AI Was Supposed to Broaden Our Thinking. What If It Does the Exact Opposite?
Can AI improve an individual idea while simultaneously narrowing the solution space? Anti-Fixation Lab explores whether controlled context isolation can help reduce fixation in human-AI collaboration.
Anti-Fixation Lab: an experimental architecture for human-AI collaboration based on memory, controlled context isolation, and deliberate “forgetting”
Generative artificial intelligence is an excellent partner for conceptual work. In a matter of seconds, it can propose new directions, develop an idea, identify its weaknesses, create alternatives, and respond to the next human comment.
This creates a highly efficient loop:
human → AI → human → AI → human → AI
Each successive response can be better than the previous one.
But this is exactly where the problem appears.
A better and better answer does not necessarily mean broader and broader thinking.
We can spend an hour having an extremely productive conversation with AI while the entire process keeps exploring only one fragment of the space of possible solutions in ever greater depth.
Worse still, both the human and the AI may feel that the process is highly creative.
At Ferway, we therefore began working on an experimental architecture whose goal is not to force AI to generate even more ideas.
We want to test something else:
Can human-AI collaboration be designed so that the process itself can detect when it is starting to think too narrowly—and deliberately leave that path?
We have provisionally named this architecture Anti-Fixation Lab.
The paradox of creativity with AI
The most natural way of working with a language model looks roughly like this:
the human proposes an idea → AI develops it → the human reacts → AI develops the reaction → the human notices another direction → AI continues.
This is no longer simply using a tool.
A co-creation loop emerges.
HUMAN
↓
idea A
↓
AI
↓
development of A
↓
HUMAN
↓
new association
↓
AI
↓
an even better A
Such a process can lead to a solution that neither the human nor the model would probably have created independently.
But exactly the same mechanism can work in the opposite direction.
The first proposal creates a frame. The next response develops it. That development affects the way the human starts looking at the problem. The next question is already formed inside that frame.
After several iterations, we do not get:
A → B → C → D
but rather:
A → A+ → A++ → A+++
The solution becomes increasingly refined.
Not necessarily increasingly original.
Research shows that the problem is real
In a study presented at CHI 2024, Samangi Wadinambiarachchi and co-authors examined the impact of generative AI on design fixation and divergent thinking in an ideation task. The experiment involved 60 participants.
People using an AI image generator showed greater fixation on the initial example. Compared with the baseline group, they also generated fewer ideas, with less diversity and lower originality.[1] DOI
This does not mean that AI always reduces creativity. The study concerned a specific type of task and a specific form of AI support.
It does, however, reveal an important mechanism:
a tool designed to support ideation can simultaneously become a source of fixation.
Another experiment, published in 2024 in Science Advances by Anil Doshi and Oliver Hauser, revealed an even more interesting paradox.
Access to AI-generated ideas caused short stories to be rated, on average, as more creative, better written, and more enjoyable. At the same time, stories created with AI support were more similar to one another than stories created solely by humans.[2] DOI
We can therefore get both at the same time:
a better individual result + less diversity across the overall space of results.
This distinction is crucial.
The problem may be the human-AI loop itself
The simplest answer is: we need better prompting.
We can tell the model:
Do not automatically agree with me.
Or:
Suggest something completely different.
Or:
Generate five alternative concepts.
This may help.
But it does not solve the whole problem.
The model still knows the earlier conversation. It knows which idea came first. It knows what interested the user. It knows which arguments were considered important. It even knows what it is supposed to be “completely different” from.
So it is still moving within the history of the same conversation.
The problem may not stem solely from human behavior or solely from AI behavior.
It may be a property of the entire coupled system:
human → AI → human → AI
The human proposes A.
AI develops A.
The development of A affects how the human understands the problem.
The human asks another question that already follows from A.
AI receives even more context related to A.
And the whole system begins to converge.
Neither participant has to make an obvious mistake.
And yet the search space narrows.
What if part of the system deliberately knew nothing?
Instead of trying to force one model to keep “thinking differently,” we can approach the problem at the architectural level.
Not every component of the system has to have access to the same history.
Some can receive a fuller context.
Others only part of it.
