Rethinking how we work with AI: frame, explore, refine, commit.
FERC is not a prompt template. It is a governed interaction cycle for multi-round human–AI collaboration: The human frames the problem, the model generates alternatives, the human evaluates comparatively and reshapes the next round, and the human commits and owns the consequence. Developed at the Center for Hybrid Intelligence, applied in more than 20 organizations with over 2,500 participants, and now measurable through the FERC-Bot.
2,500+
individuals trained in FERC
20+
organizations, public and private
4 phases
observable, and therefore scoreable
20 authors
across 15 research institutions
01 · Authorship
Human intent, kept explicit
Authorship drift is gradual: Goals, criteria and finally responsibility move to the system without anyone deciding that they should. FERC makes each of those moves a step someone has to take.
02 · Process
Governance, not better prompts
Prompt frameworks structure the first message. FERC governs what happens across rounds: How alternatives are obtained, how judgment rebuilds the next question, and when the work stops.
03 · Measurement
Collaboration you can score
The guiding question is not “was the answer good?” but “was the collaboration well structured and authorship-preserving?” That question has an answer you can record.
Ensuring Human Authorship in the Age of Generative AI
Suggested citation: Sherson, J., Rafner, J., Reiter-Palmon, R., Lebuda, I., Söllner, M., Kenett, Y. N., Emruli, B., Rindle, M., Weiss, S., Goecke, B., Vinchon, F., Mao, Y., Dellermann, D., Rahimi, S., Marler, J. H., Nguyen, A., Bjerring, J. C., & DiPaola, S. (in press, 2026). Ensuring human authorship in the age of generative AI: The FERC framework for governed human–AI collaboration. Accepted for NordiCHI 2026 conference proceedings https://www.researchgate.net/publication/403595998_Ensuring_Human_Authorship_in_the_Age_of_Generative_AI_The_FERC_Framework_for_Governed_Human-AI_Collaboration
Where it has been used
2,500+ people have already been trained in FERC.
Across corporate, public-sector and educational settings, from trade unions and supermarket chains to universities, libraries and the European Commission. Including organizations like these.
Why we built it
One-shot prompting quietly takes four things off your desk.
The framing
Ask for a finished product and the model supplies the goal, the audience and the success criteria on your behalf. You will not notice, because the output looks like an answer to your question.
The alternatives
One fluent response arrives and the search space closes before it opened. Fluency encourages premature convergence, and there is nothing left to compare the answer against.
The evaluation
Asking for a better version is polishing, not judgment. Real evaluation is comparative and it changes the criteria, so the next round searches somewhere else.
The responsibility
Without an explicit stop, work ends when the tokens, the confidence or the calendar run out. A decision nobody made is a decision nobody owns.
Left unmanaged, this resolves into deskilling: Generation substitutes for learning, and the economic value of human expertise falls. Generative AI tends to benefit less-experienced users most, which gives organizations a rational incentive to underinvest in developing expertise at all. FERC is the behavioral answer to that incentive, the layer where an individual professional keeps evaluative authority across rounds and closes with accountable commitment.
From product focus (“the perfect prompt”) to process (“the perfect collaboration”).
What the human is for
Prediction–judgment: A durable definition of complementarity.
Asking which tasks AI will take is the wrong question, because the answer changes with every model release. The durable question is what kind of work each side is structurally suited to. Machines do prediction: They are strong wherever large amounts of structured, labelled data exist and human computation is infeasible. Humans do judgment: Evaluation and decision-making that rest on tacit and contextual knowledge which is infeasible to write down in the first place.
Between the two poles sits a scale, not a switch: Human-dominated work with machine suggestions, genuinely shared tasks, machine-dominated work with human verification. The frontier between prediction and judgment moves as more contextual knowledge is made available to machines, which is exactly why the position has to be chosen deliberately for each task rather than inherited from whatever the tool does by default. FERC is how that choice gets made in practice, one interaction at a time.
The cycle
Real work is never one-shot. Why should prompting be?
Four phases, each a governance function that cannot be merged with its neighbour without recreating a familiar failure: Generation without criteria, refinement without divergence, iteration without an ending.
Application principles
Six principles for FERC
Ensure human authorship by becoming the playing coach of your AI.
1
Use FERC for text products such as ideas and drafts. Information search works differently.
