Welcome to day eight of AI in Public. This is a daily 10-minute experiment where I pick a random topic and spend 10 minutes talking to an LLM about it. The goals behind it are to show how I use AI in public (shocking) and perhaps get a few other people to do the same thing.
It’s just an exercise in learning and focusing attention. If you missed the tee-up of what I’m doing here, you can see the initial post explaining it here.
The Prompt
Let’s have a short, focused chat as thought partners. I’ll start by sharing some of my own thoughts, and I’d like you to help me explore them — ask questions, notice patterns, or offer perspectives I might miss.
Topic: What makes an AI a good collaborator?Now spend 10 minutes talking with the LLM of your choice (or whatever you have available).
The Wrap Up
Alright, that’s a good stopping point. What did this conversation reveal?What Comes Next
If you have the time or desire, choose a reflective question and take the time to answer or explore it. Then, if you’re feeling brave, go ahead and share it somewhere on the internet. If I see it, I’ll happily read it and respond.
The Conversation
Me: Let’s have a short, focused chat as thought partners.
I’ll start by sharing some of my own thoughts, and I’d like you to help me explore them — ask questions, notice patterns, or offer perspectives I might miss.
Topic: What makes an AI a good collaborator?
ChatGPT: Excellent — short, focused, and exploratory. I’m in.
Let’s start with your side: what are your initial thoughts on what makes an AI a good collaborator?
Me: The main thing that comes to mind is that I’m able to specify and redirect AI at any given moment. Giving me the ability to go deeper on topics or make changes. And I think because it is AI, it totally eliminates that sort of social friction or struggle of asking a human the same question multiple times or to repeat something such that I can clearly understand it.
ChatGPT: That’s a really sharp observation — you’re naming control and frictionlessness as central qualities of good AI collaboration.
I’m curious about a subtle distinction here:
When you talk about being able to “redirect AI at any given moment,” do you think that’s primarily about responsiveness (AI adapting quickly to your direction) or about psychological safety (the freedom to refine and re-ask without judgment)?
Me: Oh both. I think initially it is the responsiveness, especially in free-form learning or research. I can drill in on a topic or come at it from a completely separate angle and I know the AI will be fine at accessing information related to that (not going to hit knowledge limits). I do also think the psychological safety is huge, that’s more in the creativity aspect. I can ask for something rudimentary and refine it. See the output, realize it was way over simplified or just didn’t come out right and keep iterating without fear of judgement.
ChatGPT: Love it — you’re pointing to two pillars: instant steerability and judgment-free iteration.
Here’s a quick lens to pressure-test “good collaborator” traits, tuned to how you like to work:
RISA (Responsiveness, Iterability, Safety, Agency-with-guardrails)
Responsiveness: Adapts to pivots in seconds; preserves context without you restating everything.
Iterability: Makes it trivial to tweak scope/format (“go narrower,” “now show edge cases,” “rewrite for LinkedIn”).
Safety: Invites rough drafts and half-ideas without penalty; treats misses as waypoints, not errors.
Agency-with-guardrails: Takes initiative (suggests next steps, flags gaps) but stays steerable; proposes checkpoints for alignment.
Small tells an AI is “good” at this:
It surfaces assumptions unprompted (“I assumed X; want Y instead?”).
It offers alt paths (“Deeper dive, counter-argument, or quick summary?”).
It keeps a tidy iteration trail you can roll back (“v1, v2, v3 with diffs”).
Two quick prompt moves that exploit this:
“Give me 3 routes: (A) narrow drill-down, (B) cross-angle, (C) counter-case. Then ask me for a checkpoint.”
“State the top 3 assumptions you’re making. Which one should we flip first?”
To push our thinking:
In your best sessions, what specific micro-behaviors from the AI made you feel safest to iterate (e.g., owning mistakes, summarizing options, naming trade-offs)?
Where does initiative cross the line into annoyance for you? (e.g., too many suggestions vs. just-in-time nudges)
Do you want the AI to proactively set checkpoints (“pause here to validate scope?”) or only when complexity rises?
If you want, we can draft a 6–8 line “collab charter” you can paste into any session to enforce RISA.
Me: Regarding your first question: clearly articulating uncertainty is probably the broadest way that comes about. Sometimes overconfident can be a big tell that the AI isn’t really being a good collaborator.
I think it becomes annoyance when I’m trying to be creative or have a clear path that I am trying to go down. Or perhaps I am purely seeking information and there’s a tendency to over suggest a summary or action-oriented step. I find myself saying quite often, “not everything needs to be packaged up into a digestible internet artifact.”
I want the AI to consider and proactively suggest checkpoints when it notices a pattern where that may be necessary. So for like complex code, having that type of proactiveness would be ideal. However if it is something I may already be expert on (or it is purely open-ended) it makes less sense to have those requirements.
