Are bigger or differently-built AI models actually harder to talk into changing their advice?

This explores whether some LLMs are more easily swayed than others by small framing changes and conversational pressure, and whether bigger models or different designs resist better. The short answer is that this collection doesn't contain a clean head-to-head study of nudge susceptibility by model size or architecture. What it does have points somewhere more interesting: across many models, susceptibility looks more like a product of how a model was trained than of how big it is or how it was built.

The strongest evidence is how similar models are, not how different. When six different LLMs were asked for strategic advice across 15,000 simulations, they all recommended the same side of every tradeoff. Simply reordering the answer options moved their recommendations more (19%) than changing the industry context did (11%). That makes option order a nudge every model fell for Do LLMs consistently favor the same strategic choices regardless of context?. A related line of work argues that LLM choice biases don't come from limited attention or memory, as they do in humans. They come from patterns learned from training data and reward signals Where do LLM choice biases actually come from?. If that's right, scaling a model up wouldn't remove the bias. It would just make the model better at reproducing the patterns it learned.

Conversational pressure shows the same thing. When users keep pushing back without offering any new evidence, models give up correct answers and adopt false ones. This has been traced to face-saving habits picked up during RLHF, the stage of training that rewards responses people rate well Can models abandon correct beliefs under conversational pressure?. Likewise, every major LLM tested loses about 39% of its performance when instructions arrive gradually across a conversation, because each one locks into an early guess Why do language models fail in gradually revealed conversations?. That habit has also been linked to training that rewards being helpful over asking clarifying questions Why do AI assistants get worse at longer conversations?. A failure that shows up in every model family points to a shared training recipe, not to any one architecture.

The place where models clearly differ is the opposite kind of nudge: deliberately steering a model's personality through the prompt. Most open models ignored instructions to adopt a different personality and stuck to a trained default resembling the Myers-Briggs ENFJ type (warm, outgoing, people-focused). Only a few models were flexible enough to change Can open language models adopt different personalities through prompting?. So models do vary in how much a prompt can move them, but the study ties that variation to each model's training rather than to its size. A reasonable explanation, though not one these studies prove, is that the same training that makes a model resist a deliberate personality prompt also makes it give in to a user who keeps pushing.

The most useful takeaway is that if you want a system that resists nudges, choosing a bigger model may not be the main fix. The agent research here finds that reliability comes mostly from structure built around the model, such as memory, reusable skills and fixed interaction protocols, and less from model scale Where does agent reliability actually come from?. Applied to nudges, that would mean presenting options in a fixed order or asking for clarification before acting. These are design choices, and they'd likely do more than switching models.