AI in RD&I — Problem 2: Can it be too polished?

It is a classic workshop technique, used in most innovation frameworks to build consensus: small teams develop a chosen idea and present it back to the wider room. Maybe participants score the best ones with little sticky dots, and those ideas go forward to the next round.

These techniques have always carried a competitive edge. Teams want to win the room with their concept, and that competition is usually productive. It drives teams to think harder, articulate the idea more clearly, and push the concept further.

However, a recent experience of participating in one of these sessions worried me. To develop and polish their pitches, teams were all using AI image generation. The competition shifted. Where as once it focus on the strength of the idea and how to clearly identify and distil its value, now it felt more like it had become about the slickest-looking idea, and underneath that, who was best at using an AI image generator.

A group with access to AI image tools can produce a whole slew of images of the idea being used in all kinds of real-world scenarios, in the same time it used to take to do a cartoonish sketch on a flipchart and add a few sticky notes describing key features. The most polished output wins the room — and the response correlates more strongly with visual finish than with how well the concept actually answers the brief.

Just as the use of large language models can create innovation artefacts that are devoid of the necessary consensus-building and shared sense-making — see The Shell Artifact — AI image generators add visual polish that biases our perception of how “real” the idea is. The shell artifact hides missing agreement; the polish problem hides missing proof and short circuits our critical judgement.

“I see a real-looking image, therefore it must be a realistic idea.” 

But in reality, the user research is not yet done, feasibility is not yet tested, and functionality is not yet defined.

The more real it looks, the less we interrogate it

The underlying mechanism is well established in UX research. Studies going back to the early 1990s have shown that high-fidelity prototypes attract materially less critical feedback than rough ones. People critique a sketch, but they admire a render. The higher the fidelity, the more reviewers treat the artefact as settled rather than as something to interrogate.

This was a manageable problem when high-fidelity outputs took time and skill to produce. The cost of creating them acted as a natural barrier, keeping polished artefacts out of the early stages where critical scrutiny matters most.

AI image generation has removed that barrier entirely. A concept can arrive at pitch-deck quality within minutes of being thought of, at any stage of the process, including the stages where rough and provisional was the deliberate design choice.

Recent design research on “seamless AI” reinforces this point: AI output that arrives as finished invites less critical engagement than output that visibly invites elaboration. Their proposed counter-measure is generative friction — deliberately rougher or more provisional AI output.

An AI-generated image that looks like a professionally taken photograph of a person using the product or service at home has no friction. Instead, it hides a myriad of unknowns: is it technically feasible, will it be better than the alternative, and would the sort of person pictured ever want to use it, let alone pay for it?

This is particularly difficult in a workshop or funding competition setting, where, with little time and limited information to judge, the quality of the image becomes a false proxy for the quality of the idea, lending it unwarranted credibility.

Keeping the visual in its proper place

That is not to say that AI-generated images are not useful. These tools can be used to come up with a variety of early concepts, to improve the quality of a 3D render, or to create the product shots for user testing. But we need to be clear about the function of the images we create and the aesthetic used should match that function so as not to mislead i.e. in an ideation workshop limiting the tools to generating sketches and mock-up illustrations.

To use AI well we need to consider the wider unintended consequences, particularly where there is no validation before generation. It is important to remember that image generators are biased, drawing on the aesthetics and content combinations seen most frequently in their training data.. They tend towards a particular polished, high-contrast, advertising-influenced style by default - almost designed to oversell. And from an innovation perspective I worry that they tend to reflect a generic-looking past back to us, rather than radically subverting it, as you might hope the best innovation ideas would do. 

Those biases go much wider than aesthetics, including gender, race, disability and other assumptions that can cause problems if they are not picked up through proper user research. Use the AI too early and rather than prompting with a clearly researched target consumer, the visual design starts making assumptions about who they are which could bias future development. 

So, visual generation is most useful after the core conceptual work has been done, when a team is communicating a direction that has already been tested against user need and feasibility, not as a tool for deciding on the direction in the first place.

In my experience, the most useful discipline is to be explicit about what stage of assessment is happening in the room. If the team is evaluating ideas for user need, feasibility and strategic fit, then visual finish should be declared irrelevant, or a deliberately rough style of image generation should be used throughout. If the team is communicating a direction that has cleared those tests, then by all means, make it look as good as it can.

Polish has its place. But it should be the finish applied to an idea that has earned it, not the evidence that the idea is good.