Still others can work exclusively from the original formulation of the problem—without knowing what solutions were generated earlier.
The system then gains two seemingly contradictory properties:
continuity and controlled amnesia.
The latter term is a metaphor. Technically, this is about controlled access to context, not memory in the biological sense.
Fuller context makes it possible to develop earlier discoveries.
Limited context makes it possible to look at the problem again without carrying the full burden of the trajectory the system has already created for itself.
This is the central hypothesis of Anti-Fixation Lab.
What are we actually proposing?
Anti-Fixation Lab does not assume that we need a “more creative” AI model.
It assumes that diversity can be increased by changing the architecture of information flow.
The system deliberately maintains different knowledge states in parallel: fuller context, limited context, and clean context. It brings them together only after independent exploration, introduces the human before the selection stage, checks for symptoms of convergence, and—if necessary—performs a controlled reset.
The experimental variable is therefore not only the prompt or the agent persona. It is access to information over time.
Anti-Fixation Lab architecture
The simplified pipeline looks like this:
At first glance, this resembles a multi-agent system.
The most important difference, however, is not the number of agents.
What we primarily differentiate is the information each component of the system is allowed to access.
1. Synthesizer: first a map of the problem, not a solution
The first component is the Synthesizer.
Its task is not to propose a solution.
It is supposed to create a neutral map of the problem:
- goals,
- hard and soft constraints,
- unknowns,
- tensions,
- contradictions,
- assumptions,
- relevant decision axes.
This distinction matters.
If the first agent immediately presents three concrete solutions, everything that happens later may become anchored to them.
The Synthesizer should therefore primarily answer the question:
What is the problem made of?
rather than:
How should it be solved?
It should also preserve contradictions instead of resolving them too quickly.
If there is tension between two goals in the problem, there may be no single “obvious” way to resolve that tension.
Different branches may resolve it in completely different ways.
But even a neutral problem map can become a source of anchoring.
That is why the Clean Room does not receive it at all.
2. Three branches, three different knowledge states
Once the map has been created, the actual exploration begins.
We do not ask a single model to generate ten ideas.
We create several branches operating under different informational conditions.
Expander A — fuller context
The first branch knows the problem and the map created by the Synthesizer.
In later iterations, it may also know the relevant process history, earlier decisions, and permitted human feedback.
Its advantage is continuity.
It can use earlier discoveries and develop them further.
At the same time, precisely for that reason, it is the branch most exposed to reinforcing the existing trajectory.
Expander B — limited context
The second branch knows the original problem and the Synthesizer’s map.
It does not, however, know Expander A’s response.
It should also not receive information about which proposals the user liked.
This is an important distinction.
If we tell the agent:
A proposed X. You come up with something completely different,
then X still becomes the reference point.
The agent may produce anti-X rather than genuinely exploring the problem independently.
Expander B is therefore not supposed to:
complement A, criticize A, or be more original than A.
It is simply not supposed to know what A did.
Clean Room — raw problem
The third branch receives only the original problem definition.
It does not see:
- the Synthesizer’s map,
- Expander A’s response,
- Expander B’s response,
- the human’s reactions,
- previous iterations,
- the current “favorite.”
Its task is not to be more creative.
Its task is to operate in a different informational state.
This is a fundamental difference.
It is not about different experts. It is about different histories
A typical multi-agent system may look like this:
Agent 1 — creative
Agent 2 — analytical
Agent 3 — critical
Such an arrangement can be useful.
But all three agents may know exactly the same history and remain anchored in the same assumptions.
Anti-Fixation Lab tries to differentiate something else:
the state of knowledge at the moment a response is generated.
One branch knows more.
Another knows less.
The third does not know the trajectory the system has already developed.
Only later do we allow those trajectories to meet.
Current experimental environment
The first implementation of the architecture is currently being tested in ChatGPT Work using OpenAI’s GPT-6 Astra model.
OpenAI has presented GPT-6 Astra as a model intended, among other things, for complex professional work, web browsing, computer use, programming, and other demanding tasks.[5] OpenAI
This does not mean, however, that Anti-Fixation Lab is an architecture designed specifically for Astra.
Quite the opposite.
