Peer agreement is much better at knocking a correct model off course than repairing a wrong one. Social influence is easier to use for harm than for correction.

Most measured “peer pressure” survives after the peer is removed. A bare candidate answer produces much of the same effect, showing that many conformity benchmarks confound social influence with contextual suggestion.

Can a model tell helpful revision from harmful conformity? Our probes and steering methods predict whether a model will change its answer, but cannot separate corrections from mistakes.

More reviewers do not automatically make a safety system safer. One shared misleading cue pushes reviewers in the same direction, and majority voting amplifies the error instead of cancelling it.

An LLM judge is not a neutral instrument for measuring LLM conformity. The evaluator responds to the same social cues it is supposed to grade, carrying its own bias into the result.

A conformal coverage guarantee assumes the model still scores answers the way it did at calibration time. Unanimous peer pressure changes that scoring behavior instead of the underlying questions, so calibrated 90% coverage collapses to 74%, and an adversary targeting low-confidence items can push a subgroup down to 47%.

Uncertainty is not one thing. We built evidential models that separate conflicting evidence from unfamiliar inputs, helping a system distinguish a hard case from one outside what it knows.

Measuring uncertainty matters only if the system acts on it. We train text classifiers to recognize the limits of their competence and abstain on inputs that would otherwise fail silently.

Models can recognize the broad category yet fail on an unfamiliar subtype. Accuracy drops sharply while confidence barely moves, leaving models overconfident exactly where deployment becomes unfamiliar.

A safety classifier can be accurate and still be dangerously overconfident. Safety-Flag puts seven moderation benchmarks behind one protocol and tests whether LLM moderators know when their judgments are unreliable.

We built a model that turns real documents into convincing decoys while removing their sensitive meaning. In human evaluation, readers could not reliably distinguish our generated research abstracts from real ones.

A multi-agent system can fail even when no agent misbehaves. When communication drops an “unverified” label, downstream agents can mistake an unchecked claim for an approved one.

A harmful outcome can be assembled from actions that are harmless on their own. We show why local monitors have a fundamental blind spot when the risk exists only in the composition.

Can a distributed attack be stopped before the agents finish assembling it? Across 2,082 runs, our detector catches 99% of successful attacks with a median of five steps still remaining.

We built a language model specifically for political conflict and violence. Pretrained on 34 GB of conflict news and government reports, ConfliBERT substantially improves political-science NLP tasks.

Political event coding normally requires large labeled datasets for every new question. We let models use the same natural-language codebooks as human coders, enabling zero-shot classification without task-specific training examples.

Better codebooks improve accuracy—but not necessarily rule following. LLMs can produce better social-science labels while still ignoring the coding logic they are supposed to follow.

Crowds do more than label misinformation. Across 88 studies, we show how people serve as annotators, evaluators, and creators—and when each role actually helps platforms fight misinformation.

Safe dosing requires more than knowing the drug facts. We test whether LLMs can track what a person has already taken—and abstain when another dose cannot be safely determined.

The quality of an LLM’s health answer depends on the language used to ask the question. Reliability is not a fixed property of a model; it also depends on who is using it.