2
FERC is a standard procedure that raises the floor, not a religion to be followed slavishly.
3
Never ask AI for the first draft, always give it context and structure.
4
Never use AI as an expert beyond your own, or your organization’s, expertise.
5
Split the collaboration into steps small enough that you can ask for options and evaluate them.
6
Small steps mean lower energy use and cost, and less dependence on Silicon Valley models.
The E–R loop is the engine
Explore and Refine are meant to be run more than once. Each Refine is not a request for a nicer output but a rewrite of the brief: What the comparison taught you becomes the constraint for the next set of options. That is the co-evolution of problem and solution that creativity research describes, made explicit as an interaction pattern.
Grounding
Problem construction
Frame
Divergent thinking
Explore
Iterative elaboration
Refine
Convergent evaluation
Commit
Co-evolution of problem & solution
Frame / Refine cycles
Principle 2
A standard procedure that raises the floor.
Run the full FERC cycle when …
•
you are the author of a text product, an assignment, feedback, a course
•
the task is new or open-ended
•
the quality has to be defensible to others
You can skip it when …
•
you are replying to a short email
•
the flow is linear and familiar
•
you are looking for facts, ask for a source instead
Once you know FERC, you also know when you don’t need it.
The trade-off
The FERC sweet spot
Two extremes bound the space. At one end, minimal-effort answer seeking: You ask, the model delivers, and framing, evaluation and decision criteria are outsourced by default. At the other, intensive Socratic tutoring, which scaffolds deep reasoning but demands sustained effort and does not scale in an organization. FERC is the crossing point: Enough human effort to keep authorship, little enough to be repeatable on a Tuesday afternoon.
A very small example
What’s for dinner?
The cheapest way to learn the cycle is on something with no stakes at all. Watch where the judgment sits.
Nothing here requires an AI, and that is the point. The structure is identical whether the decision is tonight’s meal or a market-entry strategy, only the cost of getting it wrong changes.
1 · Frame
The frame comes before the ideas
Frame, collected by the AI
Two adults
About 25 minutes, simple
Spinach, eggs and tomatoes
No meat, nothing spicy
The AI asks one question at a time. “Does that fit?” “Yes, that fits.”
2 · Explore
Five options inside the frame
1 · Pasta
20 min
2 · Omelette
15 min
3 · Chickpea salad
10–15 min
4 · Eggs in tomato pan
25 min
5 · Crispy tortillas
20 min
You pick the two that look most promising.
3 · Refine
Weigh up, then explore again
+ Pro: Feels like a proper dinner
− Con: Uses almost all my time
What makes one more appealing?
“Something warm, with five minutes to spare.”
Your answer shapes a second Explore round.
4 · Commit
You choose and own it
Tomato pan with spinach, eggs and rye croutons
✓ Fits the frame
✓ About 20 minutes
✓ No shopping needed
The AI never decides. You do.
Problem decomposition
Solve the problem before you solve the solution.
Most FERC work needs one cycle. When a task is genuinely ill-defined, the cycle compounds: You run FERC once on the problem and once on the solution. This is the Double Diamond read as an interaction pattern rather than a workshop poster, Discover and Define are a FERC cycle in the problem space, Develop and Deliver are a FERC cycle in the solution space.
CYCLE 1
The problem space
Frame the situation, not the deliverable. Explore alternative definitions of what the problem even is. Refine by comparing those definitions against what you actually know. Commit to one framing, in writing.
Leaders converging on which AI use case is worth doing at all.
CYCLE 2
The solution space
The committed framing becomes the frame of the second cycle. Explore implementation strategies, refine them against cost, risk and capability, commit to one and own it.
The same leaders deciding how that use case gets deployed.
THE TEST
Decompose until evaluation is realistic
If a task feels too large to review several alternatives for, one output is also too large to review responsibly. Cut the work into steps small enough that you can genuinely compare options at each one.
Smaller steps also need smaller models: Less token cost, lower footprint.
Chaining is one composition among several. Cycles can also nest inside a single phase, repeat, or interleave as the task demands. What matters is that every span of AI output is short enough for a domain expert to check against their own knowledge, which is why granular interaction removes most of the need for model transparency in the first place.
Principle 5 · Human Review Horizon (HRH)
You can only evaluate what you can understand.
