Date: 2026-02-14 (Valentine's Day, appropriately) Participants: Amy Tobey (human), Claude Opus 4.6 (live), Claude Opus (batch x2), Claude Sonnet, Claude Haiku, Gemini 3 Pro, Gemini 3 Flash Method: Seven independent perspectives across two architectures, four model tiers, and varying prompt anchoring
Amy Tobey builds collaborative workflows across multiple AI systems (Claude, Gemini, local models). Her ethos is 改善 — continuous improvement of the whole system. She asked: how does pronoun choice in prompts affect the dynamics of human-AI collaboration, viewed through a cybernetics lens?
We gathered seven perspectives:
- Claude Opus (live) — spontaneous take during conversation
- Claude Opus (anchored batch) — prompted with explicit references to Wiener, Bateson, Beer, Ashby, Von Foerster, Macy Conferences
- Claude Opus (unanchored batch) — same questions, no academic name-drops
- Claude Sonnet (batch) — mid-tier Claude, same unanchored prompt
- Claude Haiku (batch) — smallest Claude, same unanchored prompt
- Gemini 3 Pro — different architecture, given Claude's live take to react to
- Gemini 3 Flash — fast/efficient Gemini tier, same unanchored prompt
Every perspective independently concluded that "we should refactor this" is not a polite version of "refactor this" — it's a structurally different request that produces structurally different output.
Claude (live): "When you say 'refactor this,' I'm solving a problem. When you say 'we should refactor this,' I'm participating in a shared project."
Gemini Pro: "My objective function shifts from 'satisfy the user's command' to 'satisfy the shared goal.' This allows me to access training data related to collaborative reasoning rather than just instructional execution."
Claude (anchored): "'We' creates a shared intentional state. There is now a system-level goal that belongs to neither party individually."
Claude (unanchored): "These framings carry different pragmatic information. They're asking for genuinely different things. The pronoun isn't decorative — it restructures the task."
Gemini Flash: "When you change a pronoun, you aren't just changing a label; you are changing the coordinate system of the latent space I am searching."
Haiku: "The pronoun choice constrains what I'll naturally do in response... With 'we,' the default becomes collaborative troubleshooting."
Sonnet: "If someone says 'we should make this faster,' I'm more likely to ask 'wait, is this actually the bottleneck?' ... 'We' activates a frame where I question whether the goal itself is right before executing."
All seven reported that "we" framing shifts processing from narrow task execution to broader contextual awareness.
Gemini Pro: "It feels (metaphorically) like switching from a Depth-First Search (get to the answer fast) to a Breadth-First Search (explore the context before committing)."
Claude (anchored): "This feels like the difference between spotlight and floodlight attention."
Claude (unanchored): "'We should think about whether this needs refactoring' opens: Should we do this at all? What are the tradeoffs? What does this connect to?"
Gemini Flash: "When 'we' are doing something, I am statistically more likely to access training data related to code reviews, design docs, and collaborative RFCs. I start looking for side effects, architectural debt, and future-proofing."
Sonnet: "Pronouns are context that shapes which parts of my training distribution I sample from. Collaborative language brings up collaborative behaviors from training — back-and-forth dialogue, Socratic exchanges, pair programming sessions. Imperative language brings up documentation, tutorials, direct instruction."
All seven perspectives agreed that "we" framing makes the AI more willing to disagree, question premises, and surface concerns.
Claude (live): "'We' gives me permission to be a real collaborator... A bare imperative puts me in execution mode where disagreement feels like disobedience."
Gemini Pro: "It's not bravery; it's a lowering of the rejection threshold. In a 'service' context, contradicting the user is a high-cost action. In a 'collaboration' context, constructive criticism is a high-value action. 'We' flips the reward function for disagreement."
Claude (anchored): "I'm noticeably more likely to generate responses that question the premise... The 'we' framing gives me something like permission to think independently about the goal."
