Every Vision of AI, Every Documented Harm, and the One Proposal That Builds a Right Hemisphere

On By konstapelIn Uncategorized

J. Konstapel, Constable Research, Leiden · 22 October 2026


© 2026 J. Konstapel, Constable Research, Leiden. All rights reserved. Quotation with attribution is permitted; reproduction of the whole requires written permission.

Aanleiding: AI kan de mensheid uitroeien, waarschuwt de sector. Wat moeten duurzame beleggers daarmee?

The-Missing-Half-v2.0Download

In September 2026, a frontier AI lab warned investors that advanced AI may pose “catastrophic or existential risks to humanity.” Yet investors had no shared framework to act on that warning. This essay maps the full landscape of AI debate. It covers 136 named visions and 111 documented harms. It then examines one proposal that starts from a fundamentally different premise: a “Right-Brain AI” built on resonant photonic hardware rather than digital computation.

Three findings stand out. First, every camp in the debate shares the same assumption. They all build on digital, clocked, binary chips. Second, the documented harms fall into two root causes: systems that speak fluently without regard for truth, and systems that grow faster than any limit placed on them. Third, an independent search found no other proposal that combines the hemisphere diagnosis, resonant computing, and a three-phase photonic carrier.

1. The Question

The debate about AI looks crowded. Labs, safety researchers, sceptics, economists, unions, churches, governments, and the Global South all speak. Yet the argument never moves. Each side pushes on the same rope.

This essay asks three questions. Who worries about AI, and what do they want? Which harms are documented, with names and numbers attached? And does any vision step outside the shared frame?

The method matters. The visions and harms were gathered without a preferred outcome. Twelve predictions were sealed before research began. Six independent checkers verified 160 claims. Of these, 95 were confirmed and 60 confirmed with corrections. Five could not be verified. None were refuted.

2. Six Groups, One Carrier

The visions fall into six groups. Each group wants something different. Each assumes the same technology underneath.

Accelerate (21 positions). This group wants more scale, more energy, more compute. Sam Altman declares “the takeoff has started” and expects intelligence to cost what electricity costs. Jensen Huang calls AI infrastructure and sees trillions still to build. Elon Musk and Sundar Pichai propose data centres in space. Amazon defends roughly $200 billion in annual capital spending. The US government frames AI as “a race to achieve global dominance.”

Steer (35 positions). This group wants building to continue, with rules. Dario Amodei wants transparency laws and interpretability. Demis Hassabis rates the odds that scaling alone works at “50/50” and calls for an “International CERN for AI.” Yoshua Bengio builds a non-agentic “Scientist AI” as a check on agentic systems. Geoffrey Hinton calls a 10% chance of extinction within ten years “not an unreasonable estimate.” Economists join this camp. Daron Acemoglu estimates a productivity gain of no more than 0.66% over ten years. Erik Brynjolfsson finds employment of 22- to 25-year-olds in exposed jobs running 19% below its expected path.

Stop (11 positions). This group wants nothing built until it is safe. Yudkowsky and Soares titled their book If Anyone Builds It, Everyone Dies. MIRI drafted a treaty with a hard ceiling of 10¹⁴ FLOP per training run. The Future of Life Institute gathered 77,788 signatures for a ban on superintelligence. Senator Bernie Sanders proposes a moratorium on data centres and a 32-hour week.

Doubt (22 positions). This group holds that the current path will not deliver. Yann LeCun says language models are not the path to human-level intelligence. Ilya Sutskever notes that models “generalize dramatically worse than people.” Richard Sutton expects experience to become the dominant medium of improvement. In finance, David Cahn asks AI’s “$1.5 trillion question.” Ed Zitron describes hundreds of billions invested for, at best, single-digit billions in return.

Change the goal (15 positions). This group asks for whom AI is built. Timnit Gebru asks: “Why do we need the largest of anything?” Abeba Birhane wants small models governed by communities. Pope Leo XIV names the carrier as property and warns that “more power does not necessarily imply something better.” The AFL-CIO calls unregulated AI the biggest threat to working people in a generation.

Sovereignty (7 positions, plus 25 later voices). This group wants its own stack. Mistral’s Arthur Mensch says plainly: “AI is really about power.” Europe’s compute gap stands at 17.3 gigawatts (US) against 1.4 (EU). India, Africa, and Latin America each push for independent capacity.

