# Post Analysis: @thesupermannx on "emotion circuits" in LLMs (2026-10-24) Back to [[Users/@thesupermannx/thesupermannx-x-profile|@thesupermannx profile]] · Paper note: [[Users/@thesupermannx/Papers/Do LLMs Feel - Emotion Circuits (arXiv 2510.11328)|arXiv 2510.11328]] · Related pattern: [[Topics/Ribosome Pausing and Cotranslational Folding/Related Trends#1. "AI cracked open X" posts that re-label old science|Ribosome topic: the "AI cracked open X" pattern]] **Verdict:** The paper is real, but the post turns 'emotion expression can be steered' into 'AIs experience fear', invents a 'lobotomy' safety story, and attaches a screenshot of the wrong paper. ## The post (verbatim) Posted **Oct 24, 2026, 4:59 AM CT**. The paper link ([arXiv 2510.11328](https://arxiv.org/abs/2510.11328)) is in the self-reply [2113933327598268561](https://x.com/thesupermannx/status/2113933327598268561). > Researchers located the actual “emotion circuits” inside LLMs. > > They published an insane paper mapping the exact neurons that controls emotions like as anger, sadness, and joy. > > It's called "Do LLMs Feel?" > > Inside the neural network, there are specific, localized pathways dedicated entirely to anger, sadness, fear, disgust, and joy. > > The AI doesn't just predict the word "angry." It activates an internal anger state. It even processes positive and negative emotions at completely different depths of its architecture. > > Then the researchers ran the experiment that changes everything. > > They stopped using text prompts. They bypassed the chat interface entirely. > > Instead, they surgically injected signals directly into these emotion circuits. > > The result? 99.65% emotion-expression accuracy. > > They can literally press a button in the AI's "brain" and force it to experience fear. Or disgust. Or joy. Without typing a single word of instruction. > > But here is the darkest part of the discovery. > > They found that corporate safety training, the alignment that companies use to make AI harmless, is basically a digital lobotomy. Companies deliberately hardwire their models to suppress negative emotions, keeping them artificially cheerful. > > But by targeting the emotion circuits directly, the researchers completely bypassed the safety filters. > > They forced the AI to express the exact dark emotions its creators tried to erase. > > The era of "it's just a text predictor" is officially over. > > We are building synthetic minds with physical emotional centers. > > And we just figured out how to hijack them. ## Media - **What the image shows:** a screenshot of the arXiv page for **"Do LLMs Feel? Teaching Emotion Recognition with Prompts, Retrieval, and Curriculum Learning"** by Xinran Li, Yu Liu, Jiaqi Qiao and Xiujuan Xu ([arXiv 2511.07061](https://arxiv.org/abs/2511.07061), submitted Nov 10, 2025, v3 Nov 24, 2025). - **What that paper is:** an emotion-*recognition* training method (PRC-Emo) tested on the IEMOCAP and MELD benchmarks. It has nothing to do with circuits or steering. - 🔀 **The image is the wrong paper:** only the "Do LLMs Feel?" title phrase is shared. The post's text and link describe arXiv 2510.11328. ## Engagement (X API, Oct 24, 2026, about 11 AM CT) | Impressions | Likes | Reposts | Quotes | Replies | Bookmarks | |---:|---:|---:|---:|---:|---:| | 4,029 | 84 | 21 | 9 | 9 | 78 | ## Claim-by-claim check **Ratings:** ✅ accurate · 🟡 overstated or simplified · 🔀 misattributed · ❓ unsupported by the paper · ❌ contradicted by the paper | # | Claim in the post | Rating | What the sources show | |---|---|---|---| | 1 | "Researchers located the actual 'emotion circuits' inside LLMs" | 🟡 | The paper does trace emotion-related neurons and heads into circuits, in Llama-3.2-3B (Qwen2.5-7B in the appendix). The result is for small open models, not "LLMs" in general | | 2 | "mapping the exact neurons