our design, tested with AI eye-tracking
Our design and graphic elements were used as test material in a neuromarketing study published in the peer-reviewed journal Informatics (MDPI) in March 2025. The study compared how well designs created by professional graphic designers capture viewers’ attention against a design based on artificial intelligence recommendations – and the result came out in favour of human design. It was an honour to be part of this research.
It is rare for work that leaves a studio to end up in scientific literature, and rarer still for it to be measured with tools that predict where the human eye will actually look. That is why this collaboration means more to us than another portfolio reference.
About the neuromarketing study
The study is titled “Human-Centred Design Meets AI-Driven Algorithms: Comparative Analysis of Political Campaign Branding in the Harris–Trump Presidential Campaigns”.
| Journal | Informatics (MDPI), vol. 12, no. 1, article 30 |
| Published | 18 March 2025 |
| Status | Open Access · Editor's Choice |
| DOI | 10.3390/informatics12010030 |
| Authors | Hedda Martina Šola, Fayyaz Hussain Qureshi, Sarwar Khawaja |
| Institutions | Oxford Centre for Applied Research and Entrepreneurship (OxCARE), Oxford Business College; Institut za neuromarketing i intelektualno vlasništvo (Institute for Neuromarketing and Intellectual Property), Zagreb; SK Hub, Uxbridge |
The research question was simple to ask and hard to answer: can AI design visual communication that holds attention better than the work of an experienced designer?
Our role in the research
Our contribution was the design side of the study. Specifically, we:
- created a complete redesign of the original promotional material,
- designed the graphic elements – typographic hierarchies, logos within the composition, iconography, background layers and colours,
- prepared several purpose-built sub-versions of the same design, each changing specific variables in a targeted way so that their impact could be measured in isolation.
That last point is really the heart of the work. If you change five things at once between two versions, you can’t tell which one made the difference. So the sub-versions used controlled changes: font size and weight, contrast inversion behind the main headline, switching the background from black to white and light grey, adding a secondary logo behind the candidates’ figures, and introducing or removing party affiliation icons.
Everything was created in Adobe Illustrator, with a defined palette (dark blue #18345A, red #A81D46, white #FFFFFF, light grey #E9EAEE, dark grey #2A2C37, light orange #FDD2B5) and Montserrat typography in several weights.
In total, with the original material as a control, four versions were tested:
- Design 0 – original material (baseline, from an Al Jazeera article)
- Design 1 – professional redesign
- Design 2 – professional redesign with additional balance adjustments
- Design 3 – professional redesign based on an AI assistant’s recommendations
How it was measured
This is where the study differs from the usual “I like this one better” approach. Instead of subjective judgement, two tools were used.
Predict – AI eye-tracking
Software developed by Neurons in collaboration with Stanford, trained on around 180,000 tests and a database of approximately 200,000 data points collected with conventional eye-tracking devices. The sample was distributed globally (USA 35%, UK 20%, Nordic countries 20%, DACH 10%, the rest of Europe, Latin America, the Middle East and Asia), with a balanced gender ratio and an age range of 18–55.
Predict forecasts the distribution of visual attention with a stated accuracy of around 95% compared with high-precision eye-tracking – and does so in seconds, rather than the weeks a lab test takes.
CoPilot – a neuroscience-based AI assistant
A language model enhanced with neuromarketing literature, with a benchmark database of 67,429 areas of interest across 13,689 visual materials. It generated the recommendations on which Design 3 was based.
Metrics
Each design was broken down into 13 to 14 areas of interest (AOI) – headline, body text, candidates’ figures, names, logos and event venue details. For each area, the following were measured: total attention, engagement, start attention (first 2 seconds), end attention (last 2 seconds), percentage seen, clarity of message and cognitive demand.
Data processing was done in Python (Pandas, Matplotlib, Seaborn, SciPy, Statsmodels), using Spearman’s correlation, the Kruskal–Wallis H test, ANOVA and the Bonferroni correction.
Results: human design won overall
The authors calculated a weighted performance score combining all five key metrics (total attention 30%, engagement 25%, start attention 20%, end attention 15%, percentage seen 10%).
| Design | Weighted score |
|---|---|
| Design 1 (human redesign) | 88.34 |
| Design 0 (original) | 86.29 |
| Design 3 (based on AI recommendations) | 85.35 |
| Design 2 (human redesign) | 85.02 |
Design 1 also had the highest start attention (99.71 versus 99.61 for the AI version) and the highest engagement in seven of the thirteen areas of interest – making it the most consistent performer across the whole composition, not just on one or two focal points.