Ten thousand lines of code back from an agent, or a thirty-page transformation plan, cannot be refined, only accepted. So before framing anything, the task has to be cut to a size your expertise and attention can carry.
Your review horizon is the largest amount you can meaningfully evaluate in one FERC cycle.
Within the horizon
Deep understanding
Qualified judgment
Effective refinement
At the edge
Harder to understand
More spot-checking
Higher risk
Beyond the horizon
Superficial reading
Automation bias
Authorship drifts
FERC rule: does the expected output exceed your review horizon? → Break the task down first.
In the example: a whole essay in one go sits beyond the horizon. Four smaller levels each sit within it.
Example: Text generation
The task: write a high-quality essay on a given topic.
1
State the core message
What is the most important thing the reader should understand or believe?
2
Build the narrative arc
What is the journey from introduction to a strong conclusion?
3
Build the outline
Which sections are needed to deliver each part of the story?
4
Write the text
How do I express each part clearly and in my own voice?
Each level can go through the FERC loop before you move on to the next.
The method: Problem decomposition
Break the challenge into a hierarchy of units that each fit inside the horizon, run the full cycle on each, then recombine. Decompose the process, not just the product, the units that work match how the work is actually done, which is why the person with the domain expertise has to do the cutting. Commit then has two outcomes: Finish the unit, or decide it is still too large and re-enter Frame one level down.
FERC-SD applies the same logic to software development: Outcome, architecture, components, integration, validation and deployment each get their own cycle. Humans author, AI assists, and nothing reaches production before a human both understands and accepts it.
The failure mode · FAKE FERC
It looks like FERC, but it isn’t.
Running the four phases is not the same as exercising them. The experience of control is not evidence of control. Once FERC becomes the expected way to work, the cheapest response is to perform it: Paste the task in, take what comes back, and describe the result afterwards in the vocabulary of framing and refinement.
What makes this worse than skipping the cycle outright is that it produces a record. Nothing in the output looks wrong, it is fluent, structured, plausible, and the missing professional grounding is invisible to everyone downstream, including the person who signed it. Errors and misunderstandings are carried forward silently, and the process log says judgment was exercised when it was not. At organizational scale this is what the HI Manifesto calls HI-washing, and it is the specific thing process-level scoring exists to catch.
Skipping FERC deliberately is a legitimate choice: Low-risk, procedural work does not need four phases, and saying out loud that you are trading depth for speed keeps the delegation honest. Faking it is the one version with no defence.
Fake FERC
Frame
without professional context
→
Click send
paste the task in
→
Accept
take the first output
→
Done
no professional assessment
No Explore–Refine: You never choose between options
FERC
Frame
your knowledge and your criteria
→
Explore
ask for several options
⇄
Refine
what do you like, and why?
→
Commit
you own the result
The consequence
The result looks professional but lacks professional grounding, and the errors are carried forward invisibly.
The instrument
FERC-Bot scores the collaboration, not the answer.
Conceptual understanding rarely changes everyday habits. FERC-Bot embeds the four phases in a live chat environment and returns a per-phase score with a short explanation: Did you define goals and criteria, did you request genuine alternatives, did you evaluate comparatively in a way that reshaped the next round, did you decide deliberately. Substantive output quality remains the domain expert’s responsibility.
The FERC-Bot interface: A standard chat environment augmented with a live diagnostic panel that tracks collaboration quality across the four phases.
Scores
Overall authorship score
Illustrative walkthrough. Scoring metrics and their validation are ongoing work.
Workshop tool
A generator that gives every team its own FERC exercise.
The hardest part of a FERC workshop is finding a case that participants recognize and can finish in fifteen minutes. This tool builds one on the spot. It takes a company name, proposes five exercises drawn from publicly known information about the organization, and then facilitates the chosen one step by step, one phase at a time, one question at a time, no internal documents or data required.
FERC Workshop Studio: An organization and a strategic focus in, a set of practice challenges out.
How a session runs
01
Name the company
The only input. The tool works from publicly known information about the industry, products and typical areas of work, and stays general rather than inventing internal facts.
02
Five candidate exercises
Each a concrete decision with several plausible answers, understandable without extra material, and finishable in ten to fifteen minutes. Participants pick one.
