NABL & ISO/IEC 17025:2017 Calibration Reference Portal
Official Lab Services: multitek.in
Solutions for: Quality Heads & Auditors Shop Floor Inspectors Tool Room Technicians Engineering Students

AI Design Fixation: Why More AI Images Can Mean Fewer Engineering Ideas


Ask an image-generative AI system for ten engineering concepts and it can produce them in minutes.

At first glance, that looks like powerful divergent thinking. Ten images appear to provide ten possible directions.

But there is an important engineering question hidden behind the image count:

How many of those outputs are genuinely different concepts, and how many are variations of the same underlying idea?

This problem is increasingly being discussed as AI design fixation.

Design fixation occurs when ideation remains anchored around a limited set of principles, features or existing examples. The output may continue changing visually while the underlying solution space changes very little.

For engineering teams using generative AI during conceptual design, that distinction matters. A different-looking enclosure is not necessarily a different mechanism. A new rendering style is not automatically a new engineering concept.

What is design fixation?

Design fixation is a long-established concept in design research. It describes the tendency to remain attached to known examples, features or solution approaches even when other alternatives may exist.

Fixation can happen to human designers.

It can also appear in the outputs produced by generative AI.

And a third possibility is particularly important: an AI-generated image can itself become an anchor that influences what a human designer thinks of next.

That means AI can participate in fixation at more than one point in the workflow.

The problem is not simply that AI repeats an identical image

AI design fixation can be much more subtle than obvious duplication.

Imagine asking an image generator to develop a new portable measuring instrument.

The system produces:

  • a black handheld instrument with a rectangular display;
  • a blue handheld instrument with a larger display;
  • a rugged handheld instrument with rubber protection;
  • a futuristic handheld instrument with a touchscreen;
  • a compact handheld instrument with rounded edges.

These images may look different.

But suppose all five use essentially the same:

  • instrument architecture;
  • sensor placement;
  • measurement principle;
  • operator interaction;
  • display arrangement;
  • data acquisition method.

From an engineering-design perspective, the AI may have produced five visual embodiments of approximately one conceptual family.

This is why counting generated images can exaggerate how widely the design space has actually been explored.

What the 2026 engineering-design study found

A 2026 study published in Artificial Intelligence for Engineering Design, Analysis and Manufacturing compared engineering-design students with six image-generative AI tools.

The researchers evaluated creative performance using four separate dimensions:

  • Quantity: number of solutions produced;
  • Variety: how meaningfully different those solutions were;
  • Novelty: how unusual or original the solutions were;
  • Quality: how effectively they addressed the design problem.

This separation is crucial.

An AI system can score strongly in one dimension without performing equally well in another.

AI's biggest advantage was quantity

There is little surprise here. Generative systems can produce many images rapidly.

However, the researchers deliberately avoided treating generation speed as proof of greater creativity. Quantity was primarily considered a process-related variable because computational systems can naturally produce output faster than human designers.

The meaningful comparison was what happened to variety, novelty and quality once output quantity was controlled.

Variety revealed a much bigger weakness

The difference became clearer when the researchers examined conceptual variety.

The human designers achieved a median variety score of 0.45 and a mean of 0.60.

When the AI systems generated four concepts, their variety scores ranged only from 0.08 to 0.17.

When they were allowed to generate 16 solutions, some systems improved. The best result reached 0.49, but none of the evaluated AI systems exceeded the human mean.

The human-AI difference in variety was statistically significant in the experiment.

This is one of the strongest warnings against evaluating an AI ideation session by image count.

More outputs expanded the design space somewhat, but did not automatically produce human-level conceptual diversity.

Novelty tells a different story

The same research did not show AI failing across every creative dimension.

AI performance in novelty was considerably more competitive.

In the four-solution condition, the AI median novelty score matched the human median. When certain systems generated 16 concepts, some reached the maximum novelty value observed among the human designers.

This creates an important distinction:

An idea can be novel without a complete set of ideas being diverse.

An AI system might occasionally produce a surprising concept while still concentrating most of its outputs around a relatively narrow family of solutions.

Engineering teams should therefore avoid combining novelty and variety into one vague judgement such as "creative."

Why generating another 20 images may not solve fixation

A common response to repetitive AI output is simply to request more.

This can help, but the research suggests it is not a complete solution.

Increasing generation from four to sixteen concepts generally improved variety across the evaluated systems, but the size of the improvement differed substantially between tools.

Some systems improved considerably. Others remained relatively repetitive.

More importantly, generating additional solutions did not reliably improve quality.

This means an engineer could end up with:

more images + more review work + only a modest increase in actual design-space coverage.

AI can also fixate the human designer

The problem is not limited to the model's output.

A separate study presented at the 2024 CHI Conference investigated how exposure to AI-generated imagery affected visual ideation.

Researchers compared participants receiving generative-AI support with other conditions.

In that experiment, participants using the AI image generator showed greater fixation on an initial example. They also produced fewer ideas with lower variety and originality than the baseline group.

The researchers concluded that the result depended partly on how people constructed their prompts and how they responded to the generated suggestions.

