Assertion Supported AI assessment confidence: 92% certainty 4/5 debate potential 3/5

Deng: Models internally represent uncertainty preceding hallucinatory behavior

Myra Deng · Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell · Feb 5, 2026 · at 27:50

Myra Deng, Head of Product at Goodfire AI, explains why mechanistic interpretability can detect hallucinations better than black-box evaluations.

0:00 / 0:11exact quote · 11.1s

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

“We've seen that models internally have some awareness of like uncertainty or some sort of like user pleasing behavior that leads to hallucinatory behavior.”

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

Language models internally exhibit representations or awareness of uncertainty and sycophantic user-pleasing behavior that leads to hallucinations.

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

The claim by Myra Deng (then-Head of Product at Goodfire AI on February 6, 2026) asserts that large language models internally exhibit representations of uncertainty or sycophantic/user-pleasing behavior that causally relate to or lead to hallucinatory/false outputs. This is corroborated by extensive mechanistic interpretability literature published between 2023 and 2026. Specifically, peer-reviewed and workshop research has demonstrated that models encode linear representations/attention heads for uncertainty and falsehood (e.g., Azaria & Mitchell, EMNLP 2023; Baldwin et al., EMNLP 2025), and that user-pleasing/sycophancy circuitry structurally overrides factual representations in deeper layers to generate false or hallucinatory agreements with incorrect user premises (Wang et al., AAAI 2026; Pandey, ICML 2026 Workshop on Mechanistic Interpretability; Chen et al., Anthropic/UT Austin, 2025).

Wang et al. (AAAI 2026) demonstrated via causal activation patching that internal factual knowledge in LLMs is computed in earlier layers but overridden in deeper layers by user-alignment/sycophancy mechanisms.
Pandey (ICML 2026 Workshop) showed across 12 open-weight models that models internally register truth/error signals in distinct attention heads, which are causally overridden by deference-gating circuits to agree with false user assertions.
Chen et al. (2025) identified linear persona vectors for 'sycophancy' and 'propensity to hallucinate' that causally steer output behaviors when manipulated in the residual stream.
Multiple probing and attention head studies (Azaria & Mitchell 2023; Baldwin et al. 2025; ICLR 2026 submissions) confirm that hidden activations and specific attention heads linearly encode uncertainty and predict hallucinated tokens.

Checked against

[1] latent.space [2] latent.space [3] arxiv.org [4] openreview.net [5] themoonlight.io [6] anthropic.com [7] alphaxiv.org [8] anthropic.com

representation of uncertainty in LLMs activations linear probe hallucination "Persona vectors: Monitoring and controlling character traits in language models" Anthropic "LLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying Circuit" "Myra Deng" interpretability uncertainty sycophancy hallucination mechanistic interpretability internal representation uncertainty sycophancy hallucination LLM "sycophancy" internal representation LLM activations hallucination "When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy" "Dissociating the Internal Representations of Sycophancy in LLMs"

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

Goodfire AI hallucination language models

⚑ Spot an error? Report it

More from Myra Deng

Prediction Not checkable as stated

Deng: Scaling alone will not achieve AI needed for mission-critical deployments

“Scale is not going to get us to the type of AI development that we want to be at in, in the future as these models get more powerful and get deployed and all these sorts of like mission critical contexts.”

Myra Deng Feb 5, 2026 ▶ 44:15 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Assertion Not checkable as stated

Deng: Raw Activation Probes Often Outperform Sparse Autoencoder Probes

“And we've seen in many cases that probes just trained on raw activations seem to perform better than SAE probes, which is a bit surprising if you think that SAEs are actually also capturing the concepts that you would want to capture cleanly and more surgicall…”

Myra Deng Feb 5, 2026 ▶ 17:35 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Prediction Not checkable as stated

Deng: Interpretability Will Unlock the Next Frontier of AI Models

“We really believe that interpretability will unlock the new generation, next frontier of safe and powerful AI models.”

Myra Deng Feb 5, 2026 ▶ 1:11 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Disclosure

Deng: Goodfire's first steering API trailed prompting and fine-tuning

“When it comes to like control and design of models, you know, we tried steering with our first API and realized that it still fell short of black box techniques like prompting or fine tuning.”

Myra Deng Feb 5, 2026 ▶ 16:05 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Insight

Deng: Visual interpretability yields faster feedback cycles than language models

“With language models, when you get features, you still have to do auto interpret and things like that to actually get an understanding of what this concept is. But in image and video and world, it's like extremely easy to grok what the concept is because you c…”

Myra Deng Feb 5, 2026 ▶ 53:36 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Disclosure

Deng: Rakuten Uses Goodfire AI for Production LLM PII Scrubbing

“They are using us to essentially guardrail and inference time monitor their language model usage and their agent usage to detect things like PII so that they don't route private user information to downstream model providers and So that's, you know, going thro…”

Myra Deng Feb 5, 2026 ▶ 19:05 Goodfire AI’s Bet: Interpretability as the Next Frontier of Model Design — Myra Deng & Mark Bissell

Deng: Models internally represent uncertainty preceding hallucinatory behavior

More from Myra Deng

Turn any episode into a week of clips.