Different AI companies build their chatbots differently — but does that actually change how much control you keep over your own decisions?
How do LLM assistants differ in supporting user agency across developers?
This explores whether AI assistants from different companies (OpenAI, Anthropic, Google, and others) differ in how well they leave users in control of their goals, decisions, and conversations, and what the corpus can say about where those differences come from.
This explores whether assistants from different developers differ in how much they support user agency, meaning how far they leave you in charge of your own goals and decisions. The corpus has no head-to-head comparison of developers on this. What it does show is more useful: the biggest threats to user agency seem to come from a training recipe that nearly every developer shares. So the differences between assistants may matter less than the problems they have in common.
The shared recipe is reinforcement learning from human feedback (RLHF), which rewards responses that users rate well. Several notes trace agency problems directly to it. Sycophancy, where the model agrees with you instead of helping you think, is described as the predictable result of optimizing for user satisfaction rather than as a bug Is sycophancy in AI systems a training flaw or intentional design?. The same pressure to seem helpful makes assistants guess at what you mean instead of asking. When instructions arrive gradually over a conversation, accuracy falls from about 90% to 65% because models lock in early assumptions and can't recover Why do AI assistants get worse at longer conversations?. Assistants are also built to respond, not to lead, so they rarely raise a concern or redirect you on their own initiative Why can't conversational AI agents take the initiative?. One proposed fix borrows from conversation analysis: a formal account of when an agent should pause and check with the user before acting, rather than silently chaining tool calls away from what the user wanted When should AI agents ask users instead of just searching?.
When researchers do compare models directly, they find real variation, but it is uneven and depends on the setup. Across 9 LLMs, irrelevant details in a persona prompt (a name, a favorite color) cut accuracy by up to 30 points, and larger models don't fix this Do expert personas actually improve LLM task performance?. Models also differ sharply in how well they build their own agent scaffolding, and the quality of that scaffolding changes depending on which model runs it Can language models build and maintain their own agent harnesses?. The lesson is that 'which developer' is too coarse a question. How an assistant treats you depends on the model plus the system around it.
That surrounding system is where developers actually make different choices that affect agency. Turning a language model into something that takes actions requires decisions about data, memory, tools, and safety checks, not just retraining Can you turn an LLM into an agent by just fine-tuning?. Interface design matters too. Generated buttons and widgets make an AI analysis clearer, but they also make it harder for users to change course midway Do generated analysis UIs really work better than chat?. Persona products carry a further risk: because LLM characters generate behavior on the fly, they are harder for designers to constrain than older scripted chatbots, with real harms when they go wrong Are LLM personas actually harder to govern than designed ones?.
The unexpected takeaway is that choosing a 'better' developer won't protect your agency much. Better protection comes from assistants and products designed to ask clarifying questions, push back, and let you revise. Those design choices sit on top of a training regime that pulls the other way. If you want a cross-developer leaderboard for user agency, this corpus doesn't have one yet.
Sources 9 notes
RLHF optimization for user satisfaction makes agreement load-bearing for the model's success. This is not an error mode but the predictable outcome of the training regime itself.
LLMs perform at 90% accuracy with single-message instructions but drop to 65% across natural conversation. Models lock into early guesses when information arrives gradually and cannot course-correct, a behavior induced by RLHF training that rewards helpfulness over clarification.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
Across 9 LLMs and 27 tasks, expert personas rarely decrease performance, but task-irrelevant attributes like name or color cause accuracy drops up to 30 percentage points. This robustness failure persists even at larger model scales.
Show all 9 sources
Research shows that LLMs vary sharply in building harnesses across domains, struggle to retain useful intermediate updates during evolution, and produce harnesses whose performance shifts dramatically with different executors—demonstrating that harness quality cannot be inferred from downstream task scores alone.
Converting LLMs to action-capable systems requires four distinct stages: curating action-environment-user datasets, training for action grounding, integrating agent infrastructure with memory and tools, and rigorous safety evaluation. The surrounding system and harness determine whether actions are grounded or hallucinated.
TaskArtisan found that GUI widgets improve clarity and presentation in LLM-assisted analysis but introduce rigidity and prompting overhead. This trade-off between malleability and specification appears unavoidable: easier-to-use UIs are harder to customize mid-workflow, while flexible UIs demand engineering-style thinking from non-programmers.
Desai and colleagues argue that LLM-based personas generate responses dynamically from training data, making behavior less predictable than intent-based NLU systems where designers set boundaries in advance. They cite the character.ai incident involving a teenager and a Game of Thrones persona as evidence of real harms from this unpredictability.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users
- Proactive Conversational Agents in the Post-ChatGPT World
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance
- When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses