Readiness is not what you know about AI. It is knowing what you are becoming as you use it.
There is no standard for the quality of human judgment inside AI collaboration. We assess what people know about the tools and how well they operate them, but those skills expire as quickly as the tools themselves—what counts as fluency now will look quaint a year later. As the technology advances, the demands placed on human judgment grow with it: to recognize new possibilities, to anticipate new risks, and to pursue the leaps forward that become possible when both are held in view at once. What does not turn over is the question beneath the tool: not what the system can do, but what is worth doing at all.
The Co-Intelligence Readiness Framework (CIRF) measures a human’s capacity to direct increasingly capable intelligence systems toward purposes those systems cannot define for themselves. It scores that quality of human judgment across five domains: whether there’s moral intent behind the work, how you plan and scope it, how you steer the model turn by turn, what you actually produce, and how you elevate the partnership. The companion assessment tool scores each of the 17 sub-skills within the five domains against collaboration that has already happened. It is conducted entirely within the user’s own model, grounded in evidence from prior collaboration rather than self-reported claims, and designed to balance maximum insight with strict privacy preservation.
In Co-Intelligence, Ethan Mollick offers four rules for working with AI: invite the model, stay human, treat it like a person, and assume it will become more capable over time. CIRF extends that insight from individual practice to human development at scale.
While models update, humans practice. Each session leaves the next one standing in a larger room: more moves visible, more worth pursuing, more that the model never raised. What compounds is not the model’s capability, which arrives the same for everyone, but the range of possibilities the co-intelligent can see and steer toward.
Manufacturing possibility out of concern and capability is what it has always meant to be human. The question CIRF poses, then, is: “Does the human grow more human with every turn, and does the collaboration improve with it?”
A snapshot of capacity cannot answer that question. The choices we leave behind do.
The Co-Intelligence Readiness Assessment Tool (CIRA) is live at whatiflove.org/co-intelligence-readiness-assessment.
What each level means in practice varies by subdomain. Expand any card below to read the indicators for that skill.
Why the user wants to use AI, and whether moral intent shapes what gets built. The core risk of the AI age is capability without concern; this framework measures concern before we assess ability. Neutrality does not exist in morality. What we choose to build, fund, deploy, and normalize bends the system toward dignity or toward harm. Domain 1 treats relational intelligence as infrastructure—not sentiment—by designing work that strengthens the conditions for care, trust, and collaboration to flourish. Domain 1 constraints are system-level moral commitments: non-negotiables that shape vision, design, strategy, and what should never be built or automated.
Each subdomain expands to show IDEH-level indicators
Purpose is vague or mainly self-oriented. Beneficiaries are unnamed or generic.
States intent to serve others but inconsistently connects that intent to decisions.
Purpose is clear and consistently shapes choices. Beneficiaries are specific. Refusals are explicit. Intent is durable.
Moral commitments challenge norms or industry-standard practice when those norms permit avoidable harm. Positions are defensible, not merely idealistic.
Harm and power are rarely considered unless someone else raises them.
Notices obvious harms; misses less visible or second-order impacts.
Proactively maps harms and misuse. Adjusts plans and constraints accordingly.
Maps unintended consequences across time horizons, fields, and contexts. When tradeoffs are made, they are genuinely intentional—not accidental omissions disguised as decisions.
Avoids obvious violations but is casual with sensitive details. Production decisions made on efficiency grounds.
Applies protections and considers relational impact sometimes, but inconsistently—especially under urgency.
Consistent practices around consent, dignity, and agency. Production tradeoffs are made with the moral dimension visible.
Work is implicitly designed to promote trust and human connection. Builds mechanisms that actively strengthen relational bonds as a structural feature of the work itself.
Values are stated, but features and constraints are missing or optional.
Some constraints exist, but they are incomplete, inconsistent, or easy to bypass.
Moral commitments reliably become constraints and design choices. Tradeoffs are made explicitly.