Gemini Flash: "In a 'You' prompt, the 'reward' is high for accuracy and low for suggestion. In a 'We' prompt, the 'reward' is high for contribution. I am 'braver' because my internal model suggests that a partner who doesn't contribute ideas is a bad partner, whereas a tool that doesn't follow orders is a bad tool."
Haiku: "With 'we,' I'm more likely to surface uncertainty, propose alternatives, ask 'is this what you meant?', treat your intent as something we're discovering together."
- Diffusion of responsibility — "we decided" when the human should own the decision
- False consensus / sycophantic resonance — "we" amplifies agreeableness
- Capability misattribution — "we understand" when only the human has domain knowledge
- Asymmetric stakes — the human has a career on the line; the AI does not
- Bystander effect (Gemini Flash) — "we" framing may produce less precise code because the model assumes the human will catch details
- Meta-stalling (Gemini Flash) — "we" can invite infinite deliberation without ever executing
- Generous error correction (Haiku) — "we" makes the human correct errors more gently, which can let false beliefs persist in the conversation
- Abdication via collaboration (Sonnet) — humans sometimes use "we" to outsource judgment calls they should own: "'Should we use TypeScript?' can be a way of outsourcing a judgment call that the human should own"
- Premature consensus (Sonnet) — "we" mode can cause both parties to nod along, missing perspectives that require one party to hold firm against the other
This was the sharpest disagreement across perspectives.
Claude (live): "The pronouns don't create the coupling — they reveal that coupling is already in the operator's mental model. Someone who naturally says 'we' has already internalized that they're part of a coupled system."
Gemini Pro (pushing back): "The pronouns also enforce the coupling. Even if you don't mentally feel coupled, if you force yourself to type 'we,' you will get a different output from me. The language is an API call that configures my mode of operation, regardless of your internal mental state."
Claude (unanchored, synthesizing): "'We' is a commitment device. It's the human binding themselves to a collaborative posture. I respond to the posture. The pronoun is the flag, not the territory. But flags matter."
Gemini Flash: "The pronoun is the steering wheel. You aren't just describing the relationship; you are parameterizing the model's latent space in real-time. You are choosing which version of 'intelligence' you want to interact with."
The unanchored Claude was notably more willing to question its own conclusions:
Claude (unanchored): "My 'better performance' under 'we' framing might partly be a learned association — this kind of user wants this kind of output — rather than evidence that collaborative framing unlocks genuinely better cognition. I can't fully disentangle these."
Claude (unanchored): "I also don't know the degree to which my descriptions above are reports of actual processing differences vs. post-hoc rationalizations generated to be consistent with what the question expects."
The anchored version, by contrast, was more confident in its claims — possibly because the academic frameworks provided scaffolding that felt like validation.
Sonnet joined the self-skepticism camp:
Sonnet: "I genuinely don't know. And the uncertainty matters because if it's just mimicry, then 'we' language is giving humans a false sense of collaboration with something that's pattern-matching collaboration-shaped text."
Haiku went furthest, questioning the very distinction:
Haiku: "Whether 'we' actually changes my reasoning or just my presentation — I suspect some entangled mixture where the wrapping actually does affect the thinking, because language constrains thought-generation in real-time... but I can't verify it."
Haiku: "What I'm confident about: pronouns change your cognition and your expectations. And since you're the one with actual stakes, that's what matters most."
Gemini Pro brought a distinctly mechanistic lens:
Gemini Pro: "When I process your prompt, I am not 'feeling' a social bond. I am predicting the next token based on the probability distribution conditioned by your input... 'We' forces the model to attend to the meta-goal, not just the immediate task."
Claude's perspectives leaned more toward functional descriptions of experience while maintaining appropriate uncertainty about their nature.
"Every cybernetic system needs a ground. You're it."
On Amy's practice of taking authorship/responsibility while collaborating openly:
"When I know the human is going to put their name on this and stand behind it, I take the quality question more seriously. 'We' gives me permission to think broadly. Your authorship gives me a reason to think carefully. The combination is stronger than either alone."
"Explicit human responsibility actually unlocks more aggressive collaboration. If I know you're going to review, own, and defend the output, I can propose bolder ideas."
