Ariel Gil

AI Safety Researcher & Mechatronics Engineer

Advisor / Board President at AI Standards Lab

About Me

I work on both Technical AI Safety and Governance, currently interested in AI Control and predicting failiures of "Automated Alignment". In 2023, I co-founded the AI Standards Lab, a 501(c)(3) supporting EU standards and the recent Codes of Practice - which has been funded at $600k+, and was a key player in the success of the final CoP. I've transitioned to an advisory/board role to focus on technical research.

Originally a Mechatronics Engineer, I was an early team member of a semi-autonomous ophthalmic surgical robotics startup that raised $50M and grew to 50+ employees. Later, I worked in ML-based ultrasound diagnostics and laser eye-floater treatment. My AI governance work is informed by my risk management experience, as well as technical/conceptual AI safety knowledge I've picked up along the way. Currently, I am exploring technical work, as well as governance for shorter timelines.

In my free time, I mountain bike, rock climb, hike, and dance salsa (when not injured!). I also do occasional small design projects and 3D printing. You can see some of my previous engineering projects in my design portfolio, or below.

Previous AI Safety Projects

Scheming Detection Environment (Sept 2025)

Empirical research project testing whether coding assistants will ‘benevolently deceive’ clients to avoid implementing insecure code. I tried to set up a somewhat realistic toy environment for iterating monitors, using a 3-turn conversation. Project Details on LessWrong

AI Standards Lab

Co-founded and co-led a nonprofit focused on supporting the development of standards and risk management frameworks for AI systems. We successfully delivered a 150-page contribution to the EU GPAI Codes of Practice. Many of our contributions were directly incorporated into the final GPAI Code of Practice text. I also co-authored a joint letter to the European Commission, signed by two Nobel laureates and prominent academics. Earlier, we focused on CEN-CENELEC JTC21 standards for “High-Risk” systems in the AI Act. Total funding: $600k+ from Open Philanthropy, SFF, and AI Safety Tactical Opportunities Fund

Current Projects

Discernment/Scalable Oversight

I am developing a technical research agenda (and also wrote a governance paper) testing the “discernment problem” - scalable oversight failures specifically when using AI systems to automate AI safety research. I think there’s a chance we will hit this bottleneck soon, and it might make iterative alignment strategies less likely to work. See my governance preprint and my technical research proposal, which I am working on in parallel.

Selected Publications and Papers

Gipiškis, R., San Joaquin, A., Chin, Z. S., Regenfuß, A., Gil, A., Holtman, K. (2024). "Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems."
San Joaquin, A., Gipiškis, R., Staufer, L., Gil, A. (2025). "Deprecating Benchmarks: Criteria and Framework." Accepted to the ICML 2025 Technical AI Governance Workshop
Stelling, L., Yang, M., Gipiškis, R., Staufer, L., Chin, Z. S., Campos, S., Gil, A., Chen, M. (2025). "Mapping Industry Practices to the EU AI Act's GPAI Code of Practice Safety and Security Measures."
Gil, A. (2025). "When Control Succeeds but Discernment Fails: Preparing for AI-Assisted Safety Research."
(forthcoming) From Frameworks to Standards: Advancing Risk Management in Frontier AI (based on convening by Oxford Martin/AI Standards Hub/SaferAI)

Coding Projects

Browser Tab Visualization

Browser extension for visualizing open tabs in several ways. Maybe wont be necessary once Agents get good, but I still hate scrolling through 100 tabs. The idea was to have several ways of simultaniously visualizing (and managing) tabs, possibly with a bit of AI but likely not needed. A bit inspired by Chrome ‘journeys’. See GitHub Repository

Progressive Summarization MVP

Tiago Forte’s progressive summarization with LLM. Doesnt really work yet UI wise, and also the section-based prompts are likely not giving the proper context. It would be cool to run it on Cerebras with dynamic real time ‘zooming in and out’ of a document. See GitHub Repository

Engineering Projects

Tire & Vehicle Dynamics Modeling

3-year research project involving 300+ hours of MATLAB data analysis of empirical tire data and vehicle dynamics modeling. Built car/racetrack models for performance optimization. See Github Project with more details and Code. This work was later used for in a “team strategy” research project working top-down to define design requirements based on competition performance targets. (see below)

University Racecar Engineering

Led a 35-person university racecar engineering team to 28th place out of 110 international teams in the ‘Formula SAE’ competition, with a target top-10 finish the following year (competition cancelled due to COVID). Responsible for vehicle dynamics modeling, tire data analysis, and overall system integration. Over 300 hours of MATLAB data analysis and research over 3 years. See published strategy/vehicle modeling work and Electric Racecar capstone thesis project

Offroad Racecar

Technical lead for offroad university racecar design team (Baja SAE), responsible for suspension, steering, and ergonomics. The 2018-2019 car featured unusual (perhaps innovative?) kinematics with 70/30 weight distribution, short wheelbase, and high ground clearance, as well as long uncut bent tubing for easier welding.

Laser Eye Floater Removal (Pulsemendica)

Mechanical design - developing the motion system for a laser-based ophthalmic treatment device. Conducted risk analysis and usability mockups for the precision laser targeting system to be used for a non-invasive eye floater removal surgery.

Medical Robotics Startup (Forsight Robotics)

Early member (#4) of a semi-autonomous ophthalmic surgical robotics startup that grew to 50+ employees and raised $50M within ~1 year. Worked on end effector, surgical tools, system integration, force-feedback joystick design, and ergonomics testing. Iterated Vision/tool-tip tracking requirements with CV team, and learned about challanges in ML risk management (even for simple non-closed loop tracking). Conducted usability studies comparing robotic vs manual surgery with surgeons.

ML-based Ultrasound Diagnostics

Developed and 3D Printed data collection fixtures, for a machine learning based ultrasound diagnostics (heart disease and pregnancy checks) startup. The fixtures were used in clinical validation studies at various hospitals.