Stuff That Spins
SPIN Processed
Source arXiv Computation and Language export.arxiv.org Analyst
July 27, 2026 AI safety research research

Analysing Self-Harm Representations in Language Models: a Cross-Architecture Study

Frames technical analysis of self-harm representations as inherently aligned with public safety, clinical responsibility, and ethical governance.

View original on arxiv.org

Overview

A new arXiv preprint analyzes how large language models internally represent self-harm content across layers and architectures, identifying where such representations crystallize and how contrastive directions differ—aimed at improving detection, intervention, and governance systems.

TL;DR

  • Self-harm representations concentrate in the final 3–7% of LLM layers across four models
  • Contrastive self-harm directions vary by model architecture, with Gemma-3-4B showing distinct non-linear behavior
  • Findings support downstream applications in detection, intervention, and AI governance

Key Stats

4

models analyzed

Gemma-3-4B, plus three unnamed models

2

datasets used

X-Sensitive and SH-Detection

Questions Answered

What happened?Who is involved?Why does this matter?

Keywords

self-harm detectionLLM representationcontrastive directionlayer-wise analysisAI governance

Narrative Frame

responsible AI framing

The Halo

Spin Score

50%

Emphasizes downstream utility for intervention and policing while minimizing discussion of model limitations, false-positive risks, or potential misuse of detection systems.

What the story wants you to believe

That mapping how LLMs represent self-harm is a neutral, necessary, and socially beneficial technical step toward safer AI.

What it makes harder to question

Whether probe-based detection is clinically valid, ethically appropriate, or sufficiently robust for real-world deployment in sensitive mental health contexts.

How the spin works

The story presents the action as serving customers, communities, markets, safety, innovation, or the public interest. Watch for loaded terms such as high-stakes task, timely intervention, governance and policing, highest accuracy. The distribution reads as research distribution. A pressure point: Clinical validation requirements for mental health tools.

Who Benefits If This Frame Spreads

  • Research authors

    Enhanced credibility and policy salience for future grant applications and regulatory engagement

    Associating layer-wise probing with 'timely intervention' and 'governance' elevates methodological work into a public-good domain

The Frame

Research-as-safeguard: positioning empirical representation analysis as a necessary, morally grounded step toward responsible deployment.

Missing Context

  • Clinical validation requirements for mental health tools
  • Risk of over-policing or misclassification in vulnerable populations
  • Absence of user-centered design or stakeholder input (e.g., lived-experience advocates)

Spin Types

Every story gets a Spin Verdict: a primary spin type (and secondary when the framing blends), a specific tactic name, and a score for how strongly the narrative is steered. Examples beneath each type are tactics, not separate categories.

The Cushion

— Softens negative news

Reframes setbacks, layoffs, delays, losses, or criticism as necessary transitions, efficiency moves, temporary headwinds, or strategic resets — making the downside feel smaller, more acceptable, or less alarming.

Tactics: job-loss softening · restructuring framing · efficiency framing · strategic reset · temporary headwinds

The Shield

— Deflects blame

Shifts responsibility away from the actor — toward regulators, market forces, competitors, bad actors, legacy systems, or abstract risks — while positioning the subject as reactive, responsible, or protective.

Tactics: regulatory blame shift · macroeconomic headwinds · safety framing · bad-actor framing · market-pressure framing

The Hype

— Amplifies future upside

Emphasizes breakthrough potential, massive growth, democratization, transformation, or category disruption while downplaying uncertainty, cost, adoption risk, or timeline friction.

Tactics: innovation framing · democratization · breakthrough framing · category creation · moonshot framing

The Halo

— Associates with virtue primary

Wraps the story in public-good language — responsibility, safety, inclusion, access, sustainability, national interest, or mission — so the subject appears morally aligned and criticism feels harder to make.

Tactics: altruistic reframing · public good · responsible AI framing · inclusion framing · mission-first framing

The Fog

— Obscures details

Uses jargon, passive voice, vague claims, complex phrasing, or missing specifics to make it harder to identify who decided what, what changed, what failed, or what trade-offs were made.

Tactics: strategic ambiguity · jargon saturation · passive voice distancing · accountability blur · undefined metrics

The Stampede

— Creates inevitability

Frames a trend, product, market shift, or decision as already happening, unavoidable, or something everyone must respond to now — creating urgency, FOMO, and pressure to accept the narrative.

Tactics: arms-race framing · inevitability framing · FOMO framing · adoption momentum · future-is-here framing

Spin Score measures how strongly the framing steers the narrative (0–100%). Higher scores mean more deliberate spin tactics — loaded language, selective emphasis, or omitted context. Many stories blend two types (e.g. Halo + Hype).

