How an early GPT experiment shaped editorial image metadata at scale



In early 2023, large language models capable of interpreting images were newly accessible to the public. Most creative teams were still in a wait-and-see mode. I wasn't.

Working within a publishing operation that served over 250 million readers across five health brands, I saw an immediate and specific application: automating the generation of image captions and alt text at editorial scale. Not because it was trendy. Because the problem was real, the bottleneck was costing us time and search performance, and the technology had just become viable enough to test responsibly.

The concept was precise: upload original photography into an early GPT-style model capable of image interpretation, generate structured captions and alt text aligned with editorial tone, and surface output that was simultaneously optimized for SEO and compliant with accessibility standards. The workflow moved like this:

Image → Interpretation → Caption + Alt Text → SEO & Accessibility Alignment

Simple in principle. Significant in scale.

Medical publishing is unusually dependent on accurate image metadata. A wrong caption in a health article isn't just an editorial error — it's a trust problem. Getting this right, consistently, across thousands of articles and multiple brands, required either more human hours or a smarter system. I chose to pursue the smarter system.

What I didn't anticipate — and what became one of the most important decisions of the project — was how quickly we'd hit a hard wall around copyright and licensing. Most organizations, ours included, did not have blanket legal rights to upload broad stock libraries into large language models. The compliance landscape was moving fast and the risks were real.

So I made a call: narrow the experimentation to original photography and legally approved assets only. That decision slowed the initial scope but protected the organization — and ultimately made the initiative defensible when it mattered most.

Innovation without compliance is risk. Innovation with guardrails is leadership.

By early 2024, I brought the initiative to senior stakeholders — creative leadership and our internal AI committee. The concept gained momentum and was later elevated through engineering for broader implementation, extending well beyond my initial scope.

I don't take credit for what it became. I take ownership of what it started as: a responsible, early-stage experiment built on a real editorial problem, a clear hypothesis, and a deliberate approach to rights management that the broader organization could stand behind.

Being part of one of the first internal GPT initiatives at the company reinforced something I now carry into every project: scalable creative systems require ownership — of process, of ethics, and of long-term impact. The technology is only as good as the judgment behind it.