Hey Bob, thanks for sharing these experiments. It's really cool to see this technology being tested for this kind of museum work.
I think the broader concerns around AI in museum tech are real, and the point above about disclosure is always going to matter. The same goes for accuracy. Using multiple tools or models to fact-check, as you've been doing, is one of the best ways to reduce bad output. And the more context you can give the model from your own collections, archives, interpretive materials, and institutional knowledge, the more you reduce the chance of generic or inaccurate content.
Tone and human insight are still the harder parts. AI can summarize, organize, draft, and generate at scale, but it doesn't replace the judgment of a trained professional who knows what matters, what's delicate, what's missing, and what needs a human touch. But with a human in the loop to review, edit, reshape, and add that insight, I think you really do get the best of both worlds.
I understand the discomfort around all of this. But most major technological and cultural changes involve tradeoffs. Here, the tradeoff is the possibility of some inaccuracy, though I suspect that risk can be made quite small with the right process, and honestly may compare favorably with ordinary human error. There's also the issue of tone, where AI often lacks the depth and texture a curator, historian, educator, or interpreter can bring.
But on the other side of the tradeoff is something very meaningful: the ability to surface and share resources from collections that might otherwise stay hidden. That's one of the ongoing realities of museum work. Visitors see what's on view, but there are often many times more objects, stories, documents, and perspectives sitting behind the scenes. Not because anyone is neglecting them, but because everyone is time-limited. There is only so much interpretive content a staff can produce by hand.
That's where I think these tools become genuinely exciting. They can help museums begin to release more of those stories, promote more of the collection, and create more points of access for the public. So while the concerns about perception, accuracy, and tone are real, I think the potential to multiply access to cultural material is probably well worth exploring.
And I completely agree that if someone is testing AI for this kind of work, they really need to be testing the best frontier models, ideally with a paid subscription. If you're only experimenting with free or older models, you're not really seeing what the technology can do.
One thing I'd add on the narration side: ElevenLabs is excellent, but even there you can often still hear the AI in the cadence. It's usually a little too polished. That's the problem with a lot of AI-generated content, whether audio, music, or visuals. It often lacks the small irregularities that make something feel human.
A trick I've used is to read the transcript myself first, then use that recording as the performance reference in ElevenLabs. You can still apply whatever consistent voice you want for the final narration, but it follows your actual pacing, pauses, emphasis, and speech rhythm. That brings back a subtle layer of human cadence that you don't quite get when you simply hand ElevenLabs a transcript and ask it to generate the voiceover from scratch.
------------------------------
Eric Holter
CEO
Cuberis
Apex NC
------------------------------
Original Message:
Sent: 06-29-2026 02:52 PM
From: Bob Pritchett
Subject: Flexibility of AI Guides vs. Museum Tone and Phrasing
Someone asked in a direct message about the AI I used... so here's the process and the editor I use for human-review/script editng.
------------------------------
Bob Pritchett
Bellingham WA
------------------------------
Original Message:
Sent: 06-27-2026 02:24 AM
From: Bob Pritchett
Subject: Flexibility of AI Guides vs. Museum Tone and Phrasing
I know I'm late to this conversation, but I've been thinking about related issues. I built a tool that creates a tour of one or more photos using AI, in "Ken Burns-style" pan-and-zoom format. The tool takes the artifact, information about the museum, and optional specific instructions and then constructs a multi-stop 'tour' of the artifact.
The tour is technically finished and ready to use (or export to video for social media) immediately, but it also opens up an editor where every piece of it can be reviewed or replaced by a human curator, even re-narrating in the same AI voice.
In my experiments I have mixed feelings; the tours can be simplistic because the AI doesn't know enough, but on the other hand almost every one I've made has highlighted a detail I had missed despite having seen the photo many times before, and it often pulls out at least one insightful connection I'd missed.
And I'm finding the models are improving; just today I switched to a newer frontier model and got noticeably more interesting results in the first draft.
Here are a couple draft examples; these are mostly AI-generated, with no content input from me.
1859 photo of the First Bank of the United States
B. B. Grocery Co. storefront circa 1905
-- Bob
------------------------------
Bob Pritchett
Bellingham WA
------------------------------