[Nhcoll-l] Where should museums draw the line with generative AI?

Thu Dec 18 10:13:44 EST 2025

Thanks for engaging in further discussion.

I'm not worried about a tool that helps an expert be more efficient.  A
transcript that can be vetted by the writer or team is useful, whether
written by a human or AI, but the final product needs to be a human
product, especially anything published in a museum context.

AI images have similar issues.  If a real artist wants to use AI to explore
compositions, great.  I don't see a difference between that and cutting our
magazine pictures or sketching from life when it's the brainstorming
stage.  But the images used in the exhibit should be creditable to a human
expert.  The SF airport example should have never made it beyond a workshop
draft.

Even analog recreations have challenges.  In one of my galleries, there is
an enlarged to show detail mouse recreation.  To my eye, it looks like a
good-quality child's toy.  Many (maybe most) visitors can't identify or
describe any particular feature that looks fake; they think it is real.  As
a result, questions about (and emotional responses to) giant mice become
central to the exhibit, distracting from the intended points.  As signs
noting that this diorama is fake get bigger and more prominent, our
pedagogical goals are better realized.

Our visitors trust us, and whatever happens in our institutions.  We have
to find ways to help them differentiate between fake and real.  So many of
our visitors engage with AI without any training in its pitfalls, yet it is
packaged in a way that seems trustworthy and fun, and it's often given
logistical preference over any other research tools.  I would not be
surprised to find (this is a study that could be done) that AI/LLM search
results pulled up at a museum are deemed more accurate than in some other
setting.

If anything, museums should be pushing against some uses of AI because it's
not possible to ensure accurate results from the interfaces they use (eg
LLMs, digital assistants).  It seems the average visitor is more likely to
misunderstand our exhibits because of most current AI use, since
there's often few useful answers but scores of red herrings and wrong
answers.  The analog training we can provide is often key to obtaining
accurate digital results.

On a public nature walk, I encouraged people to track our observations
using an app that uses AI to make identifications.  When I use the app, it
is often pretty accurate, approaching 100% for some taxa.  However, when I
reviewed the identifications with individuals on trail and later looked at
the uploaded aggregate, a large percentage of the AI-generated IDs were
wrong.  Sometimes they were wildly wrong, like an insect being IDed as a
plant.

Unsurprisingly, the quality of the photo matters and many people, despite
having amazing and simple to use cameras, can't take a good/useful
picture.  Visitors often don't know enough to input data (be it images or
queries) effectively *and* they don't know enough to vet the information AI
returns.  You might think that everyone can tell the difference between a
plant and an animal, but katydids exist so a marmorated stinkbug identified
as a sycamore means, to a user who is newly learning, that walking sycamore
leaves must be real, too.

Museums are the stewards of authentic objects that allow humanity to
objectively understand parts of the past, interpret the present, and
anticipate the future.  AI, as a tool, may support this in many ways, but
front-of-house, direct-to-consumer AI, or whole-cloth AI products in our
galleries damages all of this.

--Steve

On Thu, Dec 18, 2025 at 7:17 AM Thor Martin Jensen <thormj at gmail.com> wrote:

