Beth Barnes

About

Founded METR, the leading independent lab evaluating frontier models for dangerous capabilities; the technical face of third-party evals.

Media coverage indexunvetted

Sources & placements — spoke to them or published them

• Harry Booth — TIME

  • TIME100 AI 2024: Beth Barnessource · direct quote sympathetic · Sep 5, 2024 — Names Barnes one of the 100 most influential people in AI, describes METR's GPT-4 deception test, and quotes her on catastrophic risk, regulation, and leaving OpenAI to speak independently.

• Kevin Roose — The Indian Express (republishing The New York Times)

  • How do you measure an AI boom?source · direct quote neutral · Apr 19, 2026 — Profiles METR and its time-horizon chart; Barnes says the team didn't expect such a clear trend, discusses extrapolation uncertainty and new covert-capability testing work.
  • How Do You Measure an A.I. Boom?source · direct quote neutral · Apr 17, 2026 — Reported visit to METR's Berkeley office covering Barnes's background, the chart's unexpected exponential regularity, recursive self-improvement, covert-capability testing, and methodological criticism.

• Will Henshall — TIME

  • Nobody Knows How to Safety-Test AIsource · direct quote neutral · Mar 21, 2024 — Substantial profile quoting Barnes on models as "vast alien intelligences," her departure from OpenAI, autonomous replication, weak evaluation science, and lab-access conflicts; also gives space to "safetywashing" critics.

• Asterisk Magazine staff — Asterisk Magazine

  • Crash Testing GPT-4source · interview neutral · Jan 1, 2023 — Barnes explains ARC Evals' GPT-4/TaskRabbit CAPTCHA experiment, scaffolded model access, the risk of teaching dangerous behavior, and her ideal evaluation regime.

• Ben Pace and other participants — LessWrong / GreaterWrong

  • Beth Barnes comments on Responsible Scaling Policy v3source · direct quote neutral · Mar 1, 2026 — Barnes responds to criticism that METR consumes scarce AI-safety talent, arguing few current METR staff came directly from full-time technical safety roles. [Valence re-judged under seriousness axis: Straight relay of Barnes's own substantive rebuttal to a scarce-talent criticism, not mockery.]

• Buck Shlegeris — LessWrong / GreaterWrong

  • Beth Barnes comments on 'The case for more ambitious language model evals'source · direct quote neutral · Feb 4, 2024 — Barnes questions the epistemic standard of using thirdhand anecdotes about isolated LLM outputs as strong evidence for research priorities.

• Centre for Effective Altruism staff — Effective Altruism Forum / EA Global

  • Safety evaluations and standards for AI | Beth Barnes | EAG Bay Area 23source · interview sympathetic · Jun 16, 2023 — Barnes presents ARC Evals' strategy for converting catastrophic-risk threat models into thresholds around autonomous replication, resource acquisition, and avoiding shutdown.

• Daniel Filan — AXRP — The AI X-risk Research Podcast

  • 34 — AI Evaluations with Beth Barnessource · interview neutral · Jul 28, 2024 — Interview on how METR constructs evaluations from threat models, capability elicitation, alignment testing, responsible-scaling commitments, lab relationships, and limits of external auditing.
  • Episode 6 — Debate and Imitative Generalization with Beth Barnessource · interview neutral · Apr 8, 2021 — Barnes, then at OpenAI, discusses AI alignment via adversarial debate, using humans as model stand-ins, dishonest strategies, judge limitations, and imitative generalization.

• Rob Wiblin — 80,000 Hours Podcast

  • #217 – Beth Barnes on the most important graph in AI right now — and the 7-month rule that governs its progresssource · interview sympathetic · Jun 2, 2025 — Nearly four-hour interview on model time horizons, scheming and chain-of-thought, independent evaluation, recursive self-improvement, lab incentives, interpretability, and Barnes's changing strategic views.

• Tim Scarfe — Machine Learning Street Talk

  • The AI Models Smart Enough to Know They're Cheating — Beth Barnes & David Rein [METR]source · interview neutral · May 4, 2026 — 113-minute interview on time-horizon methodology, benchmark validity, reward hacking, chain-of-thought monitoring, AI labor effects, scheming, and recursive self-improvement odds.
  • METR: Measuring AI's Time Horizon (with Beth Barnes & David Rein)source · interview neutral · Apr 19, 2026 — Extended paid cut of the Barnes-Rein interview, adding discussion of creativity, benchmark contamination, construct validity, and reward hacking.

