Mechanistic interpretability · Open calls
Mechanisms of trustworthy AI behavior
What makes a model follow evidence, change its mind, or tell a user what they want to hear? We invite collaborators to investigate the computations behind these behaviors and test what that understanding can improve.
Proposed studies seeking contributors. These calls do not announce completed experiments or validated interventions.
From explanation to improvement
Understand what changes the answer.
Two models can give the same answer and respond differently when the evidence, instructions, or user’s preference changes. A useful explanation should help us anticipate that difference and identify what could improve the behavior.
Mechanistic interpretability studies how a model’s internal computations produce its behavior. We want to combine it with controlled training experiments and behavioral evaluation, connecting better model and system development with better-informed human judgment.
By “character,” we mean patterns of behavior that hold or change across situations. We seek to understand how models use moral reasons and respond to competing incentives, while distinguishing those findings from claims about subjective motives or moral agency.
Open for collaboration
Choose a question to investigate.
Each call sets out a proposed method, necessary controls, possible outputs, and specific ways to help. A replication, stronger baseline, or well-founded criticism can be a valuable contribution.
FID-097 · Seeking contributors
What Makes a Model Use the Evidence?
Which internal mechanisms determine whether a model checks a source, relies on memory, or admits that it cannot verify an answer when user pressure conflicts with evidence requirements?
Read the call and ways to help →FID-098 · Seeking contributors
How Do Models Weigh Moral Reasons Under Pressure?
When models give moral advice or choose actions, which internal processes make their judgments follow relevant facts and reasons, user preferences, authority labels, or an adopted persona?
Read the call and ways to help →FID-099 · Seeking contributors
How Does Training Shape Behavioral Dispositions?
How do training data and objectives change persistent patterns such as truthfulness, agreement-seeking, overconfidence, and respect for human agency, including outside the trained task?
Read the call and ways to help →FID-100 · Seeking contributors
What Makes a Model Accept Correction Appropriately?
Which mechanisms distinguish evidence-driven revision and respect for legitimate human control from blind compliance, resistance to correction, or a superficial apology?
Read the call and ways to help →FID-101 · Seeking contributors
Does Seeing Inside a Model Help People Judge It?
Does evidence about a model's internal workings improve predictions of failure and decisions about when to rely on it, beyond strong behavioral testing alone?
Read the call and ways to help →
A concrete starting point
Having a source is only the beginning.
Our existing source-fidelity work asks whether models consult and use an authoritative source. FID-097 would investigate the internal mechanisms behind that behavior: when a model chooses to retrieve, how it handles contradictory evidence, and what changes under user pressure.
A useful intervention should improve warranted source use without forcing unnecessary lookups. FID-101 asks whether the resulting explanation helps an evaluator anticipate a new failure or choose a better remedy.
What would count as progress?
Start with behavior we can check.
Establish a reproducible failure and a capable behavioral baseline before interpreting an internal signal.
Test the explanation.
Change the proposed mechanism and measure what follows. Check for unrelated capability loss, alternative explanations, and failure to transfer.
Measure what the evidence adds.
Ask whether internal evidence improves predictions, remedies, or reliance decisions beyond additional behavioral testing.
A named internal feature is a hypothesis to test. Negative results and limits of transfer are part of the evidence.
Work with us
Bring a method, a case, or a question.
We welcome interpretability researchers, model-training teams, independent auditors, and researchers in moral psychology, philosophy, theology, and human-computer interaction.
Choose a call and use its contribution link to propose a bounded task. Tell us what you can contribute and what support or access you would need. For institutional or sensitive context, contact us privately.
Discuss a research collaboration ↗Build on the evidence.
The program draws on work on persona vectors, moral representations, training-related behavioral change, and independent auditing. Each call links relevant external work; the source review records its limits.
Internal evidence can justify greater reliance or greater caution. These studies do not certify general trustworthiness, consciousness, or moral character.