MCN 2026

Many museum digital projects have an accessibility gap. Images go without alt-text not because institutions don't care, but because writing it at scale – across hundreds of objects, for rotating content, with limited staff – is genuinely hard. When Bloomberg Connects introduced AI-generated alt-text into its CMS, the goal was to close that gap. What we learned was something different: that trust in an AI tool isn't primarily about what the AI can do. It's about what the human can still decide.

This session tells that story from both sides of the table.

In a structured 9-institution pilot, we ran 450+ images through the tool and measured every interaction. Every metric outperformed its target. But the most striking result wasn't the acceptance rate or the time savings; it was that 100% of pilot partners reported feeling in control of the final alt-text. That didn't happen by accident. It was the product of specific design choices: a clear visual separation between AI suggestion and human-editable field, an explicit accept or reject step, no auto-population. The design assumed that curatorial judgment was non-negotiable, and partners felt it.

The pilot also surfaced harder questions that design decisions alone couldn't resolve. When the AI consistently omits skin tone in portraits, is that a safe default or an act of erasure? When it recognizes one religious figure and not another, what does that reveal about whose cultural knowledge is encoded in the model? Partners didn't just want to know what the tool could do: 7 out of 9 asked us to document what it wouldn't do, and why. That transparency mattered as much as the output.

By the time of this session, the tool will have been live across Bloomberg Connects' full partner network for 5 months. This panel draws on the controlled pilot findings and the real-world launch data together, giving attendees both the structured experiment and the longitudinal test of whether what worked with 9 institutions holds at scale.

The session is structured as a panel of co-presenters, not a product demo. Each presenter brings a distinct perspective and a specific artifact – a real image, a before-and-after alt-text, a moment where they overrode the AI – to ground their contribution in practice. Attendees will leave with frameworks they can apply to their own AI adoption decisions, not as a question of capability, but as a question of institutional fit, governance, and what trust actually requires.

Session structure:

10 min - Bloomberg Connects platform framing: the accessibility problem and the design hypothesis       

10 min - Wolfsonian-FIU's experience what they expected, what surprised them, anchored in a specific example from their collection

10 min - Baltimore Museum of Art's experience: where the tool held up and where it raised harder questions, anchored in a specific example from their collection

15 min - Group discussion & Q&A: building across all 3 perspectives on trust, control, and what they'd do differently