Ultimately, the architecture should be model-agnostic.
What interests us is the process: information flow, branch isolation, points of human intervention, convergence detection, and controlled resets.
In the future, one important test will be to see whether similar effects also occur with other models.
This also means that the model is one of the experimental variables that should be documented.
Implementation status: October 2026.
3. Human Gate H1: the human before selection
After the first phase of exploration is complete, the system stops.
It does not launch the Evaluator yet.
It does not choose a winner.
It does not assign scores.
It does not even ask:
Which option do you like best?
Instead, the human is given three questions:
What surprised you?
What is missing here?
What new idea, association, or question occurred to you while reading?
This is intentional.
We want to capture the moment when AI responses become a stimulus for human thinking, before the human is forced to choose.
In a conventional process, after generating several proposals, we quickly move to the question:
which one is best?
That moves us from exploration to selection.
H1 deliberately tries to delay that selection.
The human response should be preserved as literally as possible.
If the Orchestrator first summarizes it, polishes the wording, or interprets it, it may unintentionally change its meaning.
4. Anti-Fixation Controller: the agent that is not supposed to invent
Only after the human has reacted is the Anti-Fixation Controller launched.
This is an unusual role.
The Controller is not supposed to find a solution.
Nor is it supposed to identify the best idea.
It is supposed to look at the process from a meta-level.
Among other things, it checks:
Solution-family diversity
Do the individual branches actually represent different families of solutions?
Assumption diversity
Are they not based on the same hidden assumptions?
Trajectory diversity
Do they reach the goal through meaningfully different paths?
Structural diversity
Are the mechanisms different, or only the surface features?
Human novelty
After seeing the results, did the human contribute something that was not there before?
Premature convergence
Has the system begun optimizing a particular direction before alternatives were explored sufficiently?
There is, however, one important caveat:
The Controller does not “know” that fixation has occurred.
It is also a model.
It can only look for symptoms of fixation according to defined criteria.
Its diagnosis is therefore also a hypothesis that can later be compared with the course and results of the experiment.
Different products can be the same idea
Imagine three seemingly different solutions:
an app, a service, and a digital product.
On the surface, they look different.
But all of them may rely on the same structure:
online customer acquisition → sale → standardized delivery → increased volume
In that case, we do not have three fundamentally different directions.
We have three manifestations of a similar mechanism.
That is why diversity should not be measured only by the number of responses.
What interests us is the structural distance between solutions.
And this leads to another important caveat.
The goal is also not to maximize diversity at any cost.
A system that produces increasingly absurd solutions may score very well on a simple “diversity” metric while being completely useless.
What interests us is:
relevant diversity within the boundaries of the problem, without sacrificing the quality of the final solution.
5. Controlled Reset: a controlled sideways jump
If the Controller concludes that exploration has become too narrow, we do not simply ask:
Generate five more ideas.
Because the likely result may be five new versions of the same family.
The system performs a Controlled Reset.
The reset should correspond to the specific fixation mechanism that has been diagnosed.
R1 — Blind Branch
We launch an additional branch without access to the solutions generated so far.
This is the simplest form of looking at the problem again.
R2 — Assumption Reversal
We identify an assumption shared by multiple directions.
And we ask:
What happens if this assumption is false, removed, or reversed?
Often, we are not anchored in the solution itself.
We are anchored in an assumption we have not even noticed.
R3 — Forbidden Winner
The dominant family of solutions is temporarily forbidden.
Not because it is bad.
Because we want to see what exists beyond it.
R4 — Distant Analogy
The system looks for problems with a similar structure in distant domains.
The point is not to copy surface features.
What interests us is mechanism transfer.
R5 — Human Reset
AI stops generating.
The human first proposes an independent direction, question, or association.
Only then does the system bring AI back in.
Sometimes the best anti-fixation mechanism may simply be to remove AI from the loop for a moment.
R6 — Problem Reframe
Instead of looking for another answer, the system reformulates the problem while preserving its hard constraints.
Sometimes a new solution does not require a new answer.
It requires a new question.
6. Anti-fixation cannot continue forever
It is easy to fall into the opposite of the original problem.