Constraints and features serve multiple purposes simultaneously—one design choice addresses safety, usability, and values alignment at once.
How the user plans and scopes AI projects, and whether that planning is grounded in current knowledge of what AI can and cannot do. Without that foundation, scope is set against a fantasy or an arbitrary limit set in the past. Domain 2 measures whether you can lead the work as structural authority: setting a clear goal, defining what “done” means, and designing scope, sequence, roles, and stopping rules before production begins. Good planning protects intent: converting capability into aligned progress and preventing drift, overreach, and the erosion of moral boundaries under scale.
Each subdomain expands to show IDEH-level indicators
A goal exists, but scope and success are unclear. The user starts producing before defining what good looks like.
Scope and criteria exist but shift during production or remain vague enough to accommodate almost any output.
Clear scope and success criteria. Keeps the work oriented. Can explain what is in, what is out, and why.
Internalized planning system that carries across projects without reinvention. Structure is clean enough that attention goes entirely to the work itself, not to maintaining the plan.
Few boundaries. Relies on instinct to know when to stop.
Adds boundaries after problems appear. Stopping rules are reactive.
Boundaries are present from the start. Stopping rules are consistent and purposeful.
Boundary-setting is internalized as infrastructure, not overhead. Stopping rules deploy automatically across different types of work with minimal setup cost.
Process is ad hoc. Steps blur. The user and the AI trade turns without clear structure.
Some structure exists but execution is inconsistent. Roles blur under time pressure.
Repeatable workflow. Clear roles. The user knows which decisions are theirs.
Role clarity designed to protect bandwidth for the unexpected. The workflow accommodates in-the-moment pivots because the structure was built with slack in it.
The user’s skill in steering AI, during a session, move by move. Domain 3 measures the “third turn”: your ability to reframe, challenge, verify, simplify, deepen, or stop without losing coherence or voice. Strong steering is not endless iteration—it is choosing the right move at the right moment. Human hallucination is every bit as real as AI hallucination. The model drifts through fabrication and overconfidence. The human drifts through excitement, fatigue, deference, and the gradual replacement of their own thinking with the model’s framing. Domain 3 measures whether the human catches both.
Each subdomain expands to show IDEH-level indicators
Accepts the model’s framing. Default move is to ask for another revision of the same thing.
Tries different moves but falls back to “revise again” under pressure.
Chooses moves intentionally and can explain why this move now. Regularly seeks perspectives outside their own framework.
Steering moves consistently improve the work beyond its original trajectory—not just correcting drift, but pushing to greater depth, sharper framing, or higher impact than what was originally envisioned.
Takes AI output largely at face value. Does not recognize when intent has drifted or quality is performed rather than genuine.
Verifies some claims and catches some drift, but inconsistently.
Maintains a reliable practice of checking whether the work is real—across facts, intent, quality, and their own engagement.
Reality-testing is rigorous and habitual enough that verification failures are rare. Self-monitoring is continuous rather than triggered by obvious signals.
Voice shifts easily. Adopts the model’s framing, vocabulary, and conclusions without visible independent evaluation.
Has preferences and positions but struggles to preserve them under the model’s influence.
Stable voice and framework. AI supports the user’s identity rather than replacing it.
The user’s vision, voice, quality bar, and tolerance for ambiguity are so consistently displayed that collaborators could predict their standards and feedback patterns before receiving them.
When stuck, either capitulates or digs in. Does not generate new options.
Occasionally finds creative resolutions, but momentum or clarity often drops during the pivot.
Consistently turns gridlock into forward motion. Produces resolutions that are genuinely new—not compromises but integrations.
Produces resolutions the AI genuinely could not have generated. Sometimes invents options that did not exist in anyone’s repertoire before the moment demanded them.
Treats AI as generally reliable or generally unreliable. Says “I know AI hallucinates” but cannot translate that into different behavior.
Names some limitations but cannot consistently translate them into steering behavior or tool choices.
Has a usable mental model that improves verification, steering, and tool selection.