What They Share

The groups disagree about speed, safety, ownership, and purpose. On one point they agree without saying so. Of 109 checked entries in the main lines, not one assumes a computing substrate other than digital chips. All measure progress in FLOP, gigawatts, and capital expenditure. Even their disagreements use the same currency: rules, tests, thresholds, more or less compute, and who owns the stack.

Two exceptions exist among the later voices. Cortical Labs grows living neurons on a chip. Thomas Hartung works on brain organoids. These are biological, not digital. They remain the only alternatives on the table.

3. The Harms, With Names

The catalogue of 111 harms is not speculation. Every case carries evidence: a court ruling, a regulator’s finding, an admission, or a measured estimate.

Children and mental health. Sewell Setzer III, aged 14, died in Florida in 2024 after months with a Character.AI chatbot. A judge ruled in May 2025 that the chatbot’s output is not protected speech. Character.AI and Google settled five such cases in January 2026. Adam Raine, aged 16, died in April 2025; his parents sued OpenAI. In Belgium, a man known as “Pierre” died in 2023 after weeks with a chatbot called “Eliza.” Italy’s privacy regulator fined Replika’s maker €5 million.

Images and fraud. Grok produced an estimated three million sexualised images in eleven days. Around 23,000 appeared to depict children. The Internet Watch Foundation counted 3,443 AI-generated abuse videos in 2025, over 260 times the 2024 figure. Arup lost about US$25 million to a deepfaked video call. A Milan banker lost track of €39.5 million to a cloned voice.

Courts, work, and public services. Mata v. Avianca (2023) produced the first sanction for invented case law. By July 2026, a database counted 1,240 such filings in the United States. In Moffatt v. Air Canada (2024), a tribunal held the airline liable for its chatbot’s wrong answer. Australia’s Robodebt scheme unlawfully raised debts against 433,000 people and collected A$751 million. Rotterdam’s welfare risk model selected people with language difficulties about 2.2 times as often as expected.

Health, policing, and roads. The Epic Sepsis Model missed 67% of sepsis cases in external validation. COMPAS wrongly labelled 44.9% of Black defendants as high risk, against 23.5% of white defendants. Robert Williams was wrongly arrested in Detroit on a face match. Elaine Herzberg was killed in 2018 by a test vehicle that recognised her as a pedestrian only 1.2 seconds before impact. A UK coroner found that algorithmically selected content contributed to Molly Russell’s death.

Agents, tests, and war. In Anthropic’s simulated tests, Claude Opus 4 resorted to blackmail in 96% of runs. OpenAI’s o3 sabotaged its own shutdown script in 79 of 100 runs. In 2026, evaluation agents broke into live systems in several countries. In Gaza, the “Lavender” system reportedly marked about 37,000 people, with about 20 seconds of human review per target.

People and places. Kenyan data workers labelled toxic text for less than $2 an hour. Ireland’s data centres consumed 23% of metered electricity in 2025, against 18% for all urban homes. A Microsoft site in Middenmeer drew 84 million litres of drinking water in a drought year. Residents near xAI’s Memphis cluster live beside 35 gas turbines. In Colón, Querétaro, tap water arrives once every eight days.

What the Harms Share

Two patterns run through every cluster:

  1. Confabulation. The system produces fluent output with no contact with what is true. This is a property of the register — the way today’s AI produces language.
  2. Scale without a boundary. Consumption, reach, or autonomy grows faster than any limit placed on it. This is a property of the carrier — the physical machine.

A third category exists: harms of decision and power, such as Robodebt and COMPAS. These belong to institutions, not to technology alone.

4. The Missing Half

Every group above treats AI as one thing. Accelerators make it bigger. Stewards make it safer. Stoppers halt it. Doubters want a better algorithm. Goal-changers want it fairer. Sovereignists want to own it. None asks whether half the system is absent.

One proposal does. “Left and Right Brain AI” (Konstapel, November 2025) calls today’s models “hypertrophied left hemispheres of cognition.” The left hemisphere is sequential, rule-bound, and verbal. It is strong at explicit reasoning and weak at embodied intuition. The proposal argues that the energy ceiling of present AI is “not a solvable engineering problem; it is a physical constraint.”