If fixation is bad, one may conclude that the system should keep generating more and more alternatives.
That does not make sense either.
At some point, additional branches begin to deliver mainly variations of what we already know.
That is why we need an exploration stopping condition.
The process can begin moving toward convergence if:
- there are several genuinely different families of solutions,
- the main shared assumptions have been surfaced,
- there is at least one direction clearly distant from the dominant cluster,
- there is no obvious large space that the system has failed to explore,
- further attempts begin producing mainly variants.
The goal is not to discover every possible answer.
The goal is sufficiently broad exploration before selection.
7. Only now does the Evaluator appear
This is one of the most important elements of the entire architecture.
Divergence and convergence are separated.
In the first phase, we ask:
What fundamentally different possibilities exist?
Only later:
Which of them make sense?
The Evaluator can analyze, among other things:
- feasibility,
- risk,
- costs,
- uncertainty,
- dependencies,
- scalability,
- required resources,
- barriers to entry,
- side effects.
It should also distinguish between:
FACT
ASSUMPTION
HYPOTHESIS
ESTIMATE
UNKNOWN requiring validation
It should not, however, hide all of that complexity behind a single score:
A — 8.7/10
B — 7.5/10
C — 6.9/10
Such numbers look objective, but in practice they require assigning weights to criteria.
And those weights are assumptions too.
The Evaluator should therefore primarily preserve trade-offs.
A may be better in one respect.
B in another.
C may be riskier but offer greater potential.
The decision still belongs to the human.
8. Minority Report: before we discard the strange idea
Just before selection, one more safeguard appears.
Minority Report.
The system identifies an idea, mechanism, or observation that may be about to be rejected not because it has been disproved, but because it:
- is unusual,
- is less developed,
- is harder to explain,
- does not fit the emerging consensus.
Minority Report does not say:
This is the best idea.
It says:
Before you choose, notice what you are about to abandon.
This is the final safeguard before convergence.
9. Human Gate H2: the human makes the decision
Only now do we reach actual selection.
The human receives:
- different directions,
- their trade-offs,
- relevant unknowns,
- the evaluation,
- the Minority Report.
They can:
choose, reject, combine, modify, or propose their own direction.
They may also conclude:
something is still missing.
And then the process returns to exploration.
This matters.
The human is not merely a judge choosing among answers generated by AI.
They are a full-fledged source of new directions.
10. Co-Performer: now we can develop the direction
After Human Gate H2, the nature of the work changes.
We no longer need the widest possible exploration.
We need to develop the selected direction.
This is where the Co-Performer appears.
Its task is to work with the human to turn the direction into a concrete outcome:
a strategy, project, plan, product, experiment, model, specification—depending on the nature of the problem.
Earlier, AI’s tendency to develop an already existing direction was one of the main risks.
Now it can become an advantage.
First, we protect ourselves against premature:
A → A+ → A++
But once we consciously choose A, we may want exactly:
A → A+ → A++ → A+++
The problem, then, is not convergence itself.
The problem is convergence happening too early.
11. Human Gate H3: the result may change the problem
After a concrete artifact has been created, the human evaluates the result again.
They may accept it.
They may revise it.
They may reject some assumptions.
They may also notice something that changes how the original problem is understood.
Then another iteration begins.
But it should not simply pass the entire history so far to every agent.
That would once again trap all branches in the same trajectory.
12. Iteration N+1: memory and controlled amnesia
In the next iteration, information asymmetry is preserved.
The History-aware branch knows the relevant results of previous iterations.
The Restricted branch receives the problem and a limited set of information.
The Clean Room once again receives only the source problem.
The system therefore has both:
memory and controlled ignorance.
And this leads to one of the most interesting ideas in the entire architecture:
In a creative process, forgetting can be a feature, not a bug.
We often associate the development of AI systems with more context.
More information.
Longer memory.
A larger context window.
Better access to the entire history.
But in creative exploration, maximum access to history may not always be optimal.
Sometimes it may be useful for part of the system not to know what the rest has already concluded.
13. The context firewall is more important than the persona
It is not enough to say:
Act as if you do not know the previous answers.
If earlier solutions are present in the model’s active context, we do not have a true Clean Room.