This suggests a feedback loop:

initial idea → prompt → AI interpretation → generated image → human attention → next prompt

If every stage reinforces the same features, the ideation process can become progressively narrower.

Why photorealistic AI images may be especially persuasive

A rough sketch clearly looks unfinished.

A polished AI rendering is different.

It may include realistic materials, lighting, controls, fasteners, product proportions and manufacturing-like details.

That visual confidence can make an undeveloped concept feel more mature than it actually is.

An engineer may unconsciously begin solving problems within the AI-generated architecture instead of questioning whether that architecture should exist at all.

This is especially dangerous during conceptual design, where the purpose is still to explore fundamentally different approaches.

A useful test: remove the styling

One simple way to identify possible fixation is to describe each concept without appearance-related language.

Ignore:

  • colour;
  • surface finish;
  • rendering style;
  • product styling;
  • decorative geometry;
  • background;
  • camera angle.

Instead describe:

  • what physical principle performs the function;
  • how energy or information moves through the system;
  • how the operator interacts with it;
  • how components connect;
  • how sensing occurs;
  • how adjustment occurs;
  • how maintenance is performed.

If multiple AI images collapse into essentially the same description, they probably belong to the same conceptual family.

Do not ask for "another different design"

The uploaded engineering-design study is particularly useful here.

The AI tools were repeatedly requested to generate another different concept. Yet the researchers still observed limited variety across many of the outputs.

That suggests the word different is often too weak as an engineering constraint.

Instead of requesting:

"Generate another different design."

change the dimension of exploration explicitly.

A better AI defixation workflow for engineers

1. Start without an AI image

Write the problem in functional terms before generating visual concepts.

For example:

  • What function must be achieved?
  • What inputs and outputs exist?
  • What environmental constraints apply?
  • What must be measured?
  • What must be controlled?
  • What failure modes matter?
  • What user interaction is necessary?

This reduces the chance that the first attractive image silently becomes the definition of the problem.

2. Generate principles before embodiments

Ask for alternative ways of performing the required function before asking the image model to visualize them.

For example, depending on the engineering problem, alternatives might deliberately explore:

  • mechanical approaches;
  • optical approaches;
  • magnetic approaches;
  • pneumatic approaches;
  • hydraulic approaches;
  • electrical approaches;
  • passive versus active systems;
  • contact versus non-contact measurement.

These are examples of categories, not recommendations for a specific product. The appropriate principles depend on the actual engineering requirements.

3. Use one prompt per concept family

Do not let one generated image become the parent of every subsequent variation.

Create separate prompt branches for genuinely different principles.

For example:

Branch A: solve the function using mechanical displacement.

Branch B: solve it using optical sensing.

Branch C: solve it without direct contact.

Branch D: minimize moving components.

Branch E: prioritize field repairability.

The objective is to create separate regions of the design space before visual refinement begins.

4. Change constraints deliberately

Another defixation technique is to force the problem into different operating conditions.

Ask how the concept would change if:

  • space were severely limited;
  • the operator wore gloves;
  • the device operated outdoors;
  • power were unavailable;
  • maintenance access were restricted;
  • measurement had to be non-contact;
  • cleaning requirements became critical;
  • manufacturing volume changed substantially.

These constraint shifts can reveal alternatives that cosmetic prompt variations may never reach.

5. Separate generation from selection

Do not immediately improve the first plausible AI image.

First create several concept families.

Then compare them.

This maintains the separation between divergent thinking and convergent selection.

Once a team begins refining one attractive concept, the cost of abandoning it psychologically and practically can increase.

6. Create a concept-family matrix

A simple matrix can expose repetition quickly.

Concept Physical Principle Working Principle Architecture User Interaction Key Difference
A [define] [define] [define] [define] [define]
B [define] [define] [define] [define] [define]
C [define] [define] [define] [define] [define]

If several rows contain essentially the same answers, the image set is probably less diverse than it appears.

Use the four-metric check

A practical evaluation system adapted from the creativity framework used in the 2026 study is:

Quantity

How many candidate concepts exist?

Variety

How many genuinely different solution families exist?

Novelty

How unusual are the approaches?

Quality

How effectively does the best concept satisfy the actual engineering problem?

These questions prevent a common mistake:

confusing visual abundance with engineering creativity.

AI itself may exhibit design fixation

More recent research has begun examining fixation within generative systems themselves.

A 2025 ASME study involving designers examined how design fixation appears in both text-generative and image-generative AI at the level of design features.

The researchers argue that viewing generative AI through the lens of fixation can help explain why systems have difficulty consistently producing diverse and novel design outcomes.

This gives engineering teams another reason not to assume that repeated generation automatically performs exhaustive exploration.

But AI is not automatically a creativity blocker

The evidence is not one-sided.

Other research has found that generative AI can support exploration and may help users move beyond some forms of fixation when the interaction is structured appropriately.

This is an important distinction.

The useful engineering question is therefore not:

"Does AI cause fixation?"

It is:

"Under what workflow does AI broaden the design space, and under what workflow does it reinforce the first solution family?"

The role of prompting

Prompt engineering alone cannot guarantee creativity, but prompts influence which part of the solution space the system explores.

Weak prompt:

"Design another innovative version."