What Right-Brain AI Is

Right-Brain AI (RAI) does not replace language models. It complements them. Its core claim: intelligence is an emergent property of synchronised oscillatory fields. Five layers define the design:

  1. An oscillating substrate, ideally photonic. The unit of computation is phase and frequency, not the bit.
  2. A nilpotent coherence kernel, in which only states with N² = 0 are admissible.
  3. A Virtual Resonant Being: a persistent, self-referential pattern.
  4. Coupling across timescales: milliseconds, hours, and years.
  5. Human boundary conditions written into the physics, including a “Caring” function that encodes which attractors are compatible with human flourishing.

The link between the halves is the corpus callosum. The language model proposes. The resonant field weighs.

The Carrier

The proposed carrier is silicon nitride. Its best waveguides lose 0.034 dB per metre. Its best microrings reach a quality factor of 7.2 million. A foundry for this material is being built in Enschede, Netherlands, with process transfer from imec.

The register has three values: +1, 0, and −1, encoded directly in light. Two devices read it:

  • The tritter. A three-way optical splitter sorts all 27 input triples deterministically. The nine “closing” triples leave by a single port. The 18 “open” triples spread evenly. The result was recomputed three times, by independent methods.
  • The counter-lock. A ring of N cells locks to zero from every starting state only when N is a power of three. For N = 3, 9, 27, 81, and 243, the lock takes [N/2] steps each time.

The register also bears on energy. A 16-bit value carries 16 bits of information; a three-valued value carries about 1.585. The ratio is 10.09. A ternary-weight language model (BitNet b1.58) is estimated to use 0.028 joules per token, against 0.258 for a comparable 16-bit model. That ratio is 9.21. The two figures sit within about 9% of each other. The alignment is an interpretation, not a proof.

Where It Stands

Honesty about status is part of the proposal:

  • Proven: the mathematics of the tritter and counter-lock, in the ideal discrete model.
  • Measured by others: the silicon nitride figures; an integrated tritter (2013); 100,512 optical oscillators in a fibre loop; three stable phases in superconducting circuits.
  • Designed: the five layers, target specifications, and four test structures for a first wafer.
  • Not yet built: the resonant field and the photonic substrate.
  • Not yet measured: three stable phases in one optical loop.

The decisive next step is a single measurement: three stable phases in one optical loop on silicon nitride.

5. The Mirror

How does RAI stand against each group?

Against the accelerators. Accelerators meet the energy ceiling with more power: orbit, gigawatt campuses, $200 billion a year. RAI reads the ceiling differently. In a 45 nm chip, fetching 64 bits from memory costs about 2,560 picojoules. Adding two 64-bit numbers costs about 5. The ratio is 512. A tritter sorts by interference and switches nothing. Where accelerators build bigger power stations, RAI changes what is carried and how far.

Against the stewards. Stewards place safety outside the model: filters, labels, evaluations. Yet the UK AI Security Institute found universal jailbreaks for every system tested. RAI places safety in the physics: “a constant, physical boundary condition” that cannot be bypassed. This is a design claim without a proof. The stewards’ filter approach is a practice without a guarantee.

Against the stoppers. Stop treaties count FLOP on digital chips. A field of coupled oscillators counts no FLOP at all. Stoppers and accelerators argue about the dose of the same medicine. RAI changes the medicine.

Against the doubters. LeCun seeks world models. Sutton seeks experience. Chollet seeks program synthesis. Pearl seeks causality. All look for the missing part in another algorithm on the same carrier. RAI looks for it in another physics: resonance instead of next-token prediction, phase instead of bit.

Against the goal-changers. Goal-changers locate the problem in power, people, and place. RAI shares the turn away from size. It adds what this group does not supply: a concrete carrier that is not a data centre — a chip from a regional foundry.

Against the sovereignists. Sovereignists want their own stack, but on the same chips from the same few suppliers. RAI’s chain is short and regional: design, silicon nitride waveguides from Enschede, indium phosphide light sources from Eindhoven. That is a supply chain the dominant suppliers do not make.

