Architecture of trust in the age of artificial intelligence

Seterah Fanadi, Ph.D. in science and technology communication, wrote in a note: Next to every object in a museum, there is usually a small label. In the London Science Museum, the text next to the objects is about 60 words; That is, less than a short paragraph. Imagine having to explain in this space what a hundred-year-old device was, what problem it solved, why it was important and why it should be interesting for someone who stands in front of it for just a few seconds.

And here there is a prohibition: in this museum, the generative artificial intelligence tools, i.e. the same chatbots that today generate text in a few seconds, are not allowed to write the public tags of the museum, i.e. these 60 words.

This ban seems strange. The British Science Museum Group itself uses artificial intelligence and has even created its own model. So where is the problem?

Dr. Glyn Morgan, director of collections and senior curator at the London Science Museum, says the hardest part is not writing, but deleting. Among dozens of facts related to an object, which one should be brought and which one should be left out? This decision requires "judgment", "empathy with the audience" and "historical understanding". In other words, the London Science Museum has not only obtained the permission to select information, but also the permission to remove information from artificial intelligence.

To answer this question, we must first see what goes on behind the closed doors of a large museum. We need to get out of the theaters and take a look at what goes on behind the scenes of a great museum.

A museum where most of its treasures are not seen

What we see in the halls of the museum is usually a small part of the museum collection. For example, in the group of British science museums, which has six branches in this country and receives about five million visitors annually, more than seven million objects are stored; From delicate medical specimens to historic locomotives. Most of these objects are in storage and many have not yet been properly cataloged or described.

One of these treasures is the photo archive of the "Daily Herald" newspaper; A collection containing more than three million printed photographs from 1911 to 1968, which arrived at the museum in the 1980s. When the archive entered the museum, the facilities at that time were limited and only a small part of the photos could be photographed.

But the main story is behind the photos. Take an old black and white photo; For example, a scene from a football game in the 1930s when a goalkeeper comes out of the goal and takes the ball from the attacker of the opposing team. Now turn the picture over. The back is full of writing: the file number recorded with a typewriter, the name of the photographer, sometimes a short description of what happened in the photo, a newspaper stamp, and in the corner of the paper, a number that someone once wrote in pencil. The ink of some of these writings has faded, some have overlapped, and some have different handwriting.

Every photo actually has two layers: the image on the photo, and sometimes a small historical document behind it that, until someone reads these notes, it is impossible to know who, where and when the photo was taken.

It is beyond the capacity of any human team to read all this tiny and old text for millions of photos. This is where artificial intelligence comes in: models that see both the image and read the text and can turn the text behind the photos into organized information. Such models are tested and trained on tens of thousands of images scanned from both sides.

In the case of the Daily Herald archives, the result has been impressive. For example, before the use of artificial intelligence, only about 200 photos of the photographer named George Woodbine were registered in the scanned sample. After the machine read the text on the back of the photos, the number of photos confidently attributed to him increased to more than 2,300; It means almost ten times. The information was always there, but no one had the chance to read it.

Who knows when the car makes a mistake?

But this story is not without mistakes. Part of the Daily Herald's archive was related to old British railway posters. These posters encouraged people to travel by train. The researchers asked the artificial intelligence to show on the map which posters were promoting the "beach".

Imagine a map of the UK with a dot marking the location of every poster advertising the 'beach'. You would expect the dots to line up along the coastline. But the result was something else: dense clusters of points in industrial cities far from the coast and a few points in the middle of the waters of the North Sea; Where it does not make sense to install a poster there.

why Because the car had mistaken the place where the poster was installed with the place the poster was talking about. Indeed, when the museum curators reviewed the data and separated the "advert origin" from the "travel destination," the map was made.

This small example illustrates a big point: AI is quick to find patterns, but it doesn't necessarily understand their historical meaning. A subtle error can look simple and believable if no one reviews the results.

The problem is not only that the car sometimes makes mistakes; The question is where the machine learned from. Business chatbots are trained with vast amounts of text from the Internet, and the Internet is full of errors, mislabeling, and bias. If such a model were to interfere with a museum's cataloging, a false claim could enter a collection's official record and remain there for years. Morgan calls this risk "data pollution"; Like when you mix clean water with polluted water.

For this reason, the British Science Museum Group has built an in-house model called Dalia. The model was developed using an open-source software, but trained only with the museum's own verified data: official catalogues, photo archives and curators' records. No text from the public web is imported. Dahlia helps museums find duplicate photos, link handwritten texts to simple searches, and categorize collections by subject; For example, put together all the photos related to shipbuilding or agricultural machinery.

The second reason is deeper. This collection of museums was formed in the Victorian era and at the height of the British Empire, and its collections carry the imperial views and assumptions of that era. Curators today spend a part of their work reviewing these narratives. A tool learned from the Internet may reproduce the same old defaults in a confident tone. Because the machine-generated text looks fluent and unmistakable, it is easier for the reader to believe it. In simpler terms, a machine can write "believable" text, but believability is not the same as being correct.

What do other museums in the world do?