> Steven,
>
> Thank you for the detailed response. Your examples of AI-generated
> worksheets and misdirected questions are exactly the kind of failures that
> concern me. Systems generating content that sounds plausible but
> fundamentally misunderstands what the institution actually holds or teaches.
>
> I appreciate your point about trust. Museums are among the most trusted
> institutions precisely because visitors know a person with expertise made
> choices about what to show and how to explain it. Breaking that chain is
> risky.
>
> On your concern about the listserv format, you're right that BCC prevents
> "reply all." That was unintentional. I genuinely wanted discussion, not
> just promotion. I should have used a format that allowed group
> conversation. My mistake.
>
> On translation, you're absolutely right that automation has weaknesses,
> especially with specialized terminology. At Walkie Talkie, we don't just
> run content through a translator. Museums review and edit translations
> before publishing. The AI handles the first draft to make 38 languages
> feasible; curators ensure accuracy. It's a production tool, not a
> replacement for expertise.
>
> Your calculator analogy is useful. AI as efficiency tool is reasonable. AI
> as creative authority is problematic. The line matters.
> Best,
> Thor Martin Bærug
> Co-founder, Walkie Talkie
> walkietalk.ie
>
>
> On Wed, Dec 17, 2025 at 9:51 PM Sullivan, Steven <sulliv55 at miamioh.edu>
> wrote:
>
>> We've had teachers use AI programs that supposedly analyze our website to
>> generate grade-appropriate worksheets for the students to fill out while
>> visiting the galleries.  Of course 99% of the information in our galleries
>> is not on our website.  Maybe we would want to change this in the future as
>> the web and its use evolves but, presently, the bot/app has no idea what to
>> productively compose, based on our site.  This does not stop a LLM from
>> generating 10-20 questions, though.
>>
>> At best, the AI questions are answerable but don't relate to anything in
>> our institution (eg, questions about dinosaurs when we have no fossils more
>> recent than the Silurian, or magnetism when the closest we get is a
>> magnetic wall that holds pictures in place to allow visitors to tell a
>> story about the prairie).  Many of the questions are contextually
>> pointless: "Why does the museum use magnets to teach?"  Often, though, the
>> sentences are arguably grammatically correct but conceptually convoluted:
>> "How can taxidermists address resource management and climate change in
>> biomes?
>>
>> Similarly, much of the information visitors get from AI is simply wrong.
>> It seems the longer the conversation, the more incorrect or misdirected the
>> results get, the more they may support misapprehensions (like snakes will
>> chase you, bear encounters are deadly, and outdoor domesticated cats can
>> integrate with natural ecosystems.)  Granted, Wikipedia used to have a
>> mistake in nearly every natural history article and it's now often pretty
>> useful.  I expect AI to similarly improve.  That said, I won't ever publish
>> or reference a Wikipedia article as I make exhibits, nor should any AI
>> generated labels ever be used.
>>
>> I have found AI to be useful in brainstorming categories, rendering draft
>> layouts, finding grey literature or other precedent work that can be
>> inspirational to me but is hard to find online (like policy and procedure
>> papers) and otherwise automating draft tasks.  It can also interpret
>> cursive handwriting well enough to help me get unstuck, and as a
>> translation aid, but the transcriptions of both are still full of
>> mistakes.
>>
>> Like a calculator, AI might have value in the creation process.  But, the
>> way a calculator is used, it is not necessary to cite it and the operator
>> needs to understand the operations they are automating with the tool.  If
>> the use of AI goes beyond that of an efficiency tool to that of a creative
>> tool, it must be cited, but it also probably shouldn't be used.   The
>> abhorrent AI exhibit at the San Francisco airport is an example that puts
>> our whole industry at risk.  Most of our visitors find us among the most
>> trustworthy of institutions.  Many of our visitors recognize that
>> publically accessible AI information is mostly trash (and with that
>> exhibit, find the deficiencies in "professional" AI, too).  They also value
>> the human hand in our creative products.  At the very least, to maintain
>> the public trust, it's important for us to minimize AI content in any
>> published work.
>>
>> The number one question we get is "Is this real?"  We should be able to
>> answer "Yes" for both objects and their labels and data.  AI can't ever do
>> that.
>>
>> I wrote all of the above before noticing your disclaimer line.  Maybe
>> translation is useful.  I used to be a translator, though, and have found
>> weakness in all automation I've ever tried, especially when nuance and
>> specialization are required.  That said, you used our industry listserv to
>> spam us, since there is no "reply all" button; seeing that I was only
>> replying to you made me look closer at the sub-signature bits.  This makes
>> it seem like you are not interested in actual public discussion, just
