John Gibbard

Perfection has become a tell. Too smooth, too balanced, too… AI. In a world of generative everything, we’ve reached a strange inflection point: human-centred design now demands imperfection, not as a flaw, but as a feature.

Because friction is fidelity. And too much polish starts to smell synthetic.

When Random Doesn’t Feel Random

Apple’s original shuffle algorithm was mathematically pure, each track had an equal chance of playing. But users complained. It didn’t feel random. Why? Because true randomness includes clumps, repetitions, patterns that seem suspect. A couple of U2 tracks in a row and suddenly the algorithm was “broken.”

So Apple redesigned it to be less random, so that it would seem more random. Illusion of imperfection, engineered.

It’s the same with LLMs. Outputs that are too balanced, too polished, ring false to human ears. We need to start prompt-engineering flaws into our copy, because believability demands mess.

Breath Marks and Broken Grammar

I have a playlist I use to test audio (car stereos, headphones, my hi-fi separates). Not for punch or clarity. For something else. For breath, for scratches, for the hiss of it all. Those tiny artefacts left in because they mattered. Because someone chose to keep them.

Same applies to writing. I’ve airbrushed things too, of course I have. Smoothed over copy that should have stung a little. A good sub-editor knows when to let a clause run ragged. When tidiness would kill the tone. And when grammar should yield to cadence. On LinkedIn, where polish often passes for credibility, that kind of mess is rare, although the smell of all that polish is punget.

The Pratfall Effect, Real Beauty

The Pratfall Effect teaches us that we trust people (and brands) more when they’re good and a little flawed. A genius who spills coffee. A leader who admits doubt. We warm to it.

Brands have learned this too. Dove’s Real Beauty Playbook (developed by Unilever) resonates because it shows unedited reality: pimples, pores, and all. At a recent session in London, Bianca Mack (WongDoody) reminded us of the campaign’s emotional resonance and shared new research on how people respond to AI-generated images and the labelling thereof.

But who decides when an image is ‘too perfect’? That wasn’t clear.

One crucial gap: that research didn’t appear to distinguish between image types. As I explored in my previous piece, “The Complexity and Nuance in Human-Centred Design: Beyond the ‘Average’ User” (Nov 2023), we don’t experience products or content generically, we interpret them through the lens of context, emotional expectations, and domain norms. We tolerate gloss on cars and watches. But we demand scars and breath in human faces.

Against the Plinth: Notes from JLR

When I led UX at AccentureSong for Jaguar Land Rover, we had this tension constantly. The art directors wanted visual purity: architecture that gleamed, cars posed like sculpture, not vehicles. But my team pushed back. We knew that real customers didn’t experience their Range Rover on a plinth. They experienced it on wet roads, in dim light, with children kicking the back seat. Our interfaces and imagery needed to feel lived in, not gallery-lit.

There was always a pull between the pristine and the plausible. Between the brand fantasy and the user reality. The best work came when we embraced the rough edge.

This picks up the argument I made in “The Complexity and Nuance in Human-Centred Design”, that designing for an average user strips out useful extremes. Here, it’s visual: perfection may be aesthetic, but it’s not trustworthy.

Prompt Engineering with Bruises

If you’re using LLMs for writing, design, or strategy, you’ll notice: the cleaner it reads, the less it lands. That might sound odd, but the flaws make it human.

Try this instead:

• Prompt for contradiction: “Add a small, unresolved tension.”
• Prompt for failure: “Include a misstep or wrong decision.”
• Prompt for tone: “Make this sound slightly defensive.”

These aren’t weaknesses. They’re realism. They’re humanity.

You don’t say it’s real, you show it.

Which brings us back to humans. We now have a new role: not just creators, but curators of believability. If you let a model spit out 800 words of polished perfection and ship it unchecked, don’t be surprised if your readers scroll past. They know what machine-made sounds like.

In Bianca’s research, people wanted to know when something was AI-generated, but perhaps more importantly, we want to know someone’s checked it. Because a watermark is more than a label. It’s a sign of judgement.

Just as we caveat car ads, ‘closed course’, ‘professional driver’, we’re now being asked to signal artifice across digital domains. Not to apologise for it, but to own it.

It’s not easy. As an industry we’re conditioned to edit out blemishes, not protect them. Maybe we’ve all just got too good at pretending we know what authenticity looks like.

Closing

Perfection doesn’t reassure. It can repel. If you want something to feel real, it needs to breathe. To blink. To bruise. In a world of frictionless content, the rough edge is where trust begins.

This isn’t a stylistic preference. It’s a principle.

AI was used to sub-edit this piece according to my personal tone of voice guidelines, it was also used to generate the cover image, WordPress excerpt, tagging recommendations and tighten the LinkedIn tease for it.