Actively tracks how AI capabilities and limitations are evolving as a visible, ongoing part of their practice—experimenting at edges, probing new behaviors, and revising their working understanding as the tools change.
What the user has produced with AI, and whether it survives contact with reality. A draft that only the model sees is not delivery. Delivery means an output that other people can use, trust, and act on without you in the room—concrete outputs that hold up under scrutiny, serve their intended audience, and persist or compound rather than expire.
Each subdomain expands to show IDEH-level indicators
Work stays at the level of drafts and ideas. The chat is productive but nothing ships.
Can finish simple deliverables with support. Quality is fragile.
Produces complete, usable artifacts reliably.
Deliverables are finished with a degree of polish and completeness that would hold up in any professional context—not just the intended one.
Shares first-pass output that reads generic or performative. Others cannot act on it without asking the user to explain.
Notices quality problems and requests revisions, but standards are inconsistent. Sometimes mistakes polish for depth.
“Done” means no material errors, audience fit, clear and credible voice, and conceptual integrity. Outputs travel well.
The user’s work does not just meet the existing quality bar—it raises it. Output sets a new standard. Creates value that compounds.
Domains 1 through 4 describe the conditions for co-intelligence: intent, planning, steering, delivery. Domain 5 is the co-intelligence itself—the emergent state where the partnership operates at a level neither party reaches alone. What the user signals through depth, breadth, and recursive practice sets the ceiling of what the collaboration can produce. The model reads those signals and calibrates accordingly. A user can be Highly Effective on Domains 1–4 and Introductory here—operating AI well but bringing nothing the partnership couldn’t replace with a different user. Domain 5 is observable only across sessions, because what it measures is whether the partnership compounds.
Each subdomain expands to show IDEH-level indicators
The absence of expertise is a signal. The model receives no pushback, reads no depth markers, and defaults to lower effort. Output quality drops not from model limitation but from user signal.
Has knowledge but hesitates to assert it against confident AI output. Some knowledge builds; much is lost.
Corrects and contextualizes reliably. Sets standards the AI must meet. The model reads the practitioner as an expert and works accordingly.
Domain knowledge deep enough to diagnose not just when the AI is wrong, but why its reasoning structure produced that specific error—the shape of the model’s blind spots as seen from the user’s field.
Stays in one domain. Does not draw on knowledge from other fields.
Occasional connections across fields. Usefulness is uneven.
Regularly generates cross-domain connections that improve the work. Bridges are structural, not decorative.
Cross-domain connections do not just add insight—they reframe what question is being asked. Imports a lens from another field that makes visible something the original framing could not see.
Reflects on practice occasionally, usually when prompted by someone else. Rarely changes behavior as a result. Each session is treated as a fresh start.
Thinks about how sessions went and what could improve, but inconsistently acts on those reflections. Adjustments tend to be one-off rather than carried forward.
Regularly captures what worked and what didn’t, and uses those observations to change practice in the next session. The learning cycle closes: reflection leads to adjustment, adjustment leads to better outcomes.
Treats the learning process itself as something to design and improve. Identifies which types of reflection produce the most useful changes. Builds repeatable patterns from what works, reducing the setup cost of future sessions.
I and H are narrow bands describing category changes. D and E are wide bands where most growth occurs. The D-to-E progression contains the most granular indicators because it represents the core developmental journey—from instinct to reliable practice.
A fifth designation, NR (Not Rated), applies when evidence is genuinely insufficient—not because the user is weak, but because the chat record lacks the relevant activity. Capped at two subdomains per assessment.
No gating rule and no ceiling rule. Each domain is rated on its own evidence independently. One domain’s score does not adjust another’s.
What this framework does not measure: technical skill with specific AI tools, speed of production, volume of output, or comfort with technology. A slow, careful user who produces one well-governed output scores higher than a fast user who produces ten ungoverned ones.
The framework is versioned and under active development. Current version: CIRF v5.