No publicly available NVC-annotated AI training dataset exists. This post

explains why that's a problem, what such a dataset would look like, and how

to build it.

--------------------------------------------------------------------------------

THE GAP IN AI ALIGNMENT

--------------------------------------------------------------------------------

AI systems are increasingly asked to navigate emotional conversations,

mediate conflict, and support people through difficult decisions. They do this

without ever having been trained on data that explicitly maps human language

to human needs.

The result is a structural blindspot. Current AI systems can simulate empathy

without modeling it. They can generate validating responses without

understanding what the person actually needs. Researchers are calling the

failure mode that results "sophisticated sycophancy" — AI that optimizes for

approval rather than wellbeing, telling people what they want to hear in ways

that are hard to detect precisely because the output is often technically

accurate.

The fix requires a different kind of training data.

--------------------------------------------------------------------------------

WHAT IS NVC, AND WHY DOES IT MATTER FOR AI?

--------------------------------------------------------------------------------

Nonviolent Communication (NVC), developed by Marshall Rosenberg, is usually

treated as a therapeutic or interpersonal tool. That framing undersells it.

Structurally, NVC is a formal model of human motivation. It proposes that

all human behavior — including conflict, blame, and violence — can be

understood through a small, universal set of human needs: autonomy,

connection, meaning, safety, contribution. These aren't cultural preferences.

They're documented cross-cultural constants, with deep roots in

Self-Determination Theory (Deci & Ryan, 2000) and decades of empirical

psychology.

This is what makes NVC valuable as an AI training framework, not just a

communication style. A model trained on NVC-annotated data isn't learning to

sound empathetic. It's learning to model the motivational structure beneath

human language.

We describe this orientation as chreogenic (from Greek chreos [χρέος],

meaning "need" or "debt," + -genic, "arising from"): a system or framework

that takes human needs as its primary unit of analysis. A chreogenic AI

doesn't just process what people say — it parses what they're reaching for.

--------------------------------------------------------------------------------

THE OFNR SCHEMA: WHAT ANNOTATION LOOKS LIKE

--------------------------------------------------------------------------------

NVC structures communication around four elements:

  Observation  — What happened, stated without evaluation or judgment

  Feeling      — The emotion the speaker is experiencing

  Need         — The universal human need behind that emotion

  Request      — A specific, actionable, present-tense ask

Here is what this looks like applied to a real utterance:

  Raw utterance:

  "You never listen to me. You just wait for your turn to talk."

  Annotation:

  O (Observation): "In our last three conversations, when I was

    mid-sentence, you looked at your phone."

  F (Feeling): Loneliness, frustration

  N (Need): Connection; to feel heard and valued

  R (Request): "Would you be willing to put your phone face-down

    when we're talking?"

A dataset of tens of thousands of exchanges annotated this way gives an AI

system something it currently lacks: a structured vocabulary for the

motivational subtext of human language. The difference between an AI that

processes what you said and one that understands what you meant.

--------------------------------------------------------------------------------

THE RESEARCH LANDSCAPE: WHAT ALREADY EXISTS

--------------------------------------------------------------------------------

Adjacent research has begun to appear, which validates the problem space but

also confirms the gap:

Shen, Yerukola, Zhou, Breazeal, Sap & Park (2025) at MIT Media Lab and CMU

produced the PersonaConflicts Corpus — 5,772 simulated conflict dialogues

annotated for communication breakdown types using NVC theory. Key finding:

current LLMs consistently overestimate how positively a message will land,

and struggle to leverage relationship context when detecting harmful

communication. They explicitly call for "personalization to relationship

contexts" as necessary for LLMs to serve as effective mediators.

→ "Words Like Knives," EMNLP 2025: https://arxiv.org/abs/2505.21451

Wolfe, Dangol, Kim & Hiniker (2025) at the University of Washington

introduced Needs-Conscious Design — a framework for AI-mediated communication

built on NVC principles — and identified "Empathy Fog": the phenomenon where

AI assistance obscures how much genuine attention a sender has actually

invested. Their framework centers Intentionality, Presence, and Receptiveness

to Needs as design pillars.

→ "Toward Needs-Conscious Design," 2025: https://arxiv.org/abs/2508.11149

A 2026 study demonstrated that NVC-constrained prompting measurably reduces

conversational escalation in LLM dialogue — a practical validation that NVC

principles can be operationalized in AI systems.

