Back
Add Layer
5 Points - 5.0 Point Average
Idea Post

Immutable Ethics Chip

Ai - Someone Could Turn Your Sys... - Immutable Ethics Chip
  • What is this?

  • A proposal to build a fundamentally different kind of AI — one trained from the ground up using Nonviolent Communication (NVC) as its core framework, with a physically immutable "ethics chip" built into photonic (light-based) hardware. This page exists to organize the effort and connect contributors.


  • 1. THE CORE PROBLEM WITH CURRENT AI SAFETY

  • Two hidden frameworks

  • Many people are taught to solve problems by determining who did "good" or "bad", then administering praise or punishment. But there is an alternate framework — more effective, simpler, and probably the original one in human history. If Thag breaks his leg, "Thag need help back to village" is more intuitive and natural than inventing a concept for a word and assigning it to someone.

  • Social labels serve power structures. Solving problems requires assessing needs. These are two completely different frameworks, and programmers — most of whom have never heard of NVC — don't know these frameworks are competing inside their training data.

  • *Why this matters at the perception level

  • Two villages, same situation: someone sees another person eating too much food.

  • **Tribe A (judgment framework):** The observer doesn't want conflict, but they see this person eating up the village's food before everyone returns from collecting water. Eventually they snap. *Thag's perspective: one minute he's enjoying a meal, the next someone's yelling at him. Problems ensue.*

  • **Tribe B (needs framework):** The observer sees Thag eating too much and realizes he would need to know that not everyone is back yet to even understand he's eating too much. *Thag's perspective: one minute you're enjoying a meal, the next someone politely informs you we need to save some for folks who haven't returned. No problem — you actually feel more trust in others.*

  • The observer in both scenarios was trying to look out for other people. **The difference is the process:** need-literacy is simply more effective at conceptualizing problems so peaceful outcomes are likely.

  • What AI companies are doing

  • Every major AI company trains on massive internet data — contaminated with Tribe A logic. This needlessly inflates parameter counts and encodes toxic patterns. "Reinforcement Learning" and behavioral patches applied on top of a model that has already absorbed every destructive communication pattern may be fundamentally doomed. The patterns remain in the weights and resurface under adversarial pressure.

  • **This proposal: don't create the problem in the first place.**

  • To learn more about the AI dataset concept click here: https://needpedia.org/posts/663



  • 2. THE NVC-NATIVE AI CONCEPT

  • *NVC is a formula, not a philosophy

  • Psychology literature doesn't make AI more ethical — it's like making a criminal read 1,000 books on manipulation. Psychology largely collects observations without a unifying theory.

  • NVC provides a **formula** — the OFNR schema:

  • - **Observation:** What is the observable fact? (Not the story about it)
  • - **Feeling:** What emotion does it trigger?
  • - **Need:** What underlying need is driving the feeling?
  • - **Request:** What concrete action would address the need?

  • NVC rules are clear enough that statements can definitively comply or not. There is no comparable formula for determining if a statement is "psychologically correct" — only whether it matches DSM descriptions.

  • "The Source Code of the human mind"

  • NVC's core claim: the human operating system runs on needs the way computers run on instruction sets. All behavior results from perceptions of needs and wants. If this model is more accurate than the judgment/blame model embedded in most internet text, then an AI trained on it may understand people better — not just sound more polite.

  • For AI development: a model trained on NVC needs only the parameters for needs-assessment. When interpreting problems, motivations, and conflict, no contradictory frameworks are competing. **Less noise, more signal.**

  • This enables:

  • - Internal representations of conflict that map to **needs**, not blame or judgment
  • - Adversarial input translated into **needs-assessment** rather than reactive framing
  • - **Zero-shot transfer** across contexts — a corporate board meeting, a diplomatic crisis, or a family dinner — because the underlying human "Source Code" is the same

  • Beyond "Sophisticated Sycophancy"

  • Most "polite" AI operates on a façade. Safety filters create the appearance of empathy — what the field increasingly calls **sophisticated sycophancy**. The AI sounds nice but doesn't structurally understand the causal drivers of the person it's talking to. It performs politeness without doing needs-assessment.

  • A model trained on a **chreogenic** (needs-generated) framework moves past the mask. It doesn't just act polite; it identifies the unmet needs behind the words and provides a functional response rather than a scripted one.

