When AI Keeps Adding: Testing Scope Inflation in Architectural Design

A homeowner asks a design team to improve an awkward entrance. An AI assistant proposes another storage unit, a decorative screen, a bench and integrated lighting. Each suggestion sounds useful. Together, they may leave less space for the activity that prompted the request: arriving, putting things down and moving into the house.

This is a hypothetical design situation, not a recorded project failure. It frames an architectural question suggested by behavioural research: when people and AI generate improvements, do they adequately consider removing or avoiding an element before specifying another?

The research behind the question

In Nature in 2021, Gabrielle S. Adams and colleagues reported eight experiments in which participants overlooked useful subtractive changes. Explicit reminders, repeated opportunities and cognitive load affected the tendency. The finding concerned how alternatives were generated; it did not establish that fewer components always produce better outcomes.

A 2026 study by Lydia Uhler and colleagues in Communications Psychology compared humans with GPT-4 and GPT-4o on spatial and linguistic tasks. Additive responses were more prevalent in the models overall, but sensitivity to efficiency and instruction wording varied by model and task. These findings describe the tested systems and conditions, not every AI tool or architectural workflow.

A bias with important boundaries

Joshua Juvrud, Laurence Myers and Pär Nyström’s 2024 study in Scientific Reports found only partial support for a general subtraction neglect. Performance differed across tasks, ages and samples from Sweden and the United States. That variation matters: a tendency observed in one task is not a universal rule about designers.

For architecture, the relevant distinction is between rejecting a subtractive option after assessing it and never generating that option. A team may have good reasons to keep a partition, cabinet or door. The research question concerns whether those reasons are tested against alternatives or simply inherited from the first proposal.

A specific architectural test: the residential entrance

ArchUp Research Lab proposes a controlled design study using an entrance with measured circulation constraints, storage requirements and a defined privacy objective. It would examine proposals on drawings and in a mock-up before any construction decision. No trial has yet been conducted.

Possible responses might include adding a shallow cabinet, repositioning existing storage, omitting a proposed decorative divider or retaining the original arrangement. None should be presumed superior. Removing a divider could improve movement but expose the interior to visitors; reducing storage could merely relocate clutter to another room.

The brief must therefore state what success means before options are generated: required storage capacity, an agreed sightline condition, usable movement space and a realistic arrival task. Mandatory access, structural and fire-safety requirements remain constraints to be checked by qualified professionals, never variables to trade away for a smaller scope.

Separate the effect of structure from the effect of subtraction

Three groups would receive the same plans, dimensions, needs and constraints. The first would receive a conventional request to improve the entrance. The second would receive a structured brief requiring a fixed number of alternatives, each with a functional justification and a record of trade-offs. The third would use the same structure but explicitly require consideration of addition, omission or removal, rearrangement and no change.

Comparing the first two groups tests whether structure alone improves the proposals. Comparing the second and third isolates the additional influence of naming subtractive and no-change options more closely. All groups should receive the same number of proposal slots and comparable working time. Case variations and random assignment would reduce the influence of an unusually easy brief.

For an AI-assisted arm, record the model version, date, instructions and available inputs; use independent sessions and retain unsuccessful outputs. For a human arm, record relevant experience. Neither multiple answers from one model nor repeated proposals from one designer should be counted as unrelated independent participants.

Measure the result, not the number of deletions

Assessors who do not know which brief produced each option would evaluate its functional adequacy. A mock-up could test arrival with bags, opening storage, movement conflicts and the specified sightlines. Drawings alone cannot establish how occupants will use the space.

Record which option types were proposed, which survived evaluation and why the others failed. Compare capital cost, installation work and plausible maintenance commitments using stated quantities and assumptions. Keep existing components, new purchases, relocations and removals separate: a short shopping list can conceal expensive demolition and making good.

The main question is whether the explicit comparison produces more viable alternatives and a better-performing selected proposal within the same brief. A higher deletion count is not success. Neither is a lower purchase price if storage, privacy or access deteriorates. Define the evaluation criteria beforehand and use a pilot to determine the sample needed for a meaningful comparison.

What a design office could learn

The proposed output is a decision record: what was added, retained, moved or omitted; which function each decision serves; and the consequences accepted. This could distinguish a considered reduction in scope from cost cutting that transfers inconvenience to the occupant. Any claim of savings or reduced environmental impact would require project-specific assessment. The immediate research contribution is a testable question about the alternatives a design process permits, not a promise that subtraction will win.

✦ ArchUp Editorial Insight

Professional incentives can make visible additions easier to justify than a decision to leave something out. Where a commission rewards the production of options, drawings or specified products, an extra element has a recognisable deliverable; an avoided element needs an argument about performance that never became a purchase. AI could accelerate this imbalance if its suggestions enter design reviews as a ready-made list of improvements. The studies do not demonstrate that outcome in practice, but they expose a question worth testing: which alternatives reach the decision table before the team starts comparing prices? This connects with ArchUp’s analysis of individually approved information competing within one navigational scene, where acceptable components can produce an untested combined effect. In both cases, reviewing each addition separately leaves responsibility for the whole unresolved. If maintenance obligations emerge after handover, the operator may inherit consequences absent from the original selection criteria. A useful design review would therefore record the function each element serves, the alternative considered and who assumes its continuing upkeep. The entrance layout becomes the consequence of that decision structure: either an accumulation of individually defensible purchases or a coordinated response to a defined brief.

References

• Adams, Gabrielle S.; Converse, Benjamin A.; Hales, Andrew H.; Klotz, Leidy E. People systematically overlook subtractive changes. Nature, 2021.
• Juvrud, Joshua; Myers, Laurence; Nyström, Pär. People overlook subtractive changes differently depending on age, culture, and task. Scientific Reports, 2024.
• Uhler, Lydia; Jordan, Verena; Buder, Jürgen; Huff, Markus; Papenmeier, Frank. Influence of solution efficiency and valence of instruction on additive and subtractive solution strategies in humans, GPT-4, and GPT-4o. Communications Psychology, 2026.