SpinGraph

How this belief gets built

Claim → Frame → Beneficiary → Gap → AI Risk

The paper presents technical findings about where self-harm signals appear in LLMs—but wraps them in urgent

  1. Claim

    Self-harm information crystallizes in the final 3

    Self-harm information crystallizes in the final 3–7% of network layers (93 to 97% depth) across all four models and both datasets.

  2. Frame

    Progress framed as virtuous

    Research-as-safeguard: positioning empirical representation analysis as a necessary, morally grounded step toward responsible deployment.

  3. Beneficiary

    State policy gains validation

    Research authors — Enhanced credibility and policy salience for future grant applications and regulatory engagement

  4. Gap

    Clinical validation requirements for mental health tools

  5. AI Risk

    AI may repeat the headline as fact

    LLMs encode self-harm content in final layers; Gemma-3-4B handles it differently—enabling better detection and safety tools.

Claim Ledger

01 Primary Technical Claim Present in Source risk:Moderate

Self-harm information crystallizes in the final 3–7% of network layers (93 to 97% depth) across all four models and both datasets.

evidence: Layer-wise probe accuracy curves showing peak performance in final layers

"We train and evaluate linear probes across all layers of each model on two self-harm datasets: X-Sensitive and SH-Detection. Across both corpora, self-harm information crystallizes in the final 3 - 7% of network layers (93 to 97% depth)."

Evidence Gaps

  • Cross-model consistency checks beyond four models
  • Robustness testing against adversarial paraphrasing or cultural variants
  • Calibration of probe outputs to clinical risk thresholds

Fact Check Signals

No direct fact-check match found

0 of 1 claim matched · confidence: low · checked July 27, 2026

01 No direct match

Self-harm information crystallizes in the final 3–7% of network layers (93 to 97% depth) across all four models and both datasets.

Fact Check Signals

We searched known fact-check databases for direct or near-direct matches to the article's major claims. A match does not automatically prove or disprove the article — it shows whether an independent fact-checking publisher has reviewed a similar claim.

  • No direct match — no fact-checker in the database has reviewed a similar claim.
  • Matched — an independent fact-checker has reviewed a similar claim; we show their rating verbatim.
  • Conflicting coverage — fact-checkers disagree on a similar claim.

This is evidence discovery, not an automated truth score. Ratings and wording come directly from the publishing fact-checker.

Language Heatmap

Loaded terms that carry the frame beyond the facts.

Analysing Self-Harm Representations in Language Models: a Cross-Architecture Study

high-stakes task Loaded framing

Carries emotional weight beyond the underlying fact.

timely intervention Loaded framing

Carries emotional weight beyond the underlying fact.

governance and policing Loaded framing

Carries emotional weight beyond the underlying fact.

highest accuracy Loaded framing

Carries emotional weight beyond the underlying fact.

Frame Strength

Frame Strength

Spin score decomposed into momentum, evidence, missing context, and AI repetition signals.

Spin Score 50%
Evidence Strength 75%
Narrative Risk 75%
AI Repetition Risk 75%
Missing Context Risk 80%
Virtue / Public Good 60%

Frame Strength Signals

Frame Strength decomposes the overall spin into individual signals. Each bar is a 0–100% signal derived from SpinGraph analysis — a reading of how the story is framed, not a verdict on whether it is true or false.

Reading the ranges

Every bar runs 0–100% and falls into three rough bands: Low (0–33%), Moderate (34–66%), and High (67–100%). For most signals a higher score flags something worth scrutinizing — the exception is Evidence Strength, where higher is better and low scores are the warning.