We have a model simulating ignorance of something it can actually see.
That is why the architecture requires something like a context firewall.
Each branch should receive only the information it is allowed to access.
If the technical environment cannot genuinely guarantee such isolation, the system should state this explicitly.
For example:
CLEAN_ROOM_ISOLATION_NOT_GUARANTEED
This is not a technical detail.
Without this information, we do not know whether differences between responses arise from different access to context or merely from the same model playing a different role.
14. The Orchestrator can ruin the experiment
Even if the branches are properly separated, there is still one more possible source of fixation.
The Orchestrator.
If the central system starts to:
- improve responses,
- merge them,
- summarize them intelligently,
- fill in missing elements,
- select what it considers most important,
it becomes another creative participant in the process.
Its role should therefore be as procedural as possible.
It may:
route information, launch roles, enforce context access, record results, and stop the system at Human Gates.
It should not:
add its own ideas or silently merge independent branches.
Especially before H1, the human should, where practical, see the original branch outputs rather than only a synthetic summary.
Otherwise, it is the Orchestrator that decides what becomes the stimulus for the human.
15. The genealogy of an idea may matter more than the idea itself
If we want to study genuine human-AI collaboration, it is worth tracking provenance—the origin of new ideas.
A new element may come from:
AI-A
the branch with fuller context,
AI-B
the isolated branch,
CLEAN
the Clean Room,
HUMAN
the human,
RESET
a controlled reset,
or be marked as:
HYBRID
meaning a result created from a traceable combination of human and AI contributions.
An example genealogy might look like this:
CLEAN C3
│
▼
HUMAN H1-2
│
├────────── AI-A A4
│ │
└───────┬──────┘
▼
HYBRID XIn such a case, it would be imprecise to say:
AI came up with X.
But it would be equally imprecise to say:
the human came up with X.
X emerged from a specific interaction trajectory.
And that trajectory may turn out to be the most interesting object of study.
16. What comes from research, and what is our hypothesis?
This distinction is necessary.
Anti-Fixation Lab, in the form described here, is not a scientifically validated method.
It is an experimental architecture being developed at Ferway.
Research does, however, provide elements that justify asking the question in the first place.
Wadinambiarachchi and co-authors showed that, in the task they studied, using generative AI during ideation was associated with greater fixation and poorer performance on several measures of divergent thinking.[1] DOI
Doshi and Hauser showed another paradox: AI can improve the judged creativity of individual work while simultaneously increasing similarity across works at the group level.[2] DOI
In 2025, Nicholas Davis, Jacob Sherson, and Janet Rafner proposed the Co-Creative Design Framework for Hybrid Intelligence, describing human-AI co-creation through three main dimensions: agency, interaction dynamics, and communication.[3] cc.acm.org
In a 2026 publication, Juan Carlos Quiñones-Gómez describes five situational roles for AI in the design co-creation process: Synthesizer, Expander, Reframer, Evaluator, and Co-Performer. The author bases the framework on a synthesis of 13 publications and a questionnaire conducted among 16 design professionals.[4] e-publishing.cern.ch
Anti-Fixation Lab uses some of this terminology and draws on the broader idea of human-AI co-creation and hybrid intelligence.
Our experimental proposal is the specific combination of these elements with information-flow control: asymmetric context, Clean Room, Human Gates, an Anti-Fixation Controller, controlled resets, Minority Report, and the simultaneous preservation of continuity and controlled isolation.
We do not yet know whether this combination will actually produce the expected effect.
That is precisely why we treat it as a hypothesis.
17. The cost of independence
An architecture of this kind is not free.
Compared with a conventional conversation with AI, it requires:
- more model calls,
- more complex orchestration,
- control over context flow,
- recording the outputs of individual branches,
- additional stages of human interaction,
- more time.
It may also generate artificial diversity.
The Clean Room may produce solutions that are different but worse.
It may also repeatedly rediscover the same things as the branch that has access to history.
The Controller may incorrectly diagnose convergence.
A reset may open a direction that turns out to be a dead end.
And the additional complexity may fail to provide enough benefit to justify its cost.
All of this is possible.