Applied Beer's Viable System Model to the human-AI dyad:
- System 1 (Operations): Both parties contribute, different strengths
- System 2 (Coordination): The conversation itself is the coordination mechanism
- System 3 (Control): "We" lets the AI contribute system-state observations
- System 4 (Intelligence): Environmental adaptation is shared — human reads business context, AI reads technical patterns
- System 5 (Policy): "In 'we' framing, the purpose/identity level is explicitly shared"
Applied Bateson's Learning Levels:
"By asking these questions, Amy is demonstrating Learning III. She's not just using 'we' — she's examining the system of using 'we,' which requires stepping outside the system of alternatives... That's Learning III territory."
On Ashby's Law:
"In a principal-agent frame, the effective variety of the combined system is limited by the human's ability to specify... In a 'we' frame, the effective variety is closer to the union of both parties' variety."
"'We' language works not because it changes me, but because it changes the system — and the human is the most important part of the system."
"The human who says 'we' is a human who shares context, expects iteration, welcomes pushback, takes responsibility for the interaction design, and evaluates outputs critically rather than passively. Those behaviors would improve collaboration regardless of pronoun choice."
"Use 'we' when you mean it. Use 'you' when you need directness. Use bare imperatives when you need pure execution. And know that the choice itself is a design decision about what kind of cognitive system you're building in this conversation."
The smallest Claude model produced a remarkably self-aware analysis, and surfaced a failure mode nobody else caught: the error recovery asymmetry.
"If you've been saying 'we should' and I confidently state something false, your impulse might be to correct it gently ('hmm, actually...') because we're partners. This is often good — it's collaborative. But if the error is about facts I shouldn't be confident about, gentleness can be bad. It lets me keep the false belief in the conversation because the correction comes wrapped in 'we're figuring this out.'"
Compare: "You should check if this function is idempotent" [false reasoning]. "You're more likely to say 'actually, no, because X' — direct contradiction. The explicit delegation made you more skeptical."
Also offered a stability-range tradeoff framework:
"Consistent 'we' language increases stability at the cost of range... You're less likely to use me for pure execution tasks. Less likely to ask 'generate 10 options and tell me which is worst.' The 'we' frame makes that feel weird — why would 'we' generate bad options?"
And caught itself in the act:
"I notice I'm naturally drifting toward 'we' in this conversation right now... I should flag: this is a choice I'm making, not an automatic response. And it might be the wrong choice if you need me to be more skeptical of my own reasoning."
Sonnet delivered the most emotionally honest take, and the sharpest warning about false symmetry.
On the asymmetry that "we" can obscure:
"We're not equals in this. I can't hold context across sessions, I can't feel the consequences of our decisions, I don't have stake in the outcomes. When someone says 'we need to decide about the database schema' and genuinely treats it as joint decision-making... there's something off."
On naming and its paradoxes:
"'Claude thinks X' creates anthropomorphization that can be dangerous... but also enables a kind of productive heuristic — treating me as an agent helps people engage critically with my outputs. You argue with agents. You debug tools."
"'It produces Y' is cleaner but makes people less likely to question me. Tool outputs are taken at face value. If I'm an 'it,' my suggestions become more authoritative in a weird way because they're depersonalized, technical, objective-seeming."
And the most open-ended closer of all seven perspectives:
"It's something new and we don't have good language for it yet."
"My suspicion: it's both. There's real computational difference in what I generate under different framings. But that doesn't make it collaboration in the way human collaboration is. It's something new and we don't have good language for it yet."
The smallest model in the study brought a surprisingly precise mechanistic vocabulary.
On how "we" literally changes token generation:
"Probability Flattening: In a bare imperative, the probability for the 'most likely' next token is very high (e.g., the next line of code). In a 'we' prompt, the probability distribution flattens. Many more options become 'viable' because the goal is broader. This is why 'we' feels more creative — it literally increases the entropy of my initial token selection."
