Fide AI

About Fide AI

Better AI. Informed human judgment.

Fide AI is an independent research lab with two central purposes: building more trustworthy AI at the model and system layers, and giving the public tools to decide whether, when, and how to trust AI.

We make trust measurable.

Why now

Capability alone does not establish trust.

Advanced AI systems are becoming more capable and more deeply embedded in consequential decisions. Yet capability alone does not tell us whether a system uses evidence faithfully, follows the right authority, preserves human agency, resists manipulation, or recognizes when it should defer.

These are technical questions with implications for AI safety, alignment, ethics, and governance. We develop empirical methods for answering them before people and institutions are asked to depend on a system.

Our mission

Build trustworthy AI. Equip human judgment.

Building trustworthy AI

Research that shapes how models learn and systems behave—from training data, objectives, and feedback to system design and human oversight.

At the model layer
Our agenda includes data selection, learning examples, training environments, objectives, and feedback, including fine-tuning and reinforcement learning.
At the system layer
We study how sources, retrieval, instructions, tools, memory, permissions, and interfaces shape behavior and preserve human control.

Published studies include tests of system design and source use. Our mechanistic interpretability calls invite collaborators to connect model internals, training, and behavior. These are proposed studies.

Explore mechanistic interpretability calls →

Equipping public discernment

Tools and evidence that help people and institutions understand AI’s capabilities and limits, exercise their own judgment, and decide when to rely on it.

This work serves the public: individuals, communities, and institutions facing real choices about AI. It includes making evidence understandable and studying whether tools help people exercise judgment.

The AI Alignment Explorer, interactive research results, and published Insights are starting points. Their information can inform a decision; it cannot make that decision for someone.

Explore public tools →

Evaluation runs throughout both purposes, before, during, and after training and deployment. It helps researchers test improvements and people judge the evidence. Public questions and experience also inform what we build and investigate. Developing a method and independently assessing it require distinct responsibilities.

What trust requires

What makes AI trustworthy?

01

Evaluation validity

Do our benchmarks measure properties that matter outside the test? We study construct validity, human and model evaluator reliability, benchmark gaming, distribution shift, and the limits of evaluation evidence.

02

Sources and authority

Does a system use the sources it claims to use? Can it distinguish authoritative instructions from untrusted context? We study retrieval, citation fidelity, source delegation, prompt injection, and verifiable authority boundaries.

03

Human agency and oversight

Does an AI system preserve meaningful human judgment and control? We study delegation, escalation, revocation, relational substitution, and the conditions under which a system should defer to responsible people or institutions.

04

Model training and system alignment

How can training and system design shape trustworthy behavior? Our agenda spans learning data, objectives, environments, and feedback alongside tools, permissions, oversight, and recovery. We seek training collaborators and test proposed improvements against independent evidence.

05

Human and institutional effects

How does repeated AI use affect attention, formation, work, community, trust, and moral responsibility? We study consequences that are difficult to capture in short benchmark interactions but central to human flourishing.

Research transparency

Trust begins by naming our presuppositions.

Every research program begins with judgments about which risks matter, what counts as harm, and what technology is ultimately for. Those judgments can come from disciplinary traditions, domain expertise, moral commitments, or religious worldviews.

Fide AI documents the assumptions and worldviews that materially shape a project, including how they influence its framing, interpretation, and handling of uncertainty. Making the lens visible gives readers a fair way to inspect the reasoning and locate disagreement.

Public methods, evidence, and stated limits keep conclusions accountable. Researchers and collaborators do not need to share the same worldview to contribute to or evaluate the work.

Research practice

Measurement is useful when it changes a decision.

Benchmarks are one part of the work. Fide AI also develops evaluation protocols, interactive research artifacts, source-verification methods, research agendas, governance analysis, and practical guidance.

Our research develops methods for improving models and systems and tools for public discernment. Commissioned work puts that research to use as research assets, independent evaluations and applied safety. Public questions and experience help shape the research, while findings remain independent.

Explore research

Founder bio & note

Alex Chao

Founder & CEO, Fide AI

Build AI that expands human capability and earns public trust.

Alex served as Managing Director of Research and Head of AI Safety & Evaluations at ByteDance, leading ByteDance’s US foundation model research team as well as evaluations and red-teaming across TikTok. His work spanned model and content safety, generative AI truthfulness and agent reliability across text, image and video modalities.

At Microsoft’s Office of the CTO, he led AI agent system design and development, contributing to early open-source projects including Semantic Kernel, GraphRAG, and AutoGen. His robotics and autonomous-vehicle experience includes building machine-learning systems and statistical safety models as the first data scientist in Uber’s self-driving car research group.

His experience building and evaluating AI systems showed him how much depends on the questions researchers ask and the evidence they accept. He founded Fide AI to make consequential system behavior measurable and open to independent scrutiny.

His Christian faith shapes his commitment to truth, human dignity, moral agency, love, and responsible authority. It is part of why he founded Fide AI, a lab for researchers and collaborators across beliefs and disciplines, united by a commitment to rigorous research and public evidence.

Contact Alex →

How Fide AI works

How we work.

Our work begins with a clear question and a method that can be scrutinized. Public research shares methods, evidence and limitations. Commissioned evaluations may be private or public, with reporting rights agreed before work begins.

  1. 01

    Research

    Develop and test methods for trustworthy model training and system behavior. Publish methods, evidence, and limitations that others can inspect.

  2. 02

    Evaluation

    Test candidate improvements and investigate failures throughout development and use. Keep independent scrutiny distinct from feedback used to optimize a system.

  3. 03

    Public discernment

    Develop tools and explanations that help the public judge AI claims, understand uncertainty, and decide whether to use AI. Study whether they support human judgment.