A 600-person consumer study, reproduced in Elaiia.

How we verify that AI twins are accurate: a practical packaging case.

Whenever we tell an insights team that an AI panel can stand in for a survey, the same fair question comes back: it is faster and cheaper, which is fine, but how do you know it is right? It is an extremely valid question, and right now an unavoidable one. AI twins have become a hype topic, and hype tends to gloss over the fact that accuracy is not a property of the idea but of the system behind it. Some tools are little more than a thin wrapper around a language model dressed up as a consumer; others are genuine data-driven simulations grounded in real behaviour. That difference is what separates confident hallucination from decision-grade accuracy. Speed means nothing if the answer is wrong, and a simulated consumer that merely sounds plausible is worse than no data at all, precisely because it is confident. So instead of arguing the point, we test it. This edition is one of those tests.

Start with the decision the method is meant to help with, because that is what makes “right” concrete. Every pack redesign is a bet. Which cues actually earn the shopper’s glance: the material claim, the product window, the brand logo, the heritage badge? Answering that the traditional way means recruiting hundreds of category buyers and fielding for weeks, and even then the sample is not always representative. So most packs get redesigned on far thinner evidence than the decision deserves.

Here is how a team would use Elaiia for a question like that. You build a panel of twins to match the study’s screening profile, show them the pack, and ask which of ten design elements they like, the same task a human pack test uses, only you get the read the same afternoon. Useful, obviously. But the output is only ever as good as its fidelity to real people, which brings us straight back to the question: how do you know it is right?

So we checked it against the real thing. In 2023, Triumph’s everyday brand Sloggi ran a full quantitative packaging study on its Basic+ pack: 600 women aged 18 to 55 in Germany, all category buyers, no Sloggi rejecters, each marking which of ten elements she liked. Three years later we gave the identical task to 83 AI twins on Elaiia, with the same profile, the same pack, and the same ten elements, but no access to the original results. Then we put the two panels side by side.

Design-element tiers, human panel vs. Elaiia twins

The two panels agree on everything a designer would act on. Sort each element into a like-tier, top, middle, or low, and all ten land in the same tier in both panels, with an identical top five: product and material cues lead, so the stretch-cotton claim, the product window, the quantity cue, the model and the product photo come out on top, while the background, the comfort-seams claim and the heritage badge trail in both. Across all ten elements, twin and human preferences correlate at ρ = 0.90. Where the two differ, they differ on magnitude, not direction. The twins run a little hotter on the product window and cooler on the brand logo, but none of those gaps move an element into a different tier or change the brief.

The reason this works is also the reason the twins are not a photocopy of the humans, and are not meant to be. Built three years after the fact and blind to the original study, they had nothing to copy. Each twin reproduces the interplay of cognitive processes, personal traits and past experience that shapes a real choice, which is what produces both the agreement and the small divergences. Working from either dataset, Triumph would have prioritised the same cues, dropped the same weak claims, and briefed its designers the same way.

This is the part worth keeping. Trust in a simulation is not something you assume; it is something you measure. We do not ask whether a twin’s answer sounds plausible. We ask whether it reproduces how real consumers think, feel and act, checked against human results wherever we have them.

When the method holds, the payoff is simple: a pack test stops being a six-week gate at the end of the design process and becomes a check you can run on every variant, on the day you need it.

Sloggi (Triumph) tested the Basic+ packaging with 600 women in Germany, evaluating 10 design elements. We ran the same study with 83 AI Twins matched to the same profile, with no access to the original results: 10 out of 10 elements landed in the same preference tier — with far richer qualitative detail, in 1 day rather than 6 weeks.

PREFERENCE CORRELATION: ρ = 0.90