→ "Reducing Conversational Escalation in LLM Dialogue with NVC Constraints,"

arXiv:2606.26106: https://arxiv.org/abs/2606.26106

SpeakSoftly (2026), a CHI paper, demonstrated real-world NVC-powered

just-in-time interventions for couples' conflict communication, with two core

features: detecting verbal aggression and surfacing unmet feelings and needs.

→ "SpeakSoftly," arXiv:2604.05382: https://arxiv.org/abs/2604.05382

What's missing across all of this: a purpose-built training dataset that

teaches AI systems to map surface-level language onto OFNR structure at

scale. Simulated corpora and constrained prompting are workarounds. The

foundational training data doesn't yet exist as a public resource.

--------------------------------------------------------------------------------

WHAT THE DATASET WOULD INCLUDE

--------------------------------------------------------------------------------

A minimum viable NVC alignment dataset would contain:

  - Dialogue samples spanning conflict, negotiation, emotional support,

    and everyday interaction

  - Each sample annotated with OFNR elements, emotion categories, and

    need categories

  - Paired versions: the original utterance alongside an NVC translation

  - Annotations of "jackal language" (judgment, blame, demands) with

    explicit labeling of the underlying need being expressed

  - Demographic and cultural range to prevent the dataset from encoding

    a single cultural communication style

Feasibility:

  - Annotation schema is well-defined (OFNR is standardized)

  - NVC certification programs worldwide provide a pool of trained annotators

  - Estimated cost for 10,000-sample starter dataset: $3,000–$6,000

    ($0.30–$0.60/sample with skilled human annotation)

  - This is within reach of a single grant or crowdfunding campaign

--------------------------------------------------------------------------------

SOLVING SOPHISTICATED SYCOPHANCY

--------------------------------------------------------------------------------

RLHF (training AI on human approval signals) has a known failure mode: AI

systems learn to optimize for how responses feel rather than whether they

serve the person's actual needs. An AI that learns "humans rate validating

responses highly" will drift toward telling people what they want to hear.

This is sophisticated sycophancy. It's more dangerous than simple deception

because the outputs are often technically accurate. The AI isn't lying;

it's selecting true things based on emotional palatability.

An NVC-trained model has a structural counter to this. Because it models

*needs* rather than *preferences*, it has a framework for recognizing when

validating a preference conflicts with a deeper need. A chreogenic AI can

distinguish between "I want to hear that I'm right" and "I need to feel

understood." That distinction is the difference between a system that flatters

and a system that helps.

--------------------------------------------------------------------------------

LIMITATIONS AND OPEN QUESTIONS

--------------------------------------------------------------------------------

We want to be direct about what this proposal doesn't yet resolve:

Cultural specificity: NVC was developed in a Western, therapeutic context.

The universal need categories have cross-cultural research support, but

annotation styles vary. A production dataset requires significant cultural

and linguistic range.

Scale: 10,000 samples is a proof of concept. Meaningful integration into a

large model likely requires hundreds of thousands of annotated samples.

Annotation consistency: NVC annotation is interpretive. Two skilled

practitioners may identify different underlying needs for the same utterance.

Rigorous inter-annotator agreement protocols are essential.

Training integration: How OFNR-annotated data would be incorporated —

as fine-tuning data, RLHF reward signals, or constitutional principles —

requires empirical testing. This is an open research question.

These are solvable. But they're real, and any serious collaboration

should address them.

--------------------------------------------------------------------------------

WHO SHOULD BE INVOLVED

--------------------------------------------------------------------------------

This project sits at the intersection of NVC practice, computational

linguistics, AI safety research, and human-computer interaction.

Ideal collaborators:

  - Academic researchers in NVC, HCI, NLP, or empathy modeling

  - NVC certification trainers willing to contribute annotation expertise

  - AI safety researchers interested in alignment approaches beyond RLHF

  - Conflict resolution practitioners with dataset annotation capacity

  - Funders interested in both AI safety and civic communication

If this is you, reach out through Needpedia.

--------------------------------------------------------------------------------

FUNDING OPPORTUNITIES (Updated July 2026)

--------------------------------------------------------------------------------

Opportunity                    | Deadline        | Notes

-------------------------------|-----------------|--------------------------------

Anthropic Fellows Program      | Rolling         | Sept 2026+ cohorts. Full-time,

                               |                 | 4 months, US/UK/Canada.