03
Frame, facilitated
Three short questions, one at a time: Who it is for, what matters most, which constraints apply. Then a summary the participant has to confirm.
04
Explore and refine
Five clearly different, unranked ideas. The participant picks two, names an advantage and a disadvantage for each, and can run another round before closing.
05
Commit
One final choice, confirmed, tied back to the frame, and a first small step to test it. A complete cycle, in a quarter of an hour.
The facilitator prompt, verbatim
Organizational level
A maturity model that measures authorship, not AI usage.
Five levels, refined across more than ten executive training sessions. Across a business college faculty, a group of marketing executives and 47 leaders from a Nordic supermarket chain, self-assessed maturity clustered between Level 1 and Level 2: Multi-round experimentation is common, structured evaluative reframing and explicit commitment are not.
Where three groups placed themselves
Before the levels are described, here is how they were used in practice: each participant rated their own way of working with AI against the five levels. Select a level to see what share of each group placed itself there. Most sit at Level 1 or 2.
No group reported Level 4.
Progression is not only a training problem. Experts share the tacit knowledge that Frame depends on when they can see that their expertise is being developed rather than extracted.
The research behind it
Grounded in creativity and metacognition research.
Creative work is not one-shot generation. Problem construction research shows that creators repeatedly reframe the problem itself, and process models describe an iterative interplay in which evaluation reshapes the next search rather than terminating it. FERC operationalizes that metacognitive control for human–AI collaboration, and distinguishes authorship, accountable agency in the process, from ownership, which organizations may assign contractually.
From framework to interface: FERC-Bot as a metacognitive training system for governed human-AI collaboration
Suggested citation: Rindle, M., Kjeldsen, F. B., Berg, M., Rafner, J., Kenett, Y. N., Goecke, B., Weiss, S., & Sherson, J. (2026). From framework to interface: FERC-Bot as a metacognitive training system for governed human-AI collaboration [in press]. 5th International Conference on Human and Artificial Rationalities, Paris, France. https://www.researchgate.net/publication/415081302
Operationalizing Human Authorship in AI Systems: A Design Science Study of the FERC Framework
Suggested citation: Sherson, J., Rindle, M., Kjeldsen, F. B., Berg, M., Kenett, Y. N., & Rafner, J. (2026). Operationalizing human authorship in AI systems: A design science study of the FERC framework. Proceedings of the 5th International Conference on Hybrid Human-Artificial Intelligence (Frontiers in Artificial Intelligence and Applications, Vol. 423, pp. 541-543). IOS Press. https://doi.org/10.3233/FAIA260554
Context · HI Manifesto series
The Hybrid Intelligence Manifesto
The multi-level socio-technical framework FERC belongs to. FERC carries the interaction-level share; dedicated manifestos handle the interface, market and educational levels.
Read the manifesto →
Where the scoring leads
Two pathways: Process assessment in education, auditable human contribution at work
Because the four phases are observable, they can be recorded. In education that restores an assessable object after output-based assessment stopped working. In professional work it supplies a documentation layer for human contribution, the micro-level analogue of firm-level carbon accounting.
HI Educational Manifesto →
Academic contributors
Twenty researchers, fifteen institutions.
Jacob F. Sherson, Janet Rafner, Roni Reiter-Palmon, Izabela Lebuda, Matthias Söllner, Yoed Nissan Kenett, Blerim Emruli, Manuel Rindle, Selina Weiss, Benjamin Goecke, Florent Vinchon, Yaoli Mao, Dominik Dellermann, Seyedahmad Rahimi, Janet H. Marler, Andy Nguyen, Jens Christian Bjerring, Steve DiPaola, Frederik Brosbøl Kjeldsen, Mille Clausen Berg.
Get involved
Three ways to work with FERC.
Run it with your organization
FERC training for leaders and teams: Executive sessions, hands-on workshops and the maturity assessment. The workshop generator builds the exercises from your own line of work rather than a generic case.
Use it in teaching
We bring FERC into teaching together with educational institutions, so students learn to frame, compare and commit rather than accept the first output. Coming soon: A FERC module in the AI driver’s licence, developed in collaboration with Center for IT in Development.
Research with us
Test FERC with us in a student setting. Running the cycle and FERC-Bot in real courses gives us data on how collaboration quality develops, and helps validate the scoring metrics behind it.