In addition to the British Science Museum Group, London's Natural History Museum has used machine vision to digitize drawers full of insect specimens. The goal is to allow researchers to track changes in insect size over decades and link this old data to climate change research.

At Cambridge University's Zoological Museum, visitors can interact with a 'talking whale'. But this conversation is done only on the basis of a closed and selected set of verified information so that the visitor is not presented with new and unverified information.

The Louvre Museum in Paris, in collaboration with a private company, has created a "conversational guide for teenagers". Japan's National Museum of Emerging Science and Innovation (Miraikan) has placed interactive assistants directly into its permanent exhibit to explain complex scientific concepts to visitors. The German museum, Deutsches Museum in Munich, has also published a document titled "Artificial Intelligence Mission" that clarifies the rules of the museum's digital experiments and acknowledges that these activities are still evolving.

Thus, across a spectrum of approaches that spans from completely closed and controlled data to open commercial tools, there is a common question: Who is responsible for the accuracy of what the visitor hears?

There is another problem that is specifically related to science museums. For decades, London's Science Museum has housed physical models of molecules made of wood, wire, and clay; including insulin structure model. But today, one of the greatest achievements in biology, namely "alphafold", which predicts the structure of proteins and was awarded the Nobel Prize in Chemistry in 2024, has no physical body; It is a software.

How do you bring something that has no physical form to a museum? Dr. Morgan cites several ways: the manuscripts of an author who inspired scientists, the uniform of a team that participated in a protein prediction competition, the source code of previous software or the hardware on which the model was run. None of them are sufficient on their own, because each only shows a part of the story.

Another problem is time. Museums must preserve objects and documents for 150 years or more, but software can become outdated in less than ten years. The user interface crashes, files are silently corrupted, and the company that made the tool may shut down its product. Maintaining an artificial intelligence model means maintaining the entire environment in which the model was implemented.

If we distance ourselves from the details, a common pattern can be seen in all these examples. Museums have divided the work of artificial intelligence into two categories. The first category is “finding and organizing”: reading the captions behind the photos, finding duplicates, and categorizing. These tasks are left to the machine with more confidence; Of course, with human review.

The second category is "speaking in the name of the museum": the labels, narratives, and claims that the visitor trusts to the museum's credibility. This category needs much more attention. A machine can write fluent and even accurate text, but if the text is wrong, who is responsible and who can correct it?

It seems that the roots of trust have been transferred from the product to the process as well. Museums are still among the most prestigious scientific institutions in the world, but in the near future, everything may be about to change. Institutional credibility alone is not enough when fluent, expert-looking texts are produced everywhere. The institution must show who wrote this text, what sources its information came from, and who checked it. That is, it should use methods to maintain, check and display the validity of information. For this reason, a model such as Dahlia or the Cambridge Spokesman Whale is valid only to the extent that the source of its information is known.

Of course, this choice has a cost. Building a dedicated model requires data expertise, time and budget. A museum that does not have these facilities will likely turn to off-the-shelf commercial tools and take the same risk of data contamination. So the question is not only "should we use artificial intelligence?"; Another aspect of the issue is who can clearly and controllably define the boundaries of its use and who cannot. This gap will probably deepen between large and small institutions and between different countries.

It should be noted that being "written by humans" is not a guarantee of correctness by itself. Museums themselves know that human curators have also made mistakes, and imperialist, imperialist, or other biased narratives have remained on museum labels for decades. The real rule in this approach is not "man versus machine"; Rather, it is "the ability to answer and correct". A good text is a text that, if it was wrong, its origin can be identified and corrected. A famous example from the computer world says: "You can't bug a system unless it has logs."

This also applies to the scientific misinformation we see online. There, the problem is not only the existence of wrong information. Systems that work on the basis of attention and interaction, publish texts that are more "believable" and "fluent"; Even if they are not correct. The museum's 60 word rule is actually a small, slow and human model against this logic. Of course, let's be honest: the museum tag reaches thousands of people, while social media algorithms deal with millions. In fact, it may be said that the museum model is not the solution to the problem; It is an example that shows that trust should be designed obsessively.

If we follow this line of museum examples, we can reach a conclusion that goes beyond the museum. When we talk about artificial intelligence entering a field, our mind usually goes to the infrastructure and process: what tools to buy, what data to collect and how to organize the work. These are essential, but the above examples also illustrate another point: trust does not necessarily derive from the instrumentality and prior reputation of the institution. Trust must be designed.

That is, before technology comes into play, we must ask how the audience can understand where this text came from, who accepted the responsibility for it, and if it was wrong, who should be referred to and who is responsible for correcting it. These questions are not the final arrangement of the work; They are part of the main plan and are as important as infrastructure and process. This conscious design can be called "Trust Architecture".

From these museum examples, five components can be extracted for this architecture:

1. Clear division of work: it should be clear where the machine helps; For example, in finding, categorizing and offering, and where its use is prohibited; For example, speaking in the name of an institution.

2. Known source: what data did the tool learn from? Can these data and sources be checked?

3. Responsible person: For every public claim, a specific person or organization should be responsible.