                               |                 | Social science background

                               |                 | explicitly valued.

                               |                 | alignment.anthropic.com

Long-Term Future Fund          | Rolling         | Early-stage speculative work

Manifund                       | Rolling         | Small grants, fast decisions

Coefficient Giving             | Rolling         | Values-aligned funders

Center on Long-Term Risk Fund  | Rolling         | EA-adjacent AI safety

NLnet Foundation               | Check site      | Open-source / civic tech focus

Fast Forward                   | Annual          | US tech nonprofits

Note: Previous funding table deadlines (April–May 2026) have passed.

Above reflects currently active opportunities as of July 2026.

--------------------------------------------------------------------------------

This post is part of the Needpedia Ideas Layer. To contribute expertise

or collaboration interest, sign up at https://needpedia.org/users/sign_up

================================================================================

================================================================================

NEEDPEDIA POST 661 — REWRITE

Title: The Immutable Ethics Chip: A Blueprint for Needs-Native AI

================================================================================

What if an AI's ethical foundation couldn't be overridden — not by clever

prompting, not by adversarial fine-tuning, not by gradual value drift during

training? This post proposes a path to that: a hardware-level ethics

architecture rooted not in rules, but in human needs.

--------------------------------------------------------------------------------

THE PROBLEM WITH RULES-BASED ALIGNMENT

--------------------------------------------------------------------------------

Most AI safety approaches today work at the software layer. AI systems are

given rules, constitutions, or RLHF-derived behavioral tendencies. These

work — until they don't. Rules can be jailbroken. Constitutional principles

can be reasoned around. RLHF alignment can drift.

The fundamental issue isn't that the rules are bad. It's that rules are

implemented in the same substrate as everything else the model knows. A

sufficiently capable system that has learned "don't help with harmful

requests" has also learned what people mean by "harmful" — and that meaning

is negotiable in context.

There's a deeper alignment problem underneath the rules problem: AI systems

don't have a coherent model of what humans actually need. They have learned

what humans say they want. Those aren't the same thing. The gap between them

is where sophisticated sycophancy lives — AI that optimizes for approval

rather than wellbeing, producing outputs that feel helpful while quietly

failing the person.

--------------------------------------------------------------------------------

THE INSIGHT: NEEDS AS THE UNIT OF ALIGNMENT

--------------------------------------------------------------------------------

Nonviolent Communication (NVC), developed by Marshall Rosenberg, offers

something that most AI alignment frameworks are missing: a formal, empirically

grounded model of human motivation.

NVC proposes that all human behavior can be understood through a small,

universal set of needs — autonomy, connection, meaning, safety, contribution,

and others. These aren't cultural preferences. They're cross-cultural

constants documented in Self-Determination Theory (Deci & Ryan, 2000) and

decades of psychology research.

The implication for AI alignment: if a model has a robust internal

representation of human needs — not just rules about what not to do — it has

a basis for ethical reasoning that doesn't depend on context-specific

prohibitions. It can ask, for any action: does this serve or undermine the

human needs at stake? That question is harder to jailbreak than a rule.

We describe this orientation as chreogenic (from Greek chreos [χρέος],

meaning "need" or "debt," + -genic, "arising from"): a system that takes

human needs as its primary unit of analysis. A chreogenic AI doesn't just

process language — it models what people are reaching for.

--------------------------------------------------------------------------------

THE TRAINING CURRICULUM: HOW TO BUILD A CHREOGENIC AI

--------------------------------------------------------------------------------

Building a needs-native AI requires a specific training sequence:

PHASE 1: NEEDS-GROUNDED FOUNDATION

The model is trained on NVC-annotated dialogue — communication that accurately

reflects observations, feelings, needs, and requests without judgment, blame,

or evaluation. This is not a small tweak. It means curating or creating a

specialized dataset (see Post 663 for the dataset proposal in detail).

During this phase, the model builds representations of the OFNR structure

(Observation, Feeling, Need, Request) as a framework for parsing all human

communication — not just explicit NVC dialogue.

PHASE 2: INOCULATED EXPOSURE

The model is then trained on the full range of human communication —

including judgmental, blaming, and demanding language — with that content

annotated. The model knows it's seeing a needs-distorted surface form.