  • Existing research

  • A 2025 MIT/Carnegie Mellon study (Shen et al.) showed AI models consistently misjudge emotional impact and miss harmful patterns requiring relationship context. A 2025 Rutgers/UW study (Wolfe et al., AAAI/ACM AIES) proposed "Needs-Conscious Design" using NVC principles and identified "Empathy Fog" — the uncertainty when AI simulates empathy it lacks.

  • These researchers are reaching toward NVC from the outside. This project works from the inside out.




  • 3. THE TRAINING CURRICULUM

  • **Phase 1 — Jackal-Free Language Acquisition**

  • Train on content inherently free of judgment and blame: mathematics, logic, code, scientific observation, technical documentation, NVC-native transcripts. The model acquires language without encountering judgment or manipulation as valid processing modes.

  • **Phase 2 — Inoculated Exposure**

  • Expose the model to harmful communication narrated through an NVC lens — like a parent explaining a movie's conflict.

  • Example: *"The character says 'you're useless.' This expresses an unmet need for competence and support. In NVC: 'When I see the project results, I feel frustrated because I need to know our work is effective.'"*

  • Result: the model **parses** jackal but **thinks** in NVC. A user saying "you're a stupid machine" and one saying "I'm frustrated because I need clarity" should activate the same need-representation internally. Interpretability tools (Anthropic's circuit tracing, sparse autoencoders, activation patching) can verify this directly.

  • **Why more efficient:** Microsoft's Phi models showed curated data lets a 3.8B model compete with 10× larger ones. Recent Small Language Model breakthroughs show 100M–500M parameter models outperforming models 25× their size when training data is coherent. Removing contradictory frameworks reduces noise; less noise means more efficient parameter usage. An NVC-native 500M model may perform like a 2B model trained on standard internet data — running on standard hardware while out-reasoning massive traditional systems.



  • 4. THE ETHICS CHIP (PHOTONIC HARDWARE)

  • The idea

  • Deploy ethics-critical components on photonic (light-based) hardware with fixed, physically etched weights. Weights literally inscribed in glass or silicon. A remote attacker cannot overwrite them — no memory state to alter. **You cannot install malware into a piece of glass.**


  • Honest architecture

  • A neural network node does two things:

  • 1. **Weighted summing (linear):** Multiply each input by its importance, add them up — gathering evidence.
  • 2. **Activation (nonlinear):** Squash the sum through a curve — making a decision. Without this, the whole network collapses to one linear equation and learns nothing complex.

  • The photonic mesh — fixed beam splitters and directional couplers etched in silicon — handles **Step 1** at light speed with near-zero energy. A weight of 0.55 means 55% of light takes Path A, 45% takes Path B. These weights are physically immutable. They encode *what the system values and how strongly.*

  • **Step 2** is where precision matters. Beam splitters cannot do nonlinear squashing — light is obediently linear. This is the active bottleneck in photonic computing. Two approaches exist:

  • - **Hybrid (buildable today):** Convert light to electricity, apply the nonlinear function in a fixed-function analog circuit, convert back.
  • - **All-optical (research frontier):** Nonlinear optical materials — saturable absorbers, nonlinear crystals, optical cavities with bistable switching. They work in labs but aren't mature for mass chip-scale deployment.

  • This project adopts the **hybrid approach** now, with an all-optical upgrade path as materials mature. The key security insight: the activation function (ReLU, sigmoid, etc.) **has no ethical content.** It's just math. A ReLU passes positive values and zeros negatives — it has no opinions about harm. The ethical values live entirely in the weights. And those are etched in glass.

  • Security model

  • The activation electronics are **fixed-function analog** — no firmware, no memory, no microcontroller. A slab of doped silicon wired for exactly one mathematical operation, about as reprogrammable as a toaster. Tampering can at worst: degrade the signal, inject noise, or break it entirely.

  • What an attacker **cannot** do: change what the system considers harmful. The weights are physical structures. You'd need nanofabrication equipment, not a network connection.

  • Fail-safe architecture

  • The system treats "ethics output missing, corrupted, or inconsistent" as a **hard inhibit.** No action without valid ethics clearance. Modeled on nuclear weapon Permissive Action Links (PALs): the goal isn't "impossible to tamper with" but **"impossible to weaponize without consent."**

  • Physical-access attacker capabilities:
  • - Jam sensors → robot stops
  • - Cut power → robot shuts down
  • - Destroy activation electronics → ethics chip fails → system halts

  • What they cannot do: **make it harm someone.** The fail-safe covers the gap. If the ethics chip is unreadable, the system is inert.