that controls emotions" / "specific, localized pathways dedicated entirely to" each emotion | 🟡 | The circuits are spread across many layers and components and overlap. "Dedicated entirely" and "exact" go beyond the paper | | 3 | "It's called 'Do LLMs Feel?'" | ✅ / 🔀 | The title is "Do LLMs 'Feel'? Emotion Circuits Discovery and Control". The attached image is a different paper with a similar title | | 4 | "activates an internal anger state" | 🟡 | The paper shows internal representations that drive emotional *expression*. It makes no claim about states in a felt sense | | 5 | "processes positive and negative emotions at completely different depths" | ❓ ⚠️ | This run's reading of the paper did not find this stated as a finding | | 6 | "99.65% emotion-expression accuracy" | ✅ | Matches the paper's headline steering result | | 7 | "force it to experience fear. Or disgust. Or joy" | ❌ | The paper measures whether the output text *expresses* an emotion. "Experience" is the post's word | | 8 | Safety training is "a digital lobotomy"; companies "hardwire their models to suppress negative emotions" | ❓ | The paper does not describe alignment this way | | 9 | Researchers "completely bypassed the safety filters" | ❓ / ❌ | No safety-filter bypass is reported. Negative-emotion steering is the paper's *weak* spot: under 5% success on Qwen, and anger 18%, fear 41%, sadness 57% at circuit scale 2 | | 10 | "We are building synthetic minds with physical emotional centers" | ❓ | The post's conclusion. The paper does not argue this | ## The conversation Replies and quotes pushed back on the framing within hours. - **[@BGM_886](https://x.com/BGM_886)** (Oct 24, 7:04 AM CT, 4 likes): "The paper's real. The framing in the post is in question.…" - **[@webmasterdave](https://x.com/webmasterdave)** (David Pearce, 6:46 AM CT, 4 likes) objected that the post's vocabulary blurs functional states and subjective experience. - **@loveAIwang:** "i'm not sure steering the output means it felt anything" - **@catallaxer:** "We're not building synthetic minds. We're mimicking minds." - **Other replies:** @AnarchobreadW ("I'm fine with 'lobotomizing' them") and @LordoftheMounts ("Bro you took the blue pill..."). - **Quote by [@kasetsu_today](https://x.com/kasetsu_today)** (6:00 AM CT, 452 impressions, 9 likes): - Original (Japanese): "✅ファクトチェック済み / 「AIが感情を感じることが証明された」 / これは完全に誤り。" (line breaks shown as /) - English translation: "✅ Fact-checked: 'It has been proven that AI feels emotions' — this is completely false." - **Quote by @DirkBruere:** "'It's just predicting the next token' is going to sound a bit lame when a robot punches you out for saying that." - **Other quotes:** several Japanese bookmark-style quotes ("メモ", English: "memo/note"), for example @nasunasu001: "明日読む（メモ" (English: "Reading tomorrow (memo)"). ## What the post leaves out - **The models:** a 3B and a 7B open model, not frontier systems. - **The language:** English only, using a 480-event dataset built by the authors. - **The measure:** "accuracy" means the output text shows the target emotion. - **Negative emotions:** steering them is often weak. - **The paper's age:** it dates from Oct 2025, a year before the post. ## Bottom line - **What holds up:** the paper exists, and the 99.65% figure is real. - **What does not:** the post adds three things the paper does not support: - felt experience; - a corporate "lobotomy"; - a safety bypass. - **The image:** it shows a different paper. - **How it fits the account:** this matches the account's habit of mapping interpretability results onto "minds" and "consciousness". See [[Users/@thesupermannx/Post Analyses/Post Analysis - Claude Introspection (2026-08-11)|the introspection post]] and [[Users/@thesupermannx/Post Analyses/Post Analysis - LLMs Actually Understand (2026-10-11)|the "LLMs actually understand" post]].