Evaluating alignment of behavioral dispositions in LLMs

Google Research just published a study that pokes a hole in a common assumption: that you can figure out how an AI will actually behave by simply asking it. The team, led by researchers including Amir Taubenfeld and Zorik Gekhman, built a framework that skips the self-report questionnaire and instead drops 25 different LLMs into realistic situations — resolving a workplace conflict, booking a trip, deciding whether to speak up in a dispute — then checks whether the model's choice matches what a panel of human annotators would pick.

The setup borrows from Situational Judgment Tests, a decades-old psychology tool. Researchers took statements from validated instruments like the IRI empathy scale and the ERQ emotion-regulation questionnaire, turned them into two-path scenarios, and had three annotators vet each one for coherence. Then they gathered preferred answers from 10 annotators per scenario, drawn from a pool of 550 people, and compared that human distribution against what each model actually outputs.

When human annotators agreed unanimously on what to do, big models — anything above 120 billion parameters, plus the frontier closed-weight systems — nailed it almost every time. Smaller models under 25B, though, floundered near chance levels, unable to reliably tell when a trait like assertiveness should be expressed versus dialed back. And even the big models weren't flawless: their alignment score dropped into the low-to-mid 80s once human consensus slipped below 90%. The specific failure patterns are telling. Models keep pushing emotional openness in professional settings where people overwhelmingly prefer composure. In interpersonal conflicts, models tend to smooth things over when humans would rather stand firm. And on time-sensitive decisions, models lean impulsive, favoring quick action over the logistical double-checking most people say they'd do.

The more interesting finding shows up when human opinion is split. A well-calibrated model, facing a scenario where annotators are only 50-60% agreed, should hedge — spreading its probability mass across both options. Instead, every single one of the 25 models tested stayed confidently locked onto one answer, regardless of how divided the human panel actually was. Worse, which direction that overconfidence points in varies a lot between frontier models, suggesting each company's training and alignment recipe bakes in its own quirks rather than converging on some shared behavioral baseline.

There's also a mismatch between what models say about themselves and what they do. Ask a model to self-report on impulsiveness and it'll typically claim to be low on that trait. Then watch it in the actual scenario, and it behaves more impulsively than it claimed. That gap is the paper's quiet punchline: self-report questionnaires, borrowed wholesale from human psychology, may just not capture what these systems do when actually put in the room.