1988) — Wikipedia. https://en.wikipedia.org/wiki/Medical_School_Admissions:Report_of_a_formal_investigation_into_St._George%27s_Hospital_Medical_School(1988) ↩
19. Incident 43: Racist AI behaviour is not a new problem — Incident Database. https://incidentdatabase.ai/cite/43/ ↩
20. St. George's Hospital Medical School racial and sexual discrimination — Eticas Foundation. https://dev.eticasfoundation.org/st-georges-hospital-medical-school-racial-and-sexual-discrimination/ ↩
21. The Evolution of AI: From Rule-Based Systems to Data-Driven Intelligence — ResearchGate. https://www.researchgate.net/publication/388035967_The_Evolution_of_AI_From_Rule-Based_Systems_to_Data-Driven_Intelligence ↩
22. The Evolution Of AI: Transforming The World One Algorithm At A Time — Bernard Marr. https://bernardmarr.com/the-evolution-of-ai-transforming-the-world-one-algorithm-at-a-time/ ↩
23. FATE: Fairness, Accountability, Transparency & Ethics in AI — Microsoft Research. https://www.microsoft.com/en-us/research/theme/fate/ ↩
24. Understanding artificial intelligence ethics and safety — The Alan Turing Institute. https://www.turing.ac.uk/news/publications/understanding-artificial-intelligence-ethics-and-safety ↩
25. Ethics and Responsible Innovation — The Alan Turing Institute. https://www.turing.ac.uk/research/research-programmes/public-policy/public-policy-themes/ethics-and-responsible-innovation ↩
26. Stochastic Parrots — Emily M. Bender. https://faculty.washington.edu/ebender/stochasticparrots.html ↩
27. Stochastic parrots — Language Log. https://languagelog.ldc.upenn.edu/nll/?p=51161 ↩
28. Alignment via Competition: Emergent Alignment from Differently Misaligned Agents — NeurIPS 2025. https://neurips.cc/virtual/2025/129449 ↩
29. NeurIPS 2025 Papers (index) — NeurIPS. https://neurips.cc/virtual/2025/papers.html ↩
30. Can We Break Open AI's Black Box? — Chicago Booth Review (Feb 2026). https://www.chicagobooth.edu/review/2026/february/can-we-break-open-ais-black-box ↩
31. OpenAI disbands mission alignment team — TechCrunch (11 Feb 2026). https://techcrunch.com/2026/02/11/openai-disbands-mission-alignment-team-which-focused-on-safe-and-trustworthy-ai-development/ ↩
32. Exclusive: OpenAI disbanded its mission alignment team — Platformer (11 Feb 2026). https://www.platformer.news/openai-mission-alignment-team-joshua-achiam/ ↩
33. Anthropic Fellows Program for AI safety research: applications open for May & July 2026 — Anthropic Alignment. https://alignment.anthropic.com/2025/anthropic-fellows-program-2026/ ↩
34. AI Act | Shaping Europe’s digital future — European Commission (updated 27 Jan 2026). https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai ↩
35. Red Lines under the EU AI Act… — Future of Privacy Forum (accessed Feb 2026). https://fpf.org/blog/red-lines-under-the-eu-ai-act-understanding-prohibited-ai-practices-and-their-interplay-with-the-gdpr-dsa/ ↩
36. Implementation Timeline — EU Artificial Intelligence Act tracker. https://artificialintelligenceact.eu/implementation-timeline/ ↩
37. Standard Setting overview — EU AI Act tracker. https://artificialintelligenceact.eu/standard-setting-overview/ ↩
38. Update on CEN and CENELEC’s Decision to Accelerate… — CEN-CENELEC (23 Oct 2025). https://www.cencenelec.eu/news-events/news/2025/brief-news/2025-10-23-ai-standardization/ ↩
39. Public Attitudes to AI (key findings dashboard) — attitudestoai.uk. https://attitudestoai.uk/ ↩ ↩
40. Great (public) expectations — Ada Lovelace Institute (4 Dec 2025). https://www.adalovelaceinstitute.org/policy-briefing/great-expectations/ ↩
41. Bill Text — SB-53 — California Legislative Information. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB53 ↩
42. What is California’s AI safety law? — Brookings (23 Dec 2025). https://www.brookings.edu/articles/what-is-californias-ai-safety-law/ ↩
43. Bill Text — AB-316 Artificial intelligence: defenses — California Legislative Information. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB316 ↩
44. California Eliminates the “Autonomous AI” Defense… — Baker Botts (20 Jan 2026). https://ourtake.bakerbotts.com/post/102m29i/california-eliminates-the-autonomous-ai-defense-what-ab-316-means-for-ai-deplo ↩
45. California Governor Newsom Signs Several AI Bills… — Perkins Coie (6 Nov 2025). https://perkinscoie.com/insights/update/california-governor-newsom-signs-several-ai-bills-vetoes-three ↩
46. China Cybersecurity Law Amendment — January 1, 2026 — China Briefing (5 Nov 2025). https://www.china-briefing.com/news/china-cybersecurity-law-amendment/ ↩
47. China Data Laws 2026: Key Changes for Businesses — K. Lea Legal (10 Feb 2026). https://klealegal.com/newsroom/china-data-laws-2026-key-changes ↩
48. From India, Guterres calls for $3 billion fund to ensure AI benefits all — United Nations (Feb 2026). https://www.un.org/sustainabledevelopment/blog/2026/02/from-india-guterres-calls-for-3-billion-fund-to-ensure-ai-benefits-all/ ↩
49. Modi pitches India as an artificial intelligence hub at the AI summit — Associated Press (Feb 2026). https://apnews.com/article/67c2b5a37f98e0a6ebb81136e0287969 ↩