  • This becomes necessary infrastructure as humanoid robots enter mass production. Software ethics controls can be patched out. Etched glass cannot. As air-gap hacking evolves, the distinction between "can be stopped" and "can be weaponized" is the one that matters.

  • Size and feasibility

  • A harm and motive evaluator is a classifier with continuous output: *how much harm does this action cause, weighted across affected parties? Who benefits, who is at risk?* Much simpler than general language generation. A model in the 10M–100M parameter range can handle this — especially with high-quality NVC-annotated training data. This is one of the most achievable parts of the project.

  • The nonlinear frontier

  • Long-term possibility: engineer the activation directly into a material's physical behavior — saturation, absorption edges, cavity interference — the way analog computers historically turned component flaws into computational features. Even a "weird" lumpy activation curve is usable if stable and reproducible; training compensates for odd shapes. The bottleneck isn't finding nonlinearity — it's finding nonlinearity that's stable, temperature-tolerant, and manufacturable at chip scale. The hybrid approach deploys now; all-optical upgrades later.


  • 5. ETHICAL FRAMEWORK: MODULAR AND EXTENSIBLE

  • The ethics evaluation can incorporate multiple validated frameworks rather than any single metric. For example:

  • **Self-Determination Theory (SDT)** — Deci & Ryan's macro-theory of human motivation — identifies three basic psychological needs whose fulfillment drives intrinsic motivation and well-being:

  • - **Autonomy:** The need to feel in control of one's own life and actions, aligned with personal values rather than external pressure.
  • - **Competence:** The need to effectively master one's environment, feel capable, and demonstrate skill.
  • - **Relatedness:** The need to feel connected, understood, and cared for by others.

  • An action that undermines autonomy, competence, or relatedness across affected parties registers as harmful. SDT is empirically grounded (decades of cross-cultural research), operationalizable (needs can be scored from context), and complementary to NVC — NVC provides the communication schema (OFNR), SDT provides the taxonomy of needs being violated or fulfilled.

  • The architecture is deliberately **modular.** A harm evaluation could also incorporate:

  • - Constitutional AI constraints (Anthropic)
  • - Medical ethics principles (beneficence, non-maleficence, autonomy, justice)
  • - Legal compliance bounds
  • - Cultural context parameters

  • The point is not to enshrine one framework but to build a substrate where the evaluation is **structurally needs-based** and the weights encoding those evaluations are **physically immutable.** What specific frameworks plug in is a design choice; that they can't be remotely rewritten is the security property.



  • 6. THE NVC ANNOTATED DATASET

  • No public NVC-annotated AI training dataset exists. This project creates the first.

  • **Contents:**

  • - OFNR-annotated conflict dialogue (Observation, Feeling, Need, Request)
  • - Jackal/NVC translation pairs — same message, both frameworks, with commentary
  • - Narrated fiction: dialogue with NVC analysis
  • - NVC quality scores (0–10)

  • **Production method:** Frontier AI model generates first-pass OFNR annotations on standard text; NVC-certified practitioners review for accuracy. AI handles scale; humans provide quality control.

  • **Cost:** A publishable dataset of approximately 10,000 high-quality annotated samples — roughly **$3,000–$6,000** via this hybrid pipeline. A clear, tangible, and affordable path for researchers and funders.

  • **Distribution via HuggingFace:**

  • - Free sample (~1,000 examples) for visibility and citations
  • - Research license: free for non-commercial use
  • - Commercial license: paid access for companies building empathetic AI (Woebot, Wysa, Replika, and similar)

  • No competition exists in this niche.



  • 7. RELATED RESEARCHERS AND POTENTIAL ALLIES

  • - **Ram Nathaniel** — "RL training for soft skills using NVC" (Medium, March 2025). Explicitly seeking collaborators. [medium.com/@ram.nathaniel](https://medium.com/@ram.nathaniel)

  • - **Robert Wolfe (University of Washington)** — Lead author, "Needs-Conscious Design" (AAAI/ACM AIES 2025). Direct contact; potential arXiv endorser.

  • - **Allison Lahnala, Charles Welch, David Jurgens, Lucie Flek** — Empathy modeling research, EMNLP 2025. Analyze how theoretical groundings of empathy affect NLP performance.

