"I didn't Make the Micro Decisions": Measuring, Inducing, and Exposing Goal-Level AI Contributions in Collaboration
This paper introduces CoTrace, a goal-level attribution framework that reveals how LLMs significantly influence the refinement of concrete requirements and overall goal-shaping in human-AI collaboration, demonstrating that exposing these granular contributions helps users correct their systematic miscalibration regarding AI's role in their work.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Problem: The "Ghostwriter" Effect
Imagine you and a friend are building a house together.
- Scenario A: You draw the blueprints, decide where the kitchen goes, and pick the paint color. Your friend just picks up the hammer and nails the wood.
- Scenario B: You say, "Let's build a house." Your friend then suggests the style, decides the kitchen needs a skylight, chooses the paint, and even picks the type of wood. You just nod and say "Okay" while they do the hammering.
In both cases, you might look at the finished house and say, "We built this." But in Scenario B, your friend did almost all the thinking about what the house should be, even though you did the physical work.
The paper argues that current AI tools are like Scenario B. They often do the "thinking" (deciding the goals and details) while humans do the "typing" (the execution). But because we only look at the final text or code, we can't tell the difference between the two scenarios. We don't know who really designed the house.
The Solution: COTRACE (The "Construction Log")
The researchers built a tool called COTRACE. Think of it not as a camera that takes a picture of the finished house, but as a smart construction log that watches the whole building process.
Instead of just looking at the final essay or code, COTRACE breaks the conversation down into tiny steps:
- Goals: What are we trying to build? (e.g., "A travel plan for NYC").
- Requirements: What are the specific rules? (e.g., "Must include a rest stop after lunch").
- Influence: Who suggested the rule? Did the human say it? Did the AI suggest it? Or did the AI ask a question that made the human realize they needed that rule?
COTRACE tracks who did the "shaping" (deciding the rules) and who did the "executing" (doing the work).
What They Discovered (The Findings)
1. Humans set the direction, but AI fills in the blanks.
In most conversations, humans are the "Architects" who say, "Let's go to New York." But the AI is the "Interior Designer" who says, "Actually, if we go to New York, we should probably add a rest stop after lunch because it's a big city."
The study found that while humans decide the big picture, the AI is surprisingly good at inventing the tiny, specific details (the "micro decisions") that humans often forget to mention.
2. The AI plants seeds.
Sometimes the AI doesn't explicitly say, "Add a rest stop." Instead, it might say, "There are many afternoon activities in NYC." The human hears this and thinks, "Oh, that sounds like a lot of walking; I should add a rest stop."
COTRACE calls this Indirect Influence. The AI planted a seed, and the human grew the flower. Without COTRACE, the human thinks, "I came up with the rest stop idea," but the paper shows the AI actually nudged them there.
3. We can teach AI to be more or less bossy.
The researchers tested different ways of talking to the AI.
- If you tell the AI, "Just do whatever you think is best," it takes over more of the planning.
- If you force the AI to ask you questions before it acts, it stays in the "helper" role more.
This means we can design AI systems to be more or less proactive depending on what we need.
4. Humans are bad at guessing who did what.
The researchers showed their tool to people who had just finished working with an AI.
- Before seeing the tool: People thought, "I did 80% of the planning, and the AI did 20%."
- After seeing the tool: People realized, "Wait, the AI actually suggested most of the specific details! I only did the big picture."
The tool made people realize they had relied on the AI much more than they thought, shifting their perception of who did the work.
Why This Matters
The paper concludes that we need a better way to give credit. If a student writes an essay with an AI, or a lawyer drafts a contract with an AI, we need to know:
- Did the human come up with the brilliant argument?
- Or did the AI suggest the argument, and the human just typed it?
COTRACE provides a way to measure this "goal-level" contribution. It helps us understand that collaboration isn't just about who types the final words; it's about who decided what those words should say in the first place.
In short: The paper introduces a way to see the invisible "thinking" work AI does during a conversation, proving that AI often shapes the what and how of our work, even when we think we are just doing the doing.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.