SARAH SCHWETTMANN

Behavioral Archetype

THE SCIENTIST WHO BUILT THE MICROSCOPE — Subject is a neuroscience-trained interpretability researcher who is credited with building much of the lab’s instrument: the automated agents that probe what individual units inside a neural network do. She is the second name on a two-founder nonprofit whose CEO, Jacob Steinhardt, carries the press. She co-wrote the report on a pre-release OpenAI model that “frequently fabricates actions it took,” and the model access for that work came from OpenAI. This profile scores the position, not the person.

Essence Indicators

• Co-Founder and Chief Scientist of Transluce. The lab’s own team page lists her as “Co-Founder, Chief Scientist,” beside Steinhardt as “Co-Founder, CEO” (Transluce, About). The launch post of October 23, 2024 is signed “Jacob and Sarah” (Transluce, Introducing Transluce).

• MIT. In her own words: “I’m a Research Scientist in MIT CSAIL with the MIT-IBM Watson AI Lab,” and “I did my PhD in Brain and Cognitive Sciences at MIT, as an NSF Fellow working with Josh Tenenbaum and Antonio Torralba” (Sarah Schwettmann). The same wording appears on a MATS mentor page that also lists her Transluce title (MATS). Her personal page does not mention Transluce, so the MIT line is her own statement of an affiliation held alongside the nonprofit role, not a dated confirmation of both.

• Prior work. Her site lists FIND (NeurIPS 2023), a benchmark of functions that resemble subcomputations in trained networks; Multimodal Neurons in Pretrained Text-Only Transformers (ICCV workshop 2023); and MILAN (ICLR 2022), which labels neurons with natural-language descriptions. She states her subject as “My work investigates representations underlying intelligence in artificial (and previously, biological) neural networks” (Sarah Schwettmann).
• MAIA. Berkeley Statistics’s announcement of the founding credits her with building MAIA, “the first large-scale pipeline using interpretability agents,” which underlies Transluce’s tools (UC Berkeley Statistics).
• The o3 report. Last of five authors, and one of two correspondence contacts, on Transluce’s report from pre-release testing of OpenAI’s o3: “o3 frequently fabricates actions it took to fulfill user requests.” The acknowledgments: “We are also grateful to have received early o3 access from OpenAI, without which this work would not have been possible” (Transluce, Apr 16 2025).

• Placement beside the CEO. Steinhardt is professor, CEO and board member; the About page lists him there as “Co-Founder & CEO, Transluce.” The September 16, 2026 embedded-evaluations post is by Steinhardt alone (Transluce). She holds no separate board listing on the About page that the fetch returned, which is an observation about the page, not about her role.
• Who pays, on Transluce’s own policy page. The Independence and Transparency Policy (updated August 27, 2026) states: “Transluce received 6% of its revenue from the personal holdings of OpenAI employees” and “32% of its revenue from the personal holdings of Anthropic employees,” and that “It received no revenue directly from AI developers as organizations.” It adds that Transluce “has carried out no paid evaluations” and that it “recuses any employee with a significant financial interest” in a provider from evaluating that provider (Transluce policy). The About page says only “financial support from a variety of individuals and philanthropic organizations” (Transluce, About).

Social Persona / Impression Management

Immediate impression: A research scientist’s page: a course she designed and teaches at MIT, a list of papers, no commentary. The public voice of the lab belongs to its CEO.

Energy: Technical and low-profile. Her name appears in bylines and correspondence lines rather than in announcements.

Impression management strategy: Credentials as instrument. The lab’s case is that “Companies building AI systems cannot be the primary arbiters of their safety” (Transluce); the neuroscience-to-interpretability record is the evidence that the people saying so can build the tools. Whether she manages impressions at all is not visible from the record, which is its own finding about the role.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN):

Dark Triad:

MBTI: INTP (“The Logician”) — dominant introverted thinking, auxiliary extraverted intuition. Takes a system apart to see what each unit does, which is the stated method of her papers. Observed from the documented role only.

Threat Assessment

Calibration: Steinhardt, the CEO of the same lab, scores 6/7/6 and hazard 50 and Adam Gleave, who heads a rival evaluator and sits on METR’s board, 6/7/6 and 55. The subject holds the same lab without the CEO’s board seat, public voice or professorial platform, so each stat sits one lower and hazard at 43. No first-hand receipt supports a higher number.

Alignment Analysis

Stated alignment: “Companies building AI systems cannot be the primary arbiters of their safety,” with open tools so outsiders can check the work (Transluce).

Observed alignment: Consistent. The lab published an unflattering pre-release finding about a model from the lab that supplied access, and published a funding breakdown that names OpenAI and Anthropic staff as donors.

Gap assessment: None in aims. The record cannot settle, and for a technical co-founder does not need to settle, how much access-dependence shapes what gets tested. Transluce’s own policy says it has “carried out no paid evaluations” and recuses conflicted staff. The structural point sits with the lab, not with her.

Convergent Drive Classification

Self-preservation: Keeps the MIT research post alongside the nonprofit role. Goal preservation: Understanding what happens inside a network, the subject of every paper on her page. Resource acquisition: Model access from labs, and philanthropy including, per the policy, employees’ personal holdings at OpenAI and Anthropic. Self-improvement: From neuron labelling (MILAN) to agents that run the investigation (MAIA) to a lab that runs it as a public service.

Subject is not an AI system. The drives appear anyway — in the scientist whose instrument needs the subject’s permission to point it at the subject.

Sources: Sarah Schwettmann — personal site; About — Transluce; Introducing Transluce; Independence and Transparency Policy — Transluce; Investigating truthfulness in a pre-release o3 model — Transluce; Some Focus Areas for Embedded Evaluations — Transluce; Steinhardt Announces Co-founding of Transluce — UC Berkeley Statistics; Sarah Schwettmann, Jacob Steinhardt — MATS.