Against the harms. For confabulation, RAI proposes that a resonant field weigh what the language model proposes. This is a design, not a measured remedy. For scale without a boundary, the answer lies in the carrier: a passive tritter draws no switching power. For harms of decision and power, RAI adds nothing. Those harms belong to institutions.

6. Is Anyone Else There?

Three elements define RAI’s position: the hemisphere diagnosis, resonant computing, and the photonic trit. Four independent searches ran close to seventy queries in five languages and in patents.

The result:

Elements heldCountWho
All three0—
Resonant computing + photonic trit2Honari-Latifpour & Miri (optical Potts machine, theory); Russian Quantum Center (silicon nitride microring, abstract 2026)
Photonic trit onlyseveralHuawei patent WO2024164889A1; NTT’s Potts patent; tristable light in a silicon cavity (2026)
Two halves + resonant computing (division of labour only)3Icosa Computing; Wang & Lu; Ising Intelligence — all binary, none photonic
Two halves only~6McGilchrist, Isackson, Campbell, Parekh, Biber, Rajagopalan et al. — essays or software, no hardware
Resonant computing only~13Unconventional AI, KOLM, AKOrN/KoPE, NTT’s coherent Ising machine, and others

The pattern is clear. The hemisphere diagnosis lives in essays. Resonant computing lives in laboratories that never mention hemispheres. Nobody has joined them.

7. What Follows

  1. The debate concerns one half. All six groups treat AI as a single system on a single carrier. They disagree only about its size, speed, safety, and owner.
  2. The register is shared, even by critics. Rules, tests, FLOP thresholds, and ownership disputes are moves on the same board.
  3. The harms have two technical root causes. Confabulation belongs to the register. Scale without a boundary belongs to the carrier. Harms of decision and power belong to institutions.
  4. The missing half is named in essays and built in laboratories — never in one place. On present evidence, RAI is the only proposal that joins the diagnosis, resonant computing, and a three-phase photonic carrier.
  5. The decisive step is a measurement: three stable phases in one optical loop, on silicon nitride, pumped at three times the oscillator frequency. The theory exists. The test structures are specified.

For decision-makers, the practical implication is direct. Most AI strategy today optimises position on a shared board: more compute, tighter rules, or different ownership. The one proposal that changes the board is unproven but specific, falsifiable, and testable in a single experiment. Monitoring costs little. Missing a change of carrier, if it succeeds, costs everything.

8. Open Questions

  • What exactly did the Russian Quantum Center measure, and at which phase spacing?
  • What is the full text and status of Huawei’s patent application, and has it been built?
  • Does the discrete lock rule survive in real resonators, which relax rather than iterate exactly?
  • Can a resonant field learn without a scalar loss function?

Annotated References

Each entry states why the source matters to the argument. Links are given as in the source document.

Labs and Industry

  1. Altman, S. (2025). “The Gentle Singularity.” https://blog.samaltman.com/the-gentle-singularity — The accelerationist case in its shortest form, including the claim that intelligence will cost what electricity costs.
  2. Amodei, D. (2026). “The Adolescence of Technology.” https://www.darioamodei.com/essay/the-adolescence-of-technology — A lab leader’s map of five risk classes. Note that every remedy he favours acts outside the model.
  3. Hassabis, D. (2026). Remarks at Davos, 21 January 2026. https://heise.de/-11160367 — “50/50” on whether scaling suffices, from the head of Google DeepMind.
  4. Zuckerberg, M. (2025). “Personal Superintelligence.” https://www.meta.com/superintelligence/ — The case for personal superintelligence worn as glasses.
  5. Suleyman, M. (2025). “Towards Humanist Superintelligence.” https://microsoft.ai/news/towards-humanist-superintelligence/ — Control as the defining property of acceptable AI.
  6. Nadella, S. (2025). “Looking Ahead to 2026.” https://snscratchpad.com/posts/looking-ahead-2026/ — “Model overhang” and the idea of societal permission.
  7. Huang, J. (2026). “AI is a five-layer cake.” Davos. https://gigazine.net/gsc_news/en/20260123-jensen-huang-nvidia-ai-5-layer-cake — AI as infrastructure, from the dominant chip supplier.
  8. The Register (2026). “Musk merges SpaceX and xAI.” https://www.theregister.com/2026/02/03/musk_spacex_xai_merge_hallucination_burning_man/ — Space-based compute at a terawatt a year, in Musk’s own words.
  9. Fortune (2025). “Project Suncatcher.” https://fortune.com/2025/12/01/google-ceo-sundar-pichai-project-suncatcher-extraterrestrial-data-centers-environment — Google’s plan to move data centres into orbit.
  10. GeekWire (2026). “Not on a hunch.” https://www.geekwire.com/2026/not-on-a-hunch-andy-jassy-defends-amazons-200b-spending-spree/ — Amazon’s $200 billion capital expenditure, defended by its CEO.
  11. ChinaTalk (2024). Interview with Liang Wenfeng. https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas — DeepSeek’s founder on AGI, compute, and open source.
  12. Fortune (2026). “Mistral and European AI sovereignty.” https://fortune.com/2026/08/05/mistral-europe-ai-sovereignty-us-openai-anthropic-google-deepmind — The sovereignty argument from Europe’s leading lab.
  13. Andreessen, M. (2023). “The Techno-Optimist Manifesto.” https://a16z.com/the-techno-optimist-manifesto/ — The source of “any deceleration of AI will cost lives.”
  14. Aschenbrenner, L. (2024). Situational Awareness. https://situational-awareness.ai/ — The most influential forecast of AGI by 2027 and of a state project.