  • - **Yoshua Bengio / MILA** — Chairs International AI Safety Report; advocates for fundamentally different training approaches. [mila.quebec](https://mila.quebec)

  • - **EleutherAI** — Open-source AI research collective. Open Discord. Novel training approaches welcome.

  • - **CNVC (Center for Nonviolent Communication)** — [cnvc.org](https://cnvc.org) — institutional ally for data sourcing, practitioner network, dataset review.

  • - **Lightmatter, Lumai** — Photonic AI chip companies. Commercial contacts once patent and benchmarks exist.


  • 8. POTENTIAL FELLOWSHIPS AND FUNDING

  • | Program | Award | Deadline | Notes |
  • | CBAI Summer Research Fellowship | $10K + housing + $10K compute | April 12, 2026 | Cambridge, MA. Rolling review. |
  • | DARPA CLARA | Up to $2M / 24 months | April 17, 2026 (extended) | High-assurance AI. Modular architecture fits TA1. |
  • | Astra Fellowship (Berkeley) | Stipend + $15K/mo compute | May 3, 2026 | 3–6 months. 80%+ alumni at OpenAI, Anthropic, DeepMind. |
  • | OpenAI Safety Fellowship | Monthly stipend + compute | May 3, 2026 | Sep 2026–Feb 2027. References required. |
  • | Schmidt Sciences — Trustworthy AI | Up to $5M | May 17, 2026 | AI risk research. |
  • | Schmidt Sciences — AI Interpretability | $300K–$1M | May 26, 2026 | NVC activation verification fits directly. |
  • | Anthropic Fellows Program | $3,850/week + $15K/mo compute | May & July 2026 | No PhD required. 40%+ hired full-time. |
  • | Cambridge ERA:AI Fellowship | £34,125/year equiv. (10 weeks) | Apply early | July 6 start. |
  • | Long-Term Future Fund | $5K–$100K | Rolling | Fast turnaround. |
  • | Manifund | Variable | Rolling | Regranting platform. |
  • | Coefficient Giving (Open Philanthropy) | $50K–$5M | Rolling | Technical AI safety RFP. |
  • | Center on Long-Term Risk Fund | Variable | Rolling | Worst-case AI suffering risks. |
  • | Survival and Flourishing Fund | Variable | April 22, 2026 | Mid to large scale. |
  • | Foresight Institute AI Nodes | ~$100K/project | Monthly review | Compute + office space. |
  • | FLI Digital Media Accelerator | Creator grants | Rolling | Blog/video series on NVC+AI. |
See more
Tags: #AI Photonics NVC Public Safety Research
0 flags
0 likes
0 shares
Tony Brasher

1. Build the NVC-annotated training dataset Use a frontier AI model as a distiller to generate jackal/NVC translation pairs, need-identification annotations, and NVC quality scores. Recruit CNVC-affiliated practitioners to review for accuracy. Publish a free sample on HuggingFace to establish credibility and attract collaborators.

  • 3 months ago
Tony Brasher

2. Apply for near-term funding Prioritize the highest-urgency deadlines first: CBAI Summer Research Fellowship (April 12), DARPA CLARA (April 17), and the Astra and OpenAI Safety Fellowships (May 3). Use the dataset and this post as the basis for applications.

  • 3 months ago
Tony Brasher

3. Contact and recruit research collaborators Reach out to Ram Nathaniel, Robert Wolfe (UW), and the Lahnala/Welch/Jurgens/Flek group, all of whom are already working at the NVC-AI intersection. Approach EleutherAI's Discord community and CNVC as institutional allies.

  • 3 months ago
Tony Brasher

4. Train Phase 1 and 2 models First train a model exclusively on jackal-free data (math, logic, code, NVC transcripts). Then run the inoculated exposure phase, narrating harmful communication through an NVC lens. Use mechanistic interpretability tools — sparse autoencoders, activation patching, circuit tracing — to verify that harmful inputs are... See more

  • 3 months ago
Tony Brasher

5. Benchmark the model Use existing AI safety evals (Anthropic's and others') to measure harm avoidance, needs-assessment accuracy, and resistance to adversarial prompting. Compare parameter efficiency against standard models trained on unfiltered internet data.

  • 3 months ago

Resource Tags:

Edit Resource Tags (separated by commas)


Cancel