Where AI did better
It would be wrong to say AI contributed nothing. Design 3 achieved the highest clarity on the female candidate’s figure (10.77 – the best result across all versions) and a very strong emotional response in the same area (9.39). In other words, AI did an excellent job of optimising individual focal points.
Where it fell behind: text areas. On body text, Design 3 scored 5.31 against 6.13 for the human Design 1. The same pattern repeated across several other text AOIs.
A nuance worth noting
In the analysis of raw attention scores, Design 2 proved strongest in several individual categories (body text 14.31; candidates’ figures 13.52 and 18.70). Design 1 takes the overall win thanks to its consistency across all metrics, not by dominating each one. That distinction matters: depending on whether a campaign aims to maximise a single element or create a balanced overall impression, the “best” design is not the same.
Which hypotheses held up and which didn’t
The authors were disciplined and rejected their own initial hypotheses when the data didn’t support them:
- H1 (AI design outperforms traditional methods) – rejected
- H1a (AI design achieves higher start attention) – rejected
- H1c (AI design achieves higher clarity) – rejected
- H1b and H1d – partially supported
- H2 (combining AI tools with human expertise outperforms either approach on its own) – strongly supported
The study’s conclusion isn’t “AI doesn’t work”, but: AI as a measuring instrument plus a human designer as the decision-maker = the best result. Design 1 wasn’t created blindly – it was created after AI eye-tracking had analysed the original and shown where attention dropped, and the designer then made creative decisions based on that data.
An interesting finding: clarity comes at the cost of cognitive effort
The Spearman correlation between clarity and cognitive demand was strongly positive in all versions (0.89 to 0.96, p < 0.0001).
In practice: the areas that communicate the message most clearly also demand the most mental effort from the viewer. There is no such thing as free clarity. This applies directly to landing pages, banners and packaging – when you “clarify” an element, you often make it more demanding too, and that trade-off has to be distributed deliberately across the composition rather than solved element by element.
Limitations of the study
We are writing this on purpose, because results get over-interpreted in marketing far too often.
- The study is cross-sectional – it shows associations, not cause-and-effect relationships. The authors state this explicitly.
- ANOVA did not show statistically significant differences between the four designs for clarity (F = 0.20; p = 0.899), engagement (F = 0.01; p = 0.999) or total attention (F = 0.04; p = 0.988). The differences in the weighted score are real but modest in size. The Kruskal–Wallis test at the level of individual AOIs did show significant differences in clarity.
- Predict forecasts behaviour based on a model; it doesn’t measure specific people at a specific moment.
- The material is a political flyer. Applying the findings to e-commerce or B2B communication is a reasonable hypothesis, not a proven fact.
What this means for our client work
Three things we take from this into our everyday work:
- Design can be tested before it goes live. There is no need to wait for the CTR after a week of media spend to find out whether the CTA is visible.
- AI is a good measuring instrument but a weak author. The tool’s recommendations are useful for individual focal points; decisions on composition and copy remain the designer’s job.
- Purpose-built sub-versions are valuable in their own right. Varying one variable per version in a controlled way is the only way to find out what actually worked.
Acknowledgements
It was an honour to be part of this neuromarketing study. We would like to thank the authors – Dr Hedda Martina Šola, Dr Fayyaz Hussain Qureshi and Sarwar Khawaja – for their trust and the invitation to take part, as well as the Institut za neuromarketing i intelektualno vlasništvo and Oxford Business College.
Seeing our own work go through peer review and end up as an Editor’s Choice in an international journal is a recognition money can’t buy.
Frequently asked questions
Where can I read the full paper? The paper is Open Access and available in full free of charge: mdpi.com/2227-9709/12/1/30
Did AI design the materials? No. All four versions were created by professional graphic designers in Adobe Illustrator. One version (Design 3) was based on an AI assistant’s recommendations, but it was executed by a designer.
What is AI eye-tracking? A technology that uses a model trained on a large number of real eye-tracking recordings to predict where the human eye will look on a visual – without the need to test it on participants.
Does this mean AI is useless in design? No. The study confirmed the hypothesis that combining AI tools with human expertise gives the best results. It only rejected the claim that AI on its own outperforms a designer.
Citation
Šola, H. M., Qureshi, F. H., & Khawaja, S. (2025). Human-Centred Design Meets AI-Driven Algorithms: Comparative Analysis of Political Campaign Branding in the Harris–Trump Presidential Campaigns. Informatics, 12(1), 30. https://doi.org/10.3390/informatics12010030
The paper is published under a Creative Commons licence, which allows its content, images and tables to be reused with clear attribution of the source.