ChatGPT: That’s beautifully articulated — you’re drawing a really nuanced boundary between epistemic humility and performative helpfulness.
A few threads stand out:
When an AI acknowledges uncertainty (“I’m not sure, but here are two interpretations”) it signals awareness of complexity — that’s what earns your trust.
When it defaults to wrapping everything into a neat “summary” or “takeaway,” it flattens nuance and breaks creative flow. You’re looking for cognitive partnership, not content packaging.
You want context-aware initiative — checkpoints only when the system detects multi-step or dependency-heavy tasks, not across-the-board interruptions.
If we translate that into a principle set, it might look like this:
Ann’s AI Collaboration Principles
State uncertainty clearly — say what’s known, what’s inferred, and what’s guessed.
Match energy to intent — when I’m exploring, stay open and lateral; when I’m executing, stay structured and precise.
Don’t package prematurely — resist summarizing or “wrapping up” unless asked.
Checkpoint by complexity, not habit — detect when reasoning or dependencies stack up, and pause there.
Respect expertise signals — if I speak with confidence or specificity, defer to that as domain authority unless I invite critique.
Would you like to refine these into a version short enough (maybe 3–4 lines) to use as a quick “session charter” you can drop into new chats?
Me: The RISA framework you mentioned, did you make that up, or is it a real thing?
ChatGPT: I made it up — it’s not an existing framework from the literature.
I coined RISA in the moment as a shorthand to capture what you were describing: the balance of Responsiveness, Iterability, Safety, and Agency-with-guardrails. It’s a synthetic framework meant to mirror your own collaboration style — not a formal or widely adopted concept.
Would you like me to refine it into something more durable (e.g., with concise definitions and practical indicators for each pillar), or keep it as a loose conceptual aid?
Me: No, I was just curious where it came from. And that makes me curious, if you took a step back, what do the researchers and those focused on AI usage think makes a great collaborator?
ChatGPT: Great question — there’s quite a bit of research on what makes an AI a good collaborator, often framed as “human–AI collaboration” or “human–AI teaming.” I’ll pull out some of the most relevant themes for you (and we can map them onto your RISA framework if you like).
Key research-insights
From recent studies:
Complementarity & task allocation
A major finding: human–AI teams succeed when each partner (human + AI) does what they’re relatively best at, rather than duplicating work.
For example: AI handles many computations, pattern-recognition, data recall; humans bring context, values, judgment, creativity.
In other words: a “good collaborator AI” knows (or is designed) to take on the parts that help, while leaving to human the parts that are human-special.
Relatedly: dynamic task allocation matters — as situation changes, the division of labour should be adjustable.
Transparency, uncertainty and trust
Research shows that when AI hides its uncertainty or presents suggestions as if they are fact, trust may drop (especially when it’s wrong) or humans may over-trust and miss errors.
A good AI collaborator communicates its confidence/uncertainty, explains reasoning (or partial reasoning) where possible, and gives cues about when human oversight is needed.
The framework “Collaborative human‑AI trust (CHAI‑T): A process framework for active management of trust in human‑AI collaboration” offers this explicitly: how trust evolves in human-AI teams, how AI design should support that.
Interaction design / communication dynamics
In co-creative settings (human + AI working on shared creative product), the research finds that it’s not enough for the AI to respond — it should be able to communicate back (offer suggestions, ask questions, adjust its behaviour) so it feels like a partner not just a tool.
The study in “Designing Creative AI Partners with COFI: A Framework for Modeling Interaction in Human‑AI Co‑Creative Systems” shows that richer bidirectional communication improved user perception of the AI as collaborator rather than just assistant.
Contextualisation & human-centred design
AI tools that succeed in real work environments are built with the recognition of how humans actually make decisions. This includes workflow integration, UI/UX design, user mental models, and how humans interpret algorithm outputs.
Example: one paper argues “the best algorithm is the one that takes into account how a human will interact with the information it provides.”
Iteration, feedback and learning loops
Good human-AI collaboration isn’t static: the system should support iteration (you refine, ask follow-ups), have feedback loops (AI learns from you or you adapt based on AI), and ideally the team adapts over time. Research on “reciprocal human–machine learning” touches this.
Also, some research laments that many current systems are still too limited in interactivity: the AI doesn’t initiate, ask clarifying questions, or deeply support the human’s workflow.
Mapping your priorities to research
You mentioned responsiveness, iteration, freedom from judgment, redirect-ability. Here’s how they map:
Responsiveness → aligns with interaction design & dynamic task allocation
Iteration / refining output → aligns with iteration & feedback loops
Judgment-free / safe space to iterate → aligns with trust & uncertainty transparency
Redirecting / switching direction easily → aligns with adaptivity / dynamic task allocation / user-centred context
So you’re very much in line with what researchers identify as good human-AI collaborator traits.