That is why Anti-Fixation Lab should not be treated as a solution that must, by definition, be better than conventional dialogue with AI.
That is exactly what needs to be tested.
18. How do we test whether it works?
It is not enough to run the system and conclude:
interesting ideas appeared.
Such a result would be difficult to distinguish from an ordinary observer effect.
We need a comparison.
The same suitably complex problem can be solved under several conditions:
A. Conventional linear human-AI collaboration
B. A typical multi-agent system
C. Anti-Fixation Lab with controlled context isolation
The conditions should be as comparable as possible in terms of problem, time, and resources.
And we should not compare only the final answers.
At minimum, four different effects interest us.
Structural diversity
Does system C discover more genuinely different families of solutions?
Not more names.
Not more products.
More different mechanisms.
Post-reset distance
Does the Clean Room or Controlled Reset actually move exploration beyond the previous cluster?
Or does it merely produce a cosmetic variation of the dominant direction?
Human novelty
After seeing the system’s results, does the human generate new ideas, questions, or associations that were not there before?
This may be especially important.
Perhaps the greatest value of an AI response is not the response itself.
It may be the thought that the response triggers in the human.
Final quality
Diversity alone is not enough.
A system that generates dozens of very distant but weak solutions will not be a better decision-making system.
What interests us is relevant diversity without a significant loss of final-result quality.
The Anti-Fixation Lab hypothesis
At this stage, it can be formulated as follows:
Controlled context isolation combined with the detection of convergence symptoms may increase the relevant diversity of explored solutions without significantly reducing the quality of the final solution compared with linear human-AI collaboration.
The word “may” is crucial here.
It may turn out that the Clean Room does not add meaningful value.
It may increase diversity at the expense of quality.
It may work only for certain classes of problems.
It may also turn out that the biggest difference appears not in the ideas generated by AI, but in the ideas that later arise in the human.
Each of these outcomes would be informative.
Perhaps we are trying to improve the wrong component
The development of generative AI naturally directs attention toward the question:
How do we build a model that gives even better answers?
In creativity and complex problem-solving, it may be worth asking the question differently:
How do we build a human-AI system capable of reaching places that neither the human nor the AI would probably reach alone?
Then the unit we are trying to improve is no longer a single model.
It becomes the entire cognitive process.
Sometimes AI should generate.
Sometimes it should organize.
Sometimes it should challenge assumptions.
Sometimes it should hand the initiative back to the human.
And sometimes the most valuable property of the system may be that one part of it does not know what all the others came up with.
This reverses a popular intuition about AI systems.
We do not always need maximum memory.
We do not always need maximum context.
We do not always need an immediate answer.
Perhaps the next step in the development of human-AI collaboration is not a model that remembers more. Perhaps it is a system that knows when to remember, when to forget, and when to stop answering so that the human can think.
References and publications
[1] Samangi Wadinambiarachchi, Ryan M. Kelly, Saumya Pareek, Qiushi Zhou, Eduardo Velloso, The Effects of Generative AI on Design Fixation and Divergent Thinking, Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, Article 380, 1-18, 2024. DOI: 10.1145/3613904.3642919. DOI
Publication in the ACM Digital Library
[2] Anil R. Doshi, Oliver P. Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, 10(28), eadn5290, 2024. DOI: 10.1126/sciadv.adn5290. DOI
Full publication in Science Advances
[3] Nicholas Davis, Jacob Sherson, Janet Rafner, The Co-Creative Design Framework for Hybrid Intelligence, Proceedings of the 2025 Conference on Creativity and Cognition, pp. 560-572, Association for Computing Machinery, 2025. DOI: 10.1145/3698061.3726934. Pure
Publication via DOI
[4] Juan Carlos Quiñones-Gómez, Hybrid Human-AI Co-Creation in Design, CERN IdeaSquare Journal of Experimental Innovation, 10(2), 164-170, 2026. DOI: 10.23726/cij.2026.1857. e-publishing.cern.ch
Full publication in CERN IdeaSquare Journal of Experimental Innovation
[5] OpenAI, GPT-6 Astra: A new generation of intelligence, 2026. OpenAI
Official GPT-6 Astra page