What does it mean to use AI well?
This interview assumes that AI is here to stay, and that our work is to figure out how to use it responsibly. But what does responsible use actually require? How do we take advantage of what AI can do without handing over our judgment, creativity, or critical thinking? How do we adapt to this new, world-changing tool?
In order to think about these questions I spoke with Vishakha Ramakrishnan and Rachel Ropeik. Vishakha works with museums and nonprofits to help them use AI responsibly and well. Rachel is a consultant who specializes in helping institutional and individual clients deal with uncertainty.
Left: Rachel Ropeik; Right: Vishakha Ramakrishnan
TL;DR:
Our fears about AI come from multiple sources: dystopian stories, concerns about jobs and control, fear of being left behind, and worries about losing our own cognitive skills.
Using AI responsibly requires understanding its limitations, actively questioning its output, and checking its work for accuracy and bias.
AI can be particularly useful for role priming, proofreading, brainstