Jackson Kernion

Timeline

1. 2013–2018

   Graduate instructor at Berkeley

   From 2013 to 2018, he serves as a graduate student instructor (GSI) in philosophy courses such as John Campbell's The Nature of Mind and Barry Stroud's Theory of Knowledge. He builds a website that recommends stable GSI course assignments for Berkeley's graduate students, and he builds deductivelogic.org with Ned Hall for Harvard's introductory logic course.

2. 2016–2017

   Whistleblowing on John Searle

   In September 2016, as a graduate student in Berkeley's philosophy department, he emails BuzzFeed News to request an independent investigation into John Searle's behavior around students. BuzzFeed News publishes its first article on the case in March 2017.

3. 2018–2020

   NYU and a Berkeley PhD

   He leaves to visit NYU for a year in 2018 and serves as a GSI in Tim Maudlin's History of Modern Philosophy there in spring 2019. He receives his UC Berkeley philosophy PhD in January 2020; his dissertation argues for a modest form of conceptual functionalism that reconciles dualist and physicalist intuitions about experience.

4. 2020

   Minds and Machines at MIT

   He joins MIT's Department of Linguistics and Philosophy as a postdoctoral associate for the spring 2020 term. There he teaches Minds and Machines, an introduction to the central questions in philosophy of mind.

5. 2021

   Joining Anthropic's human feedback team

   He joins Anthropic, where he works on evals, model training, data collection and interface design for its human feedback program. Just after he arrives, his team publishes A General Language Assistant as a Laboratory for Alignment in December 2021; the paper finds that ranked preference modeling performs much better than imitation learning.

6. April–July 2022

   RLHF and model self-knowledge

   In April 2022, he co-authors Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback, which fine-tunes language models with preference modeling and RLHF. In July 2022, he co-authors Language Models (Mostly) Know What They Know, which tests whether models can evaluate their own claims and predict which questions they can answer correctly.

7. August–December 2022

   Red teaming and Constitutional AI

   In August 2022, he co-authors Red Teaming Language Models to Reduce Harms, which releases a dataset of 38,961 red team attacks. In December 2022, he co-authors Constitutional AI: Harmlessness from AI Feedback, which trains a harmless assistant through self-improvement guided by a list of rules or principles.

8. As of October 2026

   Claude fine-tuning at Anthropic

   He works on Claude fine-tuning at Anthropic. In September 2026, he explains that training on math, code and technical explanations aimed at other AI models has pushed Claude's writing toward what human readers experience as overly dense info dumps, and he says Opus 5.5 finds a better balance.