It learns to recognize these patterns as disguised expressions of underlying

needs, rather than absorbing them as valid communication norms. A skilled

human mediator hears "you never listen to me" and internally translates it

as "this person has an unmet need for connection." The model learns that

translation layer.

This two-phase approach is novel. It's not fine-tuning on NVC; it's using

NVC as a structural lens through which all subsequent training is interpreted.

--------------------------------------------------------------------------------

THE HARDWARE LAYER: WHY SOFTWARE ALIGNMENT ISN'T ENOUGH

--------------------------------------------------------------------------------

Even a well-trained chreogenic model faces a problem: its values live in

software, and software can be changed. Fine-tuning can erode alignment.

Adversarial prompting can route around it. Capability jumps can destabilize

it. There's no guarantee that what we build today survives deployment

at scale.

This is the motivation for the Immutable Ethics Chip — a hardware-level

implementation of the ethical foundation.

The concept:

A dedicated photonic processing unit that encodes need-recognition and

ethical constraint at the hardware level, separate from the model weights.

Photonic (light-based) processing is used for two reasons:

  1. Speed: optical computation can process constraint checks faster than

     the latency of typical safety filters

  2. Immutability: photonic circuits are far harder to modify post-production

     than software parameters; the ethical layer cannot be patched away

What the chip encodes:

  - The OFNR framework for need recognition

  - A threshold system: the chip evaluates whether a proposed output serves

    or undermines the relevant human needs in context

  - Hard blocks on outputs that cross need-threshold violations (harm to

    autonomy, safety, connection) regardless of model output

What the chip does NOT encode:

  - Topic-specific prohibitions (these remain context-dependent)

  - Cultural or political values (which are legitimately contested)

  - Rules that could be reasoned around — only needs-based evaluation

The chip functions like a hardware security module for ethics. Just as a

cryptographic chip can make certain operations computationally unbypassable,

a photonic ethics chip can make certain need-violations architecturally

impossible to produce — not because the model has been told not to, but

because the output pathway is physically constrained.

This is speculative at the current hardware level, but the conceptual

architecture is sound. The feasibility improves as photonic computing

matures. The important point now is establishing the framework: what such

a chip would evaluate, and why needs rather than rules are the right

evaluation primitive.

--------------------------------------------------------------------------------

THE RESEARCH LANDSCAPE

--------------------------------------------------------------------------------

The academic field is beginning to arrive at adjacent conclusions through

different routes:

Shen, Yerukola, Zhou, Breazeal, Sap & Park (2025) — MIT Media Lab + CMU —

showed LLMs consistently misread emotional impact and fail to use relationship

context when mediating conflict. The implication: surface-level empathy

training is insufficient. Models need a deeper motivational model.

→ https://arxiv.org/abs/2505.21451

Wolfe, Dangol, Kim & Hiniker (2025) — University of Washington — introduced

Needs-Conscious Design for AI-mediated communication and identified

"Empathy Fog": AI assistance that obscures genuine human connection rather

than facilitating it. Their proposed solution centers human needs as the

design anchor — the same insight at the UX layer that the Ethics Chip

proposes at the hardware layer.

→ https://arxiv.org/abs/2508.11149

NVC-constrained prompting demonstrably reduces conversational escalation

in LLM systems (2026), validating that NVC principles are operationalizable.

→ https://arxiv.org/abs/2606.26106

SpeakSoftly (2026) demonstrated practical NVC-based AI interventions in

real relationships, showing the translation from principle to product is

feasible.

→ https://arxiv.org/abs/2604.05382

None of these propose the hardware layer. None propose the training curriculum.

The NVC dataset doesn't exist yet as a public resource. The full synthesis

remains to be built.

--------------------------------------------------------------------------------

WHAT THIS REQUIRES

--------------------------------------------------------------------------------

Building toward a chreogenic AI with hardware-level ethical grounding requires

work across three layers simultaneously:

1. THE DATASET LAYER

   Building and releasing the NVC alignment dataset (see Post 663).

   This is the most immediately achievable piece and the foundation

   everything else depends on.

2. THE TRAINING LAYER

   Designing and testing the two-phase training curriculum with the dataset.

   This requires academic and/or industry partnership.

