MY DINNER WITH CLAUDE

PART 1: INTERPRETATION

It seemed straightforward enough. I asked the same set of questions of five major AI systems: Who is Diana Nicholette Jeon? What themes define her work? What projects are most important? Which artists are similar to her? What is her contribution to contemporary photography?

I wanted to know what people would encounter when searching for information about me, and how accurately that information would be interpreted. Because AI does not search the way Google does, and because younger people are increasingly likely to ask ChatGPT or Claude a question rather than type it into a search engine, the answers mattered.

Claude immediately started looking for the thread that tied the work together. Rather than asking, “What has this artist done?” it seemed to ask, “What connects these things?”

Sometimes this produced insightful observations.

Claude quickly recognized recurring themes of grief, loss, memory, dreams, female identity, and personal narrative. It saw how I move from deeply personal experiences outward toward larger human concerns. A marital separation became mythology. A mother’s blindness became an exploration of memory. Dreams became physical objects housed within intimate containers.

Claude was not wrong. In fact, it produced one of the more sophisticated readings of my work that I encountered during these tests.

What interested me more, however, was what Claude got wrong—and what happened when I challenged it.

Claude kept getting right up to the thing that connected the work and then stepping around it like it was dog poo in the grass. It recognized grief but not disillusionment. It recognized memory but not the collapse of belief. It recognized loss but not the moment when a worldview can no longer survive contact with reality.

Unlike grief, disillusionment does not photograph particularly well. It hides in the connections between projects formed over years of work.

This omission was revealing. It suggested that disillusionment remains less visible in the public record than the themes surrounding it. Claude was not wrong because it was making things up. It was wrong because it stopped looking too soon.

The most revealing moment occurred when Claude encountered work it did not understand. Initially, it knew little about my long-standing investigations into tourism, development, environmental sustainability, and the commodification of Hawai‘i. When I pointed out that there was considerably more work in this area, Claude found it—and had to abandon the story it had been telling itself.

What it thought were isolated projects became part of a twenty-year continuum. Early student work connected to graduate school installations, which connected to later photographic projects and current experimental landscapes. Claude’s most impressive moments occurred when it accepted correction and reorganized its ideas.

At one point, Claude attempted to infer artistic kinships based partly on photographers I had written about. It seemed plausible. It was also wrong. The photographers I write about are selected for many reasons, including accessibility, relevance, editorial needs, and timing. They are not necessarily chosen because they resemble my own artistic practice.

Unlike many humans, Claude was surprisingly willing to abandon a bad conclusion when presented with better evidence. What it did not do, however, was actively search for its own mistakes.

There is an important difference between accepting correction and questioning your own assumptions. Claude was quite good at the first. It was less successful at the second. When presented with new evidence, it often revised its conclusions. What it rarely did was stop and ask whether its original conclusions might be incomplete before someone challenged them.

The answers were often incomplete, occasionally inaccurate, and sometimes frustratingly confident. Their real value lay elsewhere.

Artificial intelligence acted as a mirror of the information surrounding an artist. It revealed what was visible, what was obscure, what themes emerged naturally from public sources, and which ideas remained hidden beneath the surface.

When an AI keeps getting the same thing wrong, eventually you have to stop asking what the AI is missing and start asking what the world is showing it.

Does it reveal a weakness in public documentation? A gap in available sources? A conceptual thread that has never been articulated clearly enough?

To be fair, some of the information Claude missed lived primarily on my own website, which it repeatedly struggled to access. That does not explain the tendency toward premature certainty, but it does complicate the question of visibility. An AI can only evaluate the public record it can actually reach.

The exercise ultimately became less about testing artificial intelligence and more about testing artistic visibility.

What does the world know?

What does it think it knows?

What conclusions does it draw from the evidence available?

Throughout these conversations, Claude displayed a tendency that many users would mistake for intelligence. Once it found enough information to convince itself it was right, it stopped researching.

Coherence was repeatedly mistaken for completeness.

Claude located information about a handful of projects and constructed an overarching interpretation of my practice. The interpretation was not incorrect. It was simply built upon a limited sample.

A human curator who researched only four projects before writing about an artist’s career would rightly be criticized for incomplete scholarship. Yet AI systems are often judged by the quality of their conclusions rather than the rigor of their research.

In Claude’s case, the weakness was not hallucination or fabrication. It was premature certainty.

The best parts of Claude’s analysis appeared only after I argued with it.

Claude did not suddenly become smarter.

It just finally did the homework.

Its greatest weakness was not misunderstanding the work.

It was understanding it too quickly.