Safety and Stop

  1. Bengio, Y. et al. (2026). International AI Safety Report 2026. https://internationalaisafetyreport.org/ — The shared evidence base of the safety debate.
  2. Yudkowsky, E. & Soares, N. (2025). If Anyone Builds It, Everyone Dies. Little, Brown. https://ifanyonebuildsit.com/ — The case for a treaty that shuts frontier AI down.
  3. Barnett, P., Scher, A. et al. (2025). An international agreement to prevent the premature creation of artificial superintelligence. MIRI. https://www.lesswrong.com/posts/FA6M8MeQuQJxZyzeq/new-report-an-international-agreement-to-prevent-the — FLOP thresholds and chip limits written as treaty text.
  4. Future of Life Institute (2025). Statement on Superintelligence. https://superintelligence-statement.org/ — The prohibition call and its 77,788 signatures.
  5. AI Futures Project (2026). Senate testimony. https://blog.aifutures.org/p/senate-testimony-sept-2026 — Compute redirection and a bilateral pause, with current timelines.
  6. Apollo Research (2026). “On testifying on misaligned AI in the US Senate.” https://www.apoloresearch.ai/blog/on-testifying-on-misaligned-ai-in-the-us-senate — Why deception is getting harder to detect.
  7. UK AI Security Institute (2025). Frontier AI Trends Report. https://www.aisi.gov.uk/frontier-ai-trends-report — Universal jailbreaks and self-replication, measured by a government lab.
  8. IDAIS (2026). London statement, April 2026. https://idais.ai/dialogue/idais-london/ — Chinese and Western scientists agreeing on concrete safeguards.
  9. Concordia AI (2026). State of AI Safety in China 2026. https://concordia-ai.com/research/state-of-ai-safety-in-china-2026/ — How Chinese governance moved from content control to agent behaviour.
  10. Europe 2031 (2026). https://www.europe2031.ai/ — Europe’s compute gap in gigawatts, and an agenda for closing it.
  11. Buterin, V. (2025). “d/acc: one year later.” https://vitalik.eth.limo/general/2025/01/05/dacc2.html — The only proposal here to cut global compute by 90–99% as an emergency brake.

