Fields Medallists vs AI Benchmarks: A Lesson for Marketers

Mathematicians do not usually write open letters about marketing. But the argument a group of Fields Medallists made in September applies almost word for word to how our industry measures AI. Their point: solving problems is a proxy, and the real goal is understanding. Confuse the two, and you can damage the very thing you were trying to improve.

What happened

According to The Next Web, on 11 September a declaration titled A Severe Misalignment of AI in Mathematics was published at mathandai.org. The site lists 25 signatories, including Terence Tao, Peter Scholze, Maryna Viazovska, Cédric Villani, Pierre Deligne, Manjul Bhargava and Martin Hairer. The declaration argues that solving problems is a tool and a proxy, while the primary goal of mathematics is conceptual understanding and insight. It warns that mass-producing true or false statements at an increasing pace "could destroy fertile ground". TNW describes it as a statement about benchmarks; it does not name any specific event.

The same week, a separate open letter criticised a student-run AI "Mathathon" at Caltech, calling the format "slop mathematics" and arguing that the labour of verifying and correcting AI results goes "uncompensated, uncredited, and unacknowledged". TNW reports that OpenAI withdrew its US$1 million in credits from the event on 10 September, while Anthropic kept its contribution, and that the organisers added requirements but did not cancel.

Why it matters

The mathematicians' complaint is about a familiar trap. When a hard, famous problem becomes a benchmark, labs optimise for producing answers to it. Answers are measurable. Understanding is not. Over time, the measurable thing crowds out the valuable thing, and someone else is left to check the output.

Marketing AI has exactly this problem. We benchmark tools on what is easy to count: number of assets generated, time saved, click-through rate, share of AI answers that mention the brand. Those numbers matter. But the goal of marketing is not output; it is understanding customers well enough to serve them, and building a brand people trust. A system that floods channels with content and wins on proxy metrics can quietly erode both.

The unpaid-verification point lands too. Generated content is cheap to produce and expensive to check. If the checking is invisible in your metrics, your "productivity gains" are partly an accounting trick.

What my research says

My recent research is built around this distinction. In our SIGIR 2026 paper, SOMIN: Agentic AI for Automating Professional Visibility and Countering Content Homogenization, the problem we address is precisely that AI-written content converges: everyone's assistant produces the same competent, generic posts. Our approach is to retrieve real context data before writing, so the output stays distinctive. Volume without distinctiveness is the marketing version of "slop".

On measurement, I have argued on e27 that every dashboard is green, yet nobody can say what worked. Proxy metrics that all point up are not the same as knowing what caused results. And in our explainable AI work, from Against Opacity to MindFuse, the aim is not just a better prediction but an explanation a marketer can learn from. That is the marketing equivalent of the mathematicians' "conceptual understanding".

What marketing teams can take from this

• Name the goal behind each metric. For every AI KPI, write one sentence on what real outcome it is supposed to stand for, and check periodically that the link still holds.
• Count the verification cost. Track the human time spent reviewing and correcting AI output and include it in your ROI.
• Measure distinctiveness. Compare your AI-assisted content with competitors'. If it is becoming indistinguishable, you are winning a proxy and losing the brand.
• Prefer explanations to scores. A tool that tells you why something worked builds team knowledge; one that only tells you that it worked builds dependency.
• Credit the humans. If people are cleaning up AI output, make that work visible and valued, not assumed.

The mathematicians are protecting their field from a benchmark culture that mistakes answers for understanding. Marketing should protect itself from the same mistake.

FAQ