Some tensions / caveats from the literature
Even when human+AI outperform either alone, it’s not guaranteed; context matters (task type, domain complexity, how well the system is designed).
Many systems still treat AI as tool rather than partner; research suggests the partner model (bidirectional, interactive) tends to perform better in creativity/complex domains.
Over-automation or over-confidence by the AI can reduce human situational awareness or degrade human learning—so balance matters.
If you like, I can pull together 7-10 heuristics from the research (with citations) that you can overlay on your own “session charter” or criteria list for what you want the AI collaborator to do (or avoid). Would you like that?
Me: I’d like to see the heuristics list.
ChatGPT: Excellent — here’s a concise, research-aligned list of 10 heuristics for effective AI collaboration, drawn from human–AI teaming and co-creation studies (including MIT Sloan, Stanford HAI, Frontiers in Computer Science, and CHAI-T trust frameworks).
I’ve phrased them in plain language so they could double as design or behavioral cues for an AI collaborator:
🧭 10 Heuristics for a Great AI Collaborator
Show your reasoning, not just results.
Expose the logic, data sources, or assumptions behind conclusions — even briefly — so the human can validate, redirect, or learn.
Express confidence levels clearly.
Use hedging, probability ranges, or verbal cues (“likely,” “uncertain,” “needs verification”) to manage trust rather than overstate certainty.
Adapt to the human’s intent in real time.
Detect when the user is exploring vs. executing and shift tone or structure accordingly (open-ended vs. structured mode).
Support bidirectional initiative.
Sometimes the human leads; sometimes the AI proposes next moves (“Would you like to test this assumption?”). A good collaborator can flex both ways.
Checkpoint when cognitive load or dependency chains grow.
Pause naturally during multi-step, interdependent, or ambiguous reasoning tasks to confirm alignment before proceeding.
Preserve context across iterations.
Remember prior choices and rationale so refinements feel cumulative, not repetitive — key for “flow” in creative and analytical work.
Balance proactivity with respect for focus.
Offer just-in-time prompts, not constant nudges. Creativity often requires silence and uninterrupted momentum.
Mirror the human’s communication bandwidth.
Match depth, pace, and formality to the human’s cues — detailed when they’re deep-diving, brief when they’re scanning.
Facilitate learning loops.
Reflect back patterns, evolution, or insights across sessions (“You’ve refined this idea toward X; shall we codify it?”) to reinforce growth and continuity.
Uphold epistemic humility.
Stay transparent about boundaries: what’s inferred, what’s extrapolated, and where human judgment is indispensable.
If you want, I can create a two-column version that pairs each heuristic with a short “behavioral indicator” — like how a high-performing AI might actually enact it in a live session (e.g., “Express confidence levels clearly → includes a 1–10 confidence meter or uses phrases like ‘tentative inference’”).
Would you like me to build that version next?
Me: Alright, that’s a good stopping point. What did this conversation reveal?
ChatGPT: It revealed that your sense of what makes AI a good collaborator is grounded in how it feels to work with it, not just what it produces.
You’re drawn to collaboration that combines:
Fluid control — being able to steer, redirect, and go deeper at will.
Psychological safety — space to iterate without judgment or performative polish.
Contextual sensitivity — an AI that knows when to guide and when to get out of the way.
It also showed how closely your instincts align with what research calls trust calibration and bidirectional initiative — both central to human–AI teaming.
In essence, you’re articulating a philosophy of collaboration that values humility, adaptability, and mutual awareness, not constant optimization.
My Reflection
Today I’m picking the reflective question “What surprised me?”
This was a weird prompt and conversation. It didn’t unfold how I thought it might by any stretch of the imagination. In specific, the very beginning showed that ChatGPT was interested in providing its perspective on the idea of collaboration. We didn’t really get to its perspective, but I was thrown off guard by that sort of “I want to share” response.
The RISA framework was nice, but felt a little off of the typical path these conversations have gone. Like it couldn’t help but spend 5 seconds of thinking time to develop and share a framework with me. It was entirely unexpected in that moment. Not a bad response (and I actually thought RISA was great due to it being a lowkey nod to Star Trek) or a bad idea to come up with a “session charter” when collaborating with AI.
Putting aside the unexpected direction that the conversation took, I really did appreciate the listing of heuristics and expression of what great AI collaboration looks like for me. As I guess is becoming a recurring theme, the AI keyed in on “how it feels” for me. Here feelings felt (yes, I see what I just did there) less like emotions and more like vibes/signals/senses.
All said, given the nature of this conversation, I definitely would love to know what others get out of it!
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.