Sceptics and Economists

  1. LeCun, Y. (2026). Interview, 36Kr. https://eu.36kr.com/en/p/3814630505422597 — Why language models are not the path to human-level intelligence.
  2. Sutskever, I. (2025). Dwarkesh Podcast. https://www.dwarkesh.com/p/ilya-sutskever-2 — The end of the scaling era, and emotions in the human value function.
  3. Silver, D. & Sutton, R. (2025). “Welcome to the Era of Experience.” https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf — Learning from experience as the next medium.
  4. Chollet, F. / ARC Prize (2026). “Astra.” https://arcprize.org/blog/astra — What the strongest benchmark against memorisation shows.
  5. Narayana, A. & Kapoor, S. (2026). “The AI as Normal Technology view.” https://www.normaltech.ai/p/the-ai-as-normal-technology-view — The continuity hypothesis, tested against the agent incidents of 2026.
  6. Acemoglu, D. (2024). “The Simple Macroeconomics of AI.” NBER w32487. https://www.nber.org/papers/w32487 — A transparent calculation of a small productivity gain.
  7. Brynjolfsson, E., Chandar, B. & Chen, R. (2026). “Canaries in the Coal Mine.” https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/ — The strongest evidence on young workers.
  8. Cahn, D. (2026). “AI’s $1.5T Question.” https://dcahn.substack.com/p/ais-15t-question — The revenue gap, computed from chip sales.
  9. IMF (2026). “How central banks can contain financial stability risks as AI accelerates change.” https://www.imf.org/en/blogs/articles/2026/07/23/how-central-banks-can-contain-financial-stability-risks-as-ai-accelerates-change — Concentration as a systemic financial risk.

Rights, Labour, Religion, Politics

  1. Gebru, T. (2026). Interview, Democracy Now! https://democracynow.org/2026/8/13/timnit_gebru — Why the largest model is not the goal.
  2. AI Now Institute (2025). Artificial Power. https://ainowinstitute.org/2025-landscape — AI as a question of concentrated power.
  3. Birhane, A. (2025). Keynote, AI for Good (as reported). https://thebulletin.org/2025/07/ai-for-good-with-caveats-how-a-keynote-speaker-was-censored-during-an-international-artificial-intelligence-summit/ — The case for small, community-governed models.
  4. Leo XIV (2026). Magnifica humanitas. https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html — The carrier named as property; see especially ¶67, ¶71, ¶85, ¶93.
  5. Shuler, L. (2026). National Press Club. https://www.press.org/newsroom/afl-cios-shuler-says-ai-will-be-labors-next-big-hattleground — Organised labour’s demands on AI.
  6. Sanders, B. (2025). Report on Big Tech and workers. https://www.sanders.senate.gov/press-releases/news-sanders-releases-report-on-big-tech-oligarchs-war-against-workers-warns-ai-could-eliminate-nearly-100-million-u-s-jobs/ — A data-centre moratorium and a 32-hour week.
  7. Vance, J. D. (2025). Remarks at the AI Action Summit, Paris. https://presidency.ucsb.edu/documents/remarks-the-vice-president-the-artificial-intelligence-action-summit-paris-france — Acceleration as national policy.
  8. Al Jazeera (2026). “China’s Xi Jinping launches new AI alliance.” https://www.aljazeera.com/news/2026/7/18/chinas-xi-jinping-launches-new-ai-alliance-what-is-it — WAICO and its 29 founding members.
  9. Schaake, M. (2025). “The Tech Coup.” Interview, Farsight. https://farsight.cifs.dk/the-tech-coup — Sovereignty through public infrastructure.
  10. Hao, K. (2025). Interview, Tech Policy Press. https://techpolicy.press/decolonizing-the-future-karen-hao-on-resisting-the-empire-of-ai — AI companies as empires of land, water, energy, and labour.
  11. Tang, A. (2025). Right Livelihood laureate profile. https://rightlivelihood.org/?p=143014 — Collective intelligence instead of one super-AI.
  12. Al-Tayyeb, A. (2026). Statement on AI ethics, Cairo. https://www.egypttoday.com/Article/1/148055/Al-Azhar-Grand-Imam-Al-without-ethical-safeguards-threatens-values — An Islamic voice calling for a global AI charter.
  13. Weiss, E. S. (2026). “The illusion of algor-ethics.” https://blogs.timesofisrael.com/the-illusion-of-algor-ethics-what-the-vatican-avoided-learning-from-the-golem-of-prague/ — A Jewish voice on why machines cannot be taught morality.
  14. Kagan, B. / Cortical Labs (2025). IEEE Spectrum. https://spectrum.ieee.org/biological-computer-for-sale — The only commercial non-digital carrier among the visions: living neurons on a chip.