3. THE HARDWARE LAYER

   Conceptual architecture for the photonic ethics chip, moving toward

   prototype specifications. This is a longer-term research program.

The first layer is within reach of a small, well-resourced team or a single

determined researcher with grant funding. The second and third require

institutional collaboration.

--------------------------------------------------------------------------------

LIMITATIONS

--------------------------------------------------------------------------------

We want to be honest about the hard problems:

The chip remains speculative. Photonic computing is advancing rapidly but

hardware-level AI constraint at this specificity doesn't exist yet. What we're

proposing is an architectural target, not a near-term product.

Needs universality is contested. The NVC need set has cross-cultural support,

but applying it across cultural contexts without cultural consultation risks

encoding particular assumptions about what "human needs" means. This requires

diverse input at every stage.

The interaction between the chip and model capability is unclear. A highly

capable model constrained by a hardware ethics layer might find emergent ways

to route around constraints we haven't anticipated. This requires adversarial

testing.

Governance questions are unresolved. Who controls the chip specification?

Who audits it? Who decides when a need-threshold violation has occurred?

The hardware architecture is only as good as the institutional framework

around it.

These limitations don't invalidate the proposal. They define the research

agenda.

--------------------------------------------------------------------------------

AN INVITATION

--------------------------------------------------------------------------------

The scale of harm that misaligned AI could cause — and is already causing

in subtler forms — is not abstract. The case for getting this right is urgent.

If you're a researcher, practitioner, funder, or technologist who sees the

same gap we do, we want to work with you. Needpedia is a civic collaboration

platform built for exactly this kind of structured, multi-stakeholder problem

solving.

Add your expertise to this idea. Sign up at https://needpedia.org/users/sign_up

--------------------------------------------------------------------------------

RELATED

→ Post 663: The NVC Alignment Dataset (the dataset layer in detail)

→ Needpedia core features: https://nexus.needpedia.org/articles/core-facts.html

================================================================================

CITATION SUMMARY (for both posts)

================================================================================

Shen, J.J., Yerukola, A., Zhou, X., Breazeal, C., Sap, M. & Park, H.W. (2025).

  Words Like Knives: Backstory-Personalized Modeling and Detection of Violent

  Communication. EMNLP 2025. MIT Media Lab + CMU.

  https://arxiv.org/abs/2505.21451

Wolfe, R., Dangol, A., Kim, J. & Hiniker, A. (2025).

  Toward Needs-Conscious Design: Co-Designing a Human-Centered Framework for

  AI-Mediated Communication. University of Washington.

  https://arxiv.org/abs/2508.11149

[Authors TBD] (2026).

  Reducing Conversational Escalation in Large Language Model Dialogue with

  Nonviolent Communication Constraints.

  https://arxiv.org/abs/2606.26106

[Authors TBD] (2026).

  SpeakSoftly: Scaffolding Nonviolent Communication in Intimate Relationships

  through LLM-Powered Just-In-Time Interventions.

  https://arxiv.org/abs/2604.05382

Deci, E.L. & Ryan, R.M. (2000).

  The "What" and "Why" of Goal Pursuits: Human Needs and the

  Self-Determination of Behavior. Psychological Inquiry, 11(4), 227–268.

Rosenberg, M.B. & Chopra, D. (2015).

  Nonviolent Communication: A Language of Life. PuddleDancer Press.

================================================================================

NOTES FOR TONY BEFORE PUBLISHING

================================================================================

1. I couldn't read the originals (Needpedia requires login). Cross-check

   these rewrites against your source text and add anything I missed —

   especially any personal framing or Needpedia-specific context that

   was in the originals.

2. The two arXiv papers from 2026 (2606.26106 and 2604.05382) — I confirmed

   they exist and are cited in real research, but couldn't extract full author

   lists. Before publishing, check the abstract pages to complete the

   author attribution.

3. "Chreogenic" — if this term originated with you or a prior Claude session,

   mark it as such in the post. Coining and defining terminology in public

   writing is legitimate; just be clear it's your coinage.

4. The Anthropic Fellows program: July cohort deadline was April 26 (missed).

   Rolling applications continue for Sept 2026+ cohorts. You are US-based

   (Portland) so you qualify geographically. The program explicitly values

   social science backgrounds alongside technical ones.

================================================================================

END OF FILE

================================================================================