Stronger conceptual instruction:

"Ignore the previous embodiment. Preserve only the required function. Develop a concept using a different physical principle, different component architecture and different user interaction. Explain those differences before producing the visualization."

The second instruction explicitly identifies what must change.

That provides a stronger basis for evaluating whether the output actually represents a new concept.

Use text AI and image AI for different jobs

A useful human-AI workflow can separate reasoning from visualization.

Stage 1: problem definition
Engineer defines requirements, constraints and uncertainties.

Stage 2: conceptual expansion
Text-based AI helps enumerate alternative principles and architectures.

Stage 3: human review
Engineer removes impossible, redundant or irrelevant directions.

Stage 4: visualization
Image-generative AI visualizes selected concept families.

Stage 5: fixation check
Concepts are compared by principle and architecture rather than appearance.

Stage 6: engineering development
Selected concepts move into calculations, CAD, simulation and detailed requirements.

Stage 7: verification
Engineering claims are tested through appropriate measurement, inspection and validation.

Where metrology enters the AI design workflow

An AI-generated image does not define a measurable product.

Before the concept can be engineered, visual features must become technical requirements.

For example:

visual concept → functional requirement → engineering characteristic → nominal value → tolerance → measurement method → uncertainty → acceptance decision

This is where metrology becomes highly relevant.

An AI image may show two components fitting perfectly together. It cannot establish from appearance alone:

  • the required clearance;
  • the permissible geometric variation;
  • the measurement uncertainty needed;
  • whether the proposed tolerance is manufacturable;
  • whether the characteristic can be inspected;
  • whether the measurement system is suitable for the acceptance decision.

[INTERNAL LINK: Measurement Uncertainty on a Calibration Certificate]

[INTERNAL LINK: Dimensional Calibration Practical Guide]

[INTERNAL LINK: Calibration vs Verification]

A practical anti-fixation checklist

Before selecting an AI-generated engineering concept, ask:

  • Have we explored more than one physical principle?
  • Have we explored more than one working principle?
  • Are the architectures genuinely different?
  • Would the concepts still look different if colour and styling were removed?
  • Did the first AI image influence every later prompt?
  • Have we deliberately changed constraints?
  • Have we generated concepts without showing the model the previous image?
  • Have humans proposed alternatives independently?
  • Have we separated ideation from concept selection?
  • Can every important design claim eventually be verified?

If most generated outputs collapse into one solution family, generate different principles rather than simply requesting additional images.

Do not optimize the rendering before choosing the principle

This may be the most practical rule.

Generative image tools make refinement extremely tempting.

Once an image looks convincing, users naturally want to improve the buttons, materials, dimensions, handles, displays and other details.

But early refinement can consume attention that should still be spent asking:

Is there a fundamentally better way to solve the problem?

Conceptual engineering design should therefore proceed from:

function → principle → architecture → embodiment → detail

not:

first attractive AI image → endless visual refinement.

Frequently Asked Questions

What is AI design fixation?

AI design fixation describes a situation in which generative-AI ideation remains concentrated around similar features, principles or solution patterns instead of exploring meaningfully different concepts.

Can AI-generated images cause human design fixation?

They can under some conditions. A controlled CHI 2024 experiment found that participants using an AI image generator showed greater fixation on an initial example and produced fewer, less varied and less original ideas than the baseline group in that experiment.

Does generating more AI images solve design fixation?

Not necessarily. The 2026 engineering-design study found that additional outputs generally improved variety, but the improvement differed between systems and did not guarantee better quality.

Is novelty the same as design variety?

No. Novelty describes how unusual an individual solution is. Variety describes how different a collection of solutions is from one another. An AI can produce a novel idea while the overall set of outputs remains conceptually narrow.

How can engineers reduce AI design fixation?

Define the function before generating images, explore different physical and working principles, use separate prompt branches, change constraints deliberately, compare concepts by architecture rather than styling, and keep human evaluation separate from image generation.

Should engineers stop using generative AI for ideation?

No general conclusion like that is supported by the evidence. Generative AI can support early exploration and novel idea generation. The important issue is how it is integrated into the design process and whether engineers actively check for fixation.

Conclusion

Generative AI has dramatically reduced the cost of producing another design image.

It has not eliminated the harder problem of exploring a design space well.

The research suggests that AI can provide valuable novelty and rapid exploration, but large numbers of images can still cluster around a limited set of underlying concepts.

The engineering response should not be to reject AI.

It should be to build a better workflow around it.

Define the function. Diversify the principle. Separate concept families. Generate visualizations. Check for fixation. Evaluate feasibility. Then measure and validate.

Used this way, generative AI becomes less of an automatic design generator and more of a controlled exploration tool inside a human-led engineering process.

Research basis

This article is primarily informed by Chulvi et al. (2026), Creative Potential of Image-Generative AI Models for Conceptual Engineering Design Tasks, together with supporting research on generative AI, divergent thinking and design fixation.

The findings should not be interpreted as permanent rankings of specific AI platforms. Generative systems change rapidly, and the 2026 study itself cautions that its results represent the evaluated systems and experimental conditions rather than a universal measure of human versus AI creativity.