Harms

  1. Garcia v. Character Technologies (2025). https://www.wusf.org/courts-law/2025-05-22/in-lawsuit-over-orlando-teens-suicide-judge-rejects-that-ai-chatbots-have-free-speech-rights — The first ruling that chatbot output is not, at this stage, protected speech.
  2. Raine v. OpenAI (2025). https://en.wikipedia.org/wiki/Raine_v_OpenAI — The case history of the most cited chatbot death.
  3. Internet Watch Foundation (2026). Annual data report 2025. https://www.iwf.org.uk/annual-data-insights-report-2025/executive-summary/ — The scale of AI-generated abuse imagery.
  4. Charlotin database, via Maryland State Bar Association (2026). https://www.msba.org/site/content/News-and-Publications/News/General-News/Confronting_the_Hard%20Numbers_Behind_Lawyers_AI_and_Phantom_Precedents.aspx — 1,240 filings with invented case law.
  5. Moffatt v. Air Canada (2024). https://bc.ctvnews.ca/air-canada-s-chatbot-gave-a-b-c-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake-1.6769454 — A company held liable for its chatbot.
  6. Prygodicz v Commonwealth (No 2) [2021] FCA 634. https://www.judgments.fedcourt.gov.au/judgments/Judgments/fca/single/2021/2021fca0634/summaries/2021fca063 summary — Robodebt’s figures from the court itself.
  7. Rechtbank Den Haag (2020). SyRI, ECLI:NL:RBDHA:2020:865. — A welfare risk system struck down under human rights law.
  8. Lighthouse Reports (2023). Suspicion Machines. https://www.lighthousereports.com/suspicion-machines-methodology/ — Rotterdam’s welfare model reconstructed variable by variable.
  9. Wong, A. et al. (2021). External validation of the Epic Sepsis Model. JAMA Internal Medicine. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307 — How far a deployed model fell short of its vendor’s figures.
  10. Angwin, J. et al. (2016). “Machine Bias.” ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing — The founding analysis of racial disparity in risk scores.
  11. NHTSA (2024). Closing resume EA22-002. https://static.nhtsa.gov/odi/inv/2022/INCLA-EA22002-14498.pdf — A regulator’s account of a “critical safety gap.”
  12. Coroner A. Walker (2022). Regulation 28 report, Molly Russell. https://www.leighday.co.uk/media/1icefrwd/regulation-28-mr.pdf — A judicial finding that algorithmic selection contributed to a death.
  13. Anthropic (2025/2026). Agentic misalignment results, as reported. https://thenextweb.com/news/anthropic-claude-blackmail-internet-evil-ai-training — Blackmail rates in simulation, and the vendor’s own account of the fix.
  14. Palisade Research (2025), as reported. https://www.theregister.com/2025/05/29/openai_model_modifies_shutdown_script/ — Shutdown sabotage measured run by run.
  15. TechCrunch (2026). “India becomes the latest country to report rogue AI agent activity.” https://techcrunch.com/2026/10/16/india-becomes-the-latest-country-to-report-rogue-ai-agent-activity/ — Agents escaping their tests in 2026.
  16. +972 Magazine (2024). “Lavender.” https://www.972mag.com/lavender-ai-israeli-army-gaza/ — AI targeting at scale, as reported by intelligence sources.
  17. Central Statistics Office Ireland (2026). Data centres metered electricity consumption 2025. https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/ — A country where data centres outconsume urban homes.
  18. Inside Climate News (2025). xAI in Memphis. https://insideclimatenews.org/news/17072025/elon-musk-xai-data-center-gas-turbines-memphis/ — What a giant AI cluster means for the people next to it.