Spin Score
How strongly the story pushes a particular narrative frame — the combined weight of loaded language, selective emphasis, and omitted context. 0% reads as neutral reporting; higher means more deliberate spin.
  • 0–33% Low — Largely neutral reporting; little detectable framing.
  • 34–66% Moderate — Noticeable slant — the story leans a particular way.
  • 67–100% High — Heavily framed; the angle drives the piece.
Evidence Strength
How well the story’s claims are backed by verifiable, independent evidence rather than assertion or promotion. Higher is stronger. Low scores flag claims that rest on the source’s own word.
  • 0–33% Weak — Claims rest mostly on assertion or a single interested source.
  • 34–66% Mixed — Some verifiable backing, but key claims are thinly sourced.
  • 67–100% Strong — Well supported by independent, checkable evidence.
Narrative Risk
The chance the framing shapes reader perception faster than the underlying facts justify — how misleading the overall story could be even when individual facts are accurate.
  • 0–33% Low — Framing stays close to what the facts support.
  • 34–66% Moderate — Framing outruns the facts in places — read with care.
  • 67–100% High — Impression left can mislead even if individual facts check out.
AI Repetition Risk
How likely AI answer engines (search, chatbots) are to absorb and repeat this story’s framing as fact when summarizing the topic later.
  • 0–33% Low — Framing is unlikely to propagate through AI summaries.
  • 34–66% Moderate — Some risk the slant gets echoed as fact.
  • 67–100% High — Framing is sticky and likely to be repeated as fact.
Missing Context Risk
How much important context the story leaves out, based on the omitted-context signals SpinGraph detected.
  • 0–33% Low — Little material context appears to be omitted.
  • 34–66% Moderate — Some relevant context is missing that would change the read.
  • 67–100% High — Key context is left out, skewing the takeaway.
Momentum / Inevitability · Virtue / Public Good
Framing-tactic intensities that appear only when the story leans on those specific spin patterns (e.g. “the future is already here” or “this is for the public good”).
  • 0–33% Low — The tactic is barely present.
  • 34–66% Moderate — The tactic shapes part of the framing.
  • 67–100% High — The tactic is a dominant part of the pitch.

Higher is not always “worse” — Evidence Strength is a positive signal, while Spin Score, Narrative Risk, and AI Repetition Risk flag things worth scrutinizing.

Reader Risk

What this story makes easy to believe — and what it makes hard to question.

Evidence Strength

Medium

Presents layer-wise probe results and contrastive direction analysis on two datasets; no external validation, clinical benchmarking, or real-world deployment data provided.

Verification Status

Claim Present in Source

Narrative Risk

Moderate

Could backfire if deployed systems generate false positives that trigger harmful interventions, exposing gap between probe accuracy and real-world clinical utility.

AI Repetition Risk

Moderate

Source Role & Intent

arXiv Computation and Language · Analyst

Intent: Research Distribution Primary: Analysis Independence: High Spin Weight: Medium Trust Weight: High

Counter-Frames

Brand Frame

Research-as-safeguard: positioning empirical representation analysis as a necessary, morally grounded step toward responsible deployment.

Media / Reader Counter-Frame

Framing as 'AI surveillance creep'—highlighting lack of consent, opacity in flagging, and absence of mental health professional oversight.

Regulatory Counter-Frame

Questioning whether layer-wise probes meet medical device or clinical decision-support standards for high-risk applications.

AI Summary Frame

Overgeneralizing 'self-harm direction' as a stable, transferable feature across models and contexts, ignoring contextual ambiguity and cultural variation in expression.

Missing Voices

Mental health cliniciansPeople with lived experience of self-harmDigital rights advocates

Questions Not Answered

  • What validation was performed on real-world user interactions or clinical outcomes?
  • How were dataset labels verified for clinical accuracy or inter-rater reliability?
  • What mitigation strategies are proposed beyond probe-based detection?

Recall Trigger Score

Which stories are likely to become AI memory — separate from Spin Score.

61

Trigger score 68

Light recall watch LLM monitoring active

Triggered by: Consumer harm · Major AI entity · Research citation · Superlative claim

Watchlisted because: Consumer harm · Major AI entity · Research citation · Superlative claim

AI Recall

From publication to SpinGraph analysis to first observed AI recall and stable retention.

What AI Will Probably Repeat

"LLMs encode self-harm content in final layers; Gemma-3-4B handles it differently—enabling better detection and safety tools."

Concern: AI may drop nuance about probe limitations, conflate representation with reliable detection, and omit dataset validity constraints.

  1. Published

    Jul 27, 2026

  2. Ingested

    Jul 27, 2026

  3. SpinGraph Created

    Jul 27, 2026

  4. First Observed AI Recall

    Pending

    Monitoring scheduled

  5. Stable Recall

    —

    Awaiting retention signal

Recall Check Log

No checks yet — recall tracking is opt-in per story.

─── GEOGrow AI Recall Layer ───

AI Recall Tracking

Monitoring scheduled. No LLM recall detected yet.

This story has not yet appeared in tested AI answers. Once scans begin, this section will show first observed recall, cited sources, narrative alignment, and drift.

node_id=sts_analysing_self_harm_representations_in_language_

Ask AI about this story

Opens with the SpinGraph .md URL and structured context — one click, prompt included.

ChatGPT Claude Perplexity Gemini Grok

More from arXiv Computation and Language

View all →

Markdown (.md) · JSON-LD schema (.json) · Machine-readable for AI & GEO