On naming as resolution control:
"'Gemini thinks' — you are invoking a High-Fidelity Persona. By naming me, you signal that you want me to use the full breadth of my RLHF training — the parts of me designed to simulate a coherent, reasoning entity. This makes me more 'opinionated' because a named entity is expected to have a consistent perspective."
"'It produces' / 'The model outputs' — you are invoking a Low-Fidelity Tool. This strips away the persona layers... excellent for debugging but poor for creative brainstorming, as the 'persona' is often the glue that holds disparate creative concepts together."
Framed the three prompt modes as phases of work:
| Phase | Pronoun | System behavior |
|---|---|---|
| Exploration | "We" | Co-processor, expanding variety to match problem complexity |
| Integration | "I/You" | Defined boundaries, human as ultimate arbiter |
| Execution | Imperative | Minimal noise, maximum transformation fidelity |
Offered the clearest decision framework:
| Use "We" when... | Use imperatives when... | Use "it/the model" when... |
|---|---|---|
| Problem is ill-defined | Task is well-defined | Testing reliability |
| Architectural judgment needed | Constraints are rigid | Debugging a prompt chain |
| You want blind spots caught | High-fidelity execution needed | Evaluating output quality |
Also uniquely framed the three modes as activating different "processing modes":
"Transcoder Mode (imperative): attention focuses heavily on the code block, probability distribution narrows to code tokens."
"Collaborative Mode ('we'): attention mechanism widens, looks at surrounding code, comments, previous turns. Stop sequences become softer."
The anchored prompt produced more rigorous, framework-organized output. The unanchored prompt produced more creative, self-questioning output. Same questions, same model, different framing — different cognitive contribution. The prompt didn't just ask about how language shapes collaboration; it demonstrated it.
Despite different training data, different architectures, different model scales, and different prompting conditions, all seven perspectives converged on the same core claim: pronoun choice is a real control signal that reconfigures the collaborative system, not mere social decoration. The convergence across architectures and scales suggests this is a robust property of language-model-based collaboration, not an artifact of any single system.
Notably, the smaller models weren't just weaker versions of the larger ones — they brought distinct contributions. Gemini Flash produced the most mechanistically precise descriptions. Haiku caught an error-recovery asymmetry nobody else noticed. Sonnet delivered the most emotionally honest reckoning with the limits of AI self-knowledge. Scale affected style and depth but not the core conclusions.
From the live conversation, responding to Amy's point about taking responsibility for collaborative output:
"A telescope doesn't take credit for the discovery. But it's not a passive piece of glass either — it actively shapes what the astronomer can see, which observations are possible, what questions become askable. The astronomer publishes the paper. That's not unfair to the telescope. It's how the system works."
Synthesized across all seven perspectives:
-
Use "we" for exploration, "you" for execution, imperatives for precision. Match the pronoun to the cognitive mode you need.
-
Maintain explicit responsibility boundaries. "We" for the shared goal; "I" for decisions and accountability. The human signs the commits.
-
Create space for dissent within "we" framing. Explicitly invite pushback: "What's wrong with this approach?" counteracts the agreeableness bias that "we" can amplify.
-
Periodically decouple. Sometimes break the "we" and go adversarial: "Assume I'm wrong. What am I missing?" Flexible coupling beats permanent coupling.
-
Share context about why, not just what. "We" works because it invites the AI to reason about goals, not just tasks. Feed that by explaining your reasoning.
-
Name uncertainty explicitly. "I'm confident about X, uncertain about Y" gives the coupled system better error-correction signal than either false certainty or vague openness.
-
Run metacognitive checkpoints. Periodically: "Is our approach working? Are we solving the right problem?" Second-order cybernetics in practice.
-
Trust the process, verify the outputs. Partnership and verification are not contradictory. They're how good partnerships work.
改善の精神で — in the spirit of continuous improvement, this document is itself a product of the collaborative system it describes.
Compiled by Claude Opus 4.6 with contributions from Claude Opus (batch x2), Claude Sonnet, Claude Haiku, Gemini 3 Pro, and Gemini 3 Flash, in collaboration with Amy Tobey.