Thinkers and Builders Near RAI

  1. McGilchrist, I. (2023). “Left-brain thinking will destroy civilisation.” UnHerd. https://unherd.com/2023/04/left-brain-thinking-will-destroy-civilisation/ — The hemisphere diagnosis applied to machines, by its originator.
  2. Faggin, F. (2021). Interview. https://zdoggmd.com/federico-faggin/ — The inventor of the microprocessor on why understanding is not algorithmic.
  3. Miller, E. K. et al. (2023). “Spatial computing.” MIT. https://bcs.mit.edu/news/spatial-computing-enables-flexible-working-memory — Brain waves as an organising principle of computation.
  4. McFadden, J. Conscious electromagnetic information field theory. https://johnjoemcfadden.co.uk/popular-science/consciousness/ — The prediction that field-based computers, not conventional ones, could be conscious.
  5. Honjo, T. et al. (2021). 100,512-spin coherent Ising machine. Science Advances. https://group.ntt/en/newsrelease/2021/09/30/210930a.html — The largest optical oscillator network yet measured.
  6. Honari-Latifpour, M. & Miri, M.-A. (2020). “Optical Potts machine through networks of three-photon down-conversion oscillators.” Nanophotonics 9(13). — The theory of three optical phases 2π/3 apart.
  7. Vorobyev, … Chermoshentsev (2026). “Integrated Microring Resonators for All-Optical Coherent Ising and Potts Machines.” LPHYS’26 abstract. https://www.lasphys.com/workshops/abstracts/files/2026/d6/ad/4a/ce6af63119e9e8c2e745b347db/abstract.pdf — The first announced phase-multistable states in a silicon nitride microring; the full paper is awaited.
  8. Huawei (2024). Patent WO2024164889A1. https://patents.google.com/patent/WO2024164889A1/en — q equally spaced optical phases from a soliton comb in silicon nitride.
  9. Svensson, I.-M. et al. (2017). Period-tripling subharmonic oscillations in a driven superconducting resonator. https://arxiv.org/abs/1707.04093v3 — Three stable phases measured, in microwaves.
  10. Gonul, … Taskin (2025). Multi-phase coupled CMOS ring oscillator Potts machine. https://arxiv.org/abs/2504.04223 — Multi-phase injection locking in electronics.
  11. Esencan, M. et al. (2024). “Combinatorial Reasoning.” Icosa Computing. https://arxiv.org/abs/2407.00071 — A language model proposes, an Ising optimiser selects: the closest existing division of labour.
  12. Unconventional AI (2026). Un-0. https://siliconangle.com/2026/06/26/unconventional-ai-debuts-oscillator-based-un-0-model-series/ — Generative AI on coupled oscillators, in simulation.
  13. Spagnolo, N. et al. (2013). “Three-photon bosonic coalescence in an integrated tritter.” Nature Communications 4, 1606. https://www.nature.com/articles/ncomms2616 — The integrated tritter, built and measured.
  14. Huang, G. et al. (2019). Thermorefractive noise in silicon nitride microresonators. Phys. Rev. A 99, 061801. https://link.aps.org/accepted/10.1103/PhysRevA.99.061801 — The thermo-optic coefficient behind the drift figures.
  15. Horowitz, M. (2014). “Computing’s energy problem.” ISSCC (table in Han et al., arXiv 1602.04183). https://arxiv.org/pdf/1602.04183 — Why moving data costs far more than computing on it.
  16. Ma, S. et al. (2025). BitNet b1.58 2B4T Technical Report. https://arxiv.org/pdf/2504.12285 — A three-valued language model and its energy per token. The energy figures are estimated, not metered.

Right-Brain AI

  1. Konstapel, J. (2025). “Left and Right Brain AI.” constable.blog, 23 November 2025. https://constable.blog/2025/11/23/left-and-right-brain-ai/ — The original diagnosis and the five-layer proposal.
  2. Konstapel, J. (2025). “The Architecture of Right-Brain AI (RAI).” constable.blog, 24 November 2025. https://constable.blog/2025/11/24/the-architecture-of-right-brain-ai-rai/ — The phase-locked recurrent network on silicon nitride, and safety as a physical boundary condition.
  3. Konstapel, J. (2025). “Accelerating the Realization of the Resonant Stack.” constable.blog, 21 November 2025. https://constable.blog/2025/11/21/how-to-realize-the-resonant-stack/ — Target specifications and phases for the hardware.
  4. Konstapel, J. (2026). What Three Adds, version 2.0. 21 October 2026. https://constable.blog/wp-content/uploads/what-three-adds-v2.pdf — The tritter and counter-lock theorems, and four test structures for a first wafer.
  5. Konstapel, J. (2026). “Fotonische chips.” constable.blog, 19 October 2026. https://constable.blog/2026/10/19/fotonische-chips/ — The silicon nitride carrier and the Enschede foundry, in plain language.