Assertion Supported AI assessment confidence: 95% certainty 4/5 debate potential 4/5

Bissell: CCP bias is identifiable in Qwen and DeepSeek-R1 representation spaces

Mark Bissell · Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell · Feb 5, 2026 · at 10:08

Mark Bissell, Member of Technical Staff at Goodfire AI, discusses using mechanistic interpretability to identify and surgically extract political bias vectors from Chinese foundation models like Qwen and DeepSeek-R1.

0:00 / 0:06exact quote · 6.9s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“Well, there's, there are certainly internal, yeah, parts of the representation space where you can sort of see where that lives.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

In AI models such as Qwen and DeepSeek-R1, political bias aligned with the Chinese Communist Party exists in identifiable, specific regions of the model's internal representation space.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

Multiple empirical mechanistic interpretability and representation engineering studies published between 2024 and 2026 have isolated and localized specific internal representation vectors, circuits, and features corresponding to Chinese Communist Party (CCP)-aligned censorship and narrative steering in both Qwen and DeepSeek-R1 model families. A 2025 University of Virginia study identified causal representation steering vectors governing refusal and 'thought suppression' on CCP-sensitive topics in DeepSeek-R1-Distill models. Independent representation engineering research (AUGMXNT, 2024-2026) isolated CCP-specific refusal directions in Qwen models, and multi-model circuit analyses in early to mid-2026 located middle-layer activation circuits and sparse autoencoder features specifically routing sensitive topics to CCP-aligned outputs.

Cyberey and Evans (University of Virginia, 2025) identified internal representation steering vectors in DeepSeek-R1-Distill models that exert direct causal control over refusal and thought suppression on CCP-sensitive topics.
Representation engineering research by Leonard Lin (AUGMXNT, 2024–2026) and subsequent tooling identified isolated activation directions responsible for CCP-sensitive refusals across the Qwen family.
A March 2026 multi-model study (arXiv:2603.18280) and a May 2026 circuit localization analysis of Qwen models mapped distinct internal directions and middle-layer circuits (Layers 11–20) governing PRC-sensitive framing.
Research presented at the ICML 2026 Mechanistic Interpretability Workshop used crosscoder sparse autoencoder (SAE) feature diffing to isolate specific features controlling Chinese state narrative alignments in frontier Chinese LLMs.

Checked against

[1] github.com [2] github.io [3] themoonlight.io [4] alphaxiv.org [5] openreview.net [6] github.com [7] shisa.ai [8] huggingface.co

"Goodfire" "Mark Bissell" "DeepSeek" OR "Qwen" "abliteration" "Qwen" "CCP" OR "censorship" OR "bias" "steering vector" "DeepSeek-R1" "political" OR "censorship" OR "bias" "refusal direction" OR "refusal vector" "DeepSeek" OR "Qwen" "CCP" OR "censorship" "mechanistic interpretability" "Qwen" "political bias" OR "censorship" "Goodfire" "DeepSeek" OR "Qwen" "bias" OR "censorship" representation engineering Chinese Communist Party bias Qwen DeepSeek

assessed as of 2026-09-02

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

Chinese Communist Party DeepSeek-R1 Goodfire AI Mark Bissell Qwen

⚑ Spot an error? Report it

More from Mark Bissell

Insight

Bissell: Activation Steering and In-Context Learning Are Quantitatively Equivalent

“He actually has a paper that, as well as some, you know, others from the team and elsewhere, that go into the essentially equivalence of activation steering and in-context learning, and how those are from a, he thinks of everything in a cognitive neuroscience …”

Mark Bissell Feb 5, 2026 ▶ 30:57 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Insight

Bissell: Probing internal model features matches LLM-as-a-judge quality at 500x lower cost

“If you ask that model, try to like use it as an LLM as a judge, it's not very good. But if you probe its mind and you sort of detect when the features related to personally identifiable information are firing, that gets you the highest recall of anything. It's…”

Mark Bissell Dec 31, 2025 ▶ 9:53 [State of MechInterp] SAEs in Production, Circuit Tracing, AI4Science, "Pragmatic" Interp — Goodfire

Opinion

Bissell: Subliminal learning typically only affects models sharing initial random seeds

“I think it only applies to models that were initialized from the same starting Z. Usually, yes.”

Mark Bissell Feb 5, 2026 ▶ 13:24 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Assertion Supported

Bissell: Goodfire performs activation steering on 1-trillion parameter Kimi K2

“Here you're going to see steering on a one trillion parameter model. This is Kimi K two.”

Mark Bissell Feb 5, 2026 ▶ 22:32 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Assertion Supported

Bissell: Steering Experiments Can Predict Examples Needed for Jailbreaks

“What's in this in context learning and activation steering equivalence paper is you can like predict the number of examples that you will need to put in there in order to jailbreak the model. By doing steering experiments and using this sort of like equivalenc…”

Mark Bissell Feb 5, 2026 ▶ 32:22 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Opinion

Bissell: Current AI Training and Post-Training Methods Are Primitive

“I hope that we look back at how we're currently training models and post training models and just think what a primitive way of doing that right now. Like there's no intentionality really in.”

Mark Bissell Feb 5, 2026 ▶ 35:33 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Bissell: CCP bias is identifiable in Qwen and DeepSeek-R1 representation spaces

More from Mark Bissell

Turn any episode into a week of clips.