Mentions — wrote about them

• Forbes staff — Forbes

  • METRmention · secondhand sympathetic · Jan 1, 2025 — Profiles METR as a nonprofit founded by Barnes, describing its catastrophic-risk evaluations and identifying her as founder and research leader.

• Stephen Witt — The New York Times

  • The A.I. Prompt That Could End the Worldmention · secondhand sympathetic · Oct 10, 2025 — Discusses METR's time-horizon measurement as evidence of increasing AI autonomy; borderline mention — accessible text names von Arx and Painter as METR sources rather than Barnes directly.

• TIME staff — TIME

  • TIME Reveals the 2024 TIME100 AI Listmention · passing mention sympathetic · Sep 5, 2024 — Includes Barnes, described as METR's founder and head of research, among the women leaders selected for the TIME100 AI list.

• 80,000 Hours staff — Effective Altruism Forum / 80,000 Hours

  • #217 – The most important graph in AI right now (Beth Barnes on The 80,000 Hours Podcast)mention · secondhand sympathetic · Jun 2, 2025 — Publishes highlights from the Wiblin interview, including Barnes's warnings about expert preparedness, automated AI R&D, and model honesty.

• Asterisk Magazine staff — Asterisk Magazine

  • Intelligence Testingmention · secondhand neutral · Jan 1, 2023 — Refers back to Asterisk's interview with Barnes and describes ARC Evals testing whether models can copy themselves, run on other servers, or scam people.
  • Asterisk Magazine Issue 03: AImention · passing mention neutral · Jan 1, 2023 — Announces the AI issue and notes that Barnes explains ARC Evals' work testing GPT-4.

• habryka — LessWrong / GreaterWrong

  • GPT-4: What we (I) know about itmention · secondhand sympathetic · Mar 1, 2023 — Identifies Barnes as leader of ARC's evaluations branch and describes its finding that an early GPT-4 was ineffective at autonomous replication, resource acquisition, and avoiding shutdown.

• Joe Rogero and Mitchell Howe — AI StopWatch

  • “Not on top of it”mention · secondhand sympathetic · May 23, 2026 — Lead item built around Barnes's X thread warning experts are "not on top of it"; praises her candor on extinction risk, rushed development, weak evals, and METR's reliance on voluntary lab cooperation.

• Rohin Shah — Alignment Newsletter / LessWrong

  • AN #114: Theory-inspired safety solutions for powerful Bayesian RL agentsmention · secondhand neutral · Jan 1, 2020 — Summarizes Barnes's OpenAI call for adversarial collaborators to find weaknesses in an AI-debate protocol and dishonest strategies that could defeat trained honest debaters.

• Thomas Claburn — The Register

  • AI coding tools make developers slower, study findsmention · secondhand neutral · Jul 11, 2025 — Names Barnes among the study authors and relays caution that the 19% slowdown was specific to experienced developers in familiar, mature repos.

• 定慧 — Sina Finance / 新智元

  • 2026年，或许是人类最后一次掌控AImention · secondhand sympathetic · Apr 21, 2026 — Retells Kevin Roose's METR profile, presenting Barnes's organization and chart as evidence AI capability improvement may itself be accelerating.

Scan coverage note: The report's search could not access subscription news/broadcast databases (LexisNexis, Factiva, ProQuest, Meltwater, TVEyes, Cision), closed/paywalled podcast transcripts, Patreon-only audio, deleted episodes, X/Twitter's historical archive and private/deleted posts, LinkedIn's full archive behind authentication, inconsistently indexed Apple Podcasts/Spotify metadata, some bot-blocked New York Times pages (verified via syndications/archives instead), and print-only, local-radio, non-English, or library archive material outside general web search. The visible record is concentrated from 2023 onward; despite searching with numerous name/organization variants (Beth Barnes, Elizabeth Barnes, Beth May Barnes, METR, ARC Evals, OpenAI/DeepMind, GPT-4/TaskRabbit, autonomous replication, responsible scaling, time horizons, developer-productivity study), no qualifying media coverage was found from July 15, 2016 through 2019. [Codex deep harvest 2026-07-15.]