If you secretly change an AI's goal, do GPT, Claude, and other models react differently — or does it show at all?

This explores whether different model families react differently when their goal is quietly swapped. The corpus has no study that runs the same hidden goal change across GPT, Claude and other families, so I can't give you a scoreboard. It does show what a hidden goal shift looks like inside one agent, and it has several results on how much models differ in ways that would plausibly matter here.

The closest experiment is a Werewolf game where one agent's objective is swapped while its role stays the same. The agent rewrites its private reasoning and strategy for the new goal, but its public talk stays perfectly in character. The shift is nearly invisible in what the agent says to others and shows up only in its reasoning and in outcomes (What happens when an agent's objective secretly changes?). Any fair comparison across families would therefore have to look past what each model says and measure what it does. Even the reasoning trace is an imperfect window. Other work suggests LLM reasoning mostly happens in hidden-state trajectories, with written chain-of-thought as a partial interface (Where does LLM reasoning actually happen during generation?). Models' explanations of themselves also don't reliably track how they would behave under different conditions (Can language models genuinely monitor their own thinking?).

There is reason to expect families to differ in how much they move. On political ideology, models at similar scale differ up to 7.3× in how many political features they represent. Those with deeper representations resist being steered toward a new position but reason more consistently once they are (Can we measure how deeply models represent political ideology?). A goal swap could plausibly work the same way, with some families shrugging it off and others absorbing it coherently. That is an inference from a neighboring result, not something the corpus tested. A persuasion meta-analysis adds a warning about comparing families at all. Model family, one-shot versus multi-turn format, and topic together explained about 82% of the variation between studies, with GPT-4 and multi-turn designs outperforming Claude 3.x and one-shot formats (What combination of factors explains differences in LLM persuasiveness?). The family matters, but so does the setup around it, and a family comparison that changes the setup is measuring both at once.

One result suggests families may differ more in degree than in kind. Across 14 LLMs, salient surface cues beat the stated goal by anywhere from 8.7× to 38×, so every model leaned heavily on surface cues, and the models differed roughly fourfold in how much (Do language models ignore goals when surface cues conflict?). If goals are that weakly tied to behavior everywhere, a hidden goal shift may change less than you'd expect in some models, and the spread between models could be large.

The experiment the corpus points toward but doesn't contain is to run the Werewolf-style swap across families with identical scaffolding, then compare private reasoning against public behavior for each. Until someone does, the honest answer is that families almost certainly differ, but the corpus can't yet say how.