>> promotion.  This seems to align with the perception that AI, its users, and
>> creators, are largely bad actors in a variety of ways.  I have not included
>> the listserv address because maybe you just don't know how to work the
>> technology and didn't intend to spam us and will rectify the issue and
>> engage in actual discussion.  It's an important topic that is proceeding
>> faster than policy or awareness and so far, at least in my institution,
>> causes more problems than solutions.
>>
>> --Steve
>>
>> On Wed, Dec 17, 2025 at 1:00 PM Thor Martin Jensen <thormj at gmail.com>
>> wrote:
>>
>>> Hi all,
>>>
>>> A growing number of startups are building museum solutions built
>>> entirely on AI-generated interpretation. Visitors ask questions, and a
>>> language model generates personalized explanations on the fly.
>>>
>>> This fundamentally changes what a museum is. Instead of encountering
>>> institutional expertise, visitors receive algorithmic predictions optimized
>>> for engagement. The museum’s voice, built on research and accountability,
>>> gets replaced by pattern matching across training data that the institution
>>> never reviewed.
>>>
>>> These systems cannot cite their sources because there are no sources.
>>> They generate probable-sounding answers, not verified ones. When they get
>>> something wrong, no one is accountable because no person actually made the
>>> claim.
>>>
>>> A few questions I keep returning to:
>>> Where does production end and interpretation begin? Translation and
>>> transcription clearly help museums reach more people. But generating
>>> explanations of objects based on visitor questions?
>>>
>>> Who is responsible when AI interpretation misleads visitors? The museum?
>>> The vendor? The curator who approved the tool?
>>> Should museums disclose when interpretation is AI-generated? If we hide
>>> it, we break trust. If we reveal it, do visitors trust it less?
>>>
>>> What happens to institutional authority when knowledge becomes
>>> untraceable to human expertise?
>>>
>>> I wrote more here:
>>> https://www.linkedin.com/posts/walkietalk-ie_museums-curation-audioguides-activity-7407049070478852096-KUfB
>>>
>>> How are others thinking about this?
>>>
>>> Best,
>>> Thor Martin Bærug
>>> Co-founder, Walkie Talkie
>>> walkietalk.ie
>>>
>>> *Full disclosure: I run Walkie Talkie, a multilingual audio guide
>>> platform for museums. We use AI for translation and text-to-speech but keep
>>> all interpretation with museum staff. I have a commercial interest in how
>>> this question gets answered, but I genuinely want to hear how the field is
>>> thinking about these boundaries.*
>>> _______________________________________________
>>> Nhcoll-l mailing list
>>> Nhcoll-l at mailman.yale.edu
>>> https://mailman.yale.edu/mailman/listinfo/nhcoll-l
>>>
>>> _______________________________________________
>>> NHCOLL-L is brought to you by the Society for the Preservation of
>>> Natural History Collections (SPNHC), an international society whose
>>> mission is to improve the preservation, conservation and management of
>>> natural history collections to ensure their continuing value to
>>> society. See http://www.spnhc.org for membership information.
>>> Advertising on NH-COLL-L is inappropriate.
>>>
>>
>>
>> --
>>
>> <http://miamioh.edu/>
>>
>> Steven M. Sullivan
>>
>> Director | Hefner Museum of Natural History
>> <http://miamioh.edu/cas/academics/centers/hefner-museum/>
>>
>> Miami University
>>
>> 100 Upham Hall
>>
>> 100 Bishop Circle
>>
>> Oxford, OH 45056
>>
>> Museum: 513-529-4617
>>
>> Cell: 708-937-6253
>>
>>
>> The Museum is open to the public weekdays 9-4, free admission.
>>
>> Support the Museum today!
>> <https://securelb.imodules.com/s/916/lg21/form.aspx?sid=916&gid=1&pgid=6010&cid=11236&dids=292&bledit=1>
>>
>>
>> *Connecting you to the nature in your neighborhood...and the world.*
>>
>>
>> Miami's many museums and collections
>> <https://www.miamioh.edu/cca/museums-miami/index.html> provide unique,
>> cross-disciplinary opportunities for students, educators, and the public.
>>
>>
>> The Hefner Museum recognizes the Myaamia and Shawnee people who, along
>> with other indigenous groups, were the first stewards of this land's
>> biodiversity. <http://miamioh.edu/diversity-inclusion/land/index.html>
>>
>>
>>
>>

-- 

<http://miamioh.edu/>

Steven M. Sullivan

Director | Hefner Museum of Natural History
<http://miamioh.edu/cas/academics/centers/hefner-museum/>

Miami University

100 Upham Hall

100 Bishop Circle

Oxford, OH 45056

Museum: 513-529-4617

Cell: 708-937-6253

The Museum is open to the public weekdays 9-4, free admission.

Support the Museum today!
<https://securelb.imodules.com/s/916/lg21/form.aspx?sid=916&gid=1&pgid=6010&cid=11236&dids=292&bledit=1>

*Connecting you to the nature in your neighborhood...and the world.*

Miami's many museums and collections
<https://www.miamioh.edu/cca/museums-miami/index.html> provide unique,
cross-disciplinary opportunities for students, educators, and the public.

The Hefner Museum recognizes the Myaamia and Shawnee people who, along with
other indigenous groups, were the first stewards of this land's
biodiversity. <http://miamioh.edu/diversity-inclusion/land/index.html>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.yale.edu/pipermail/nhcoll-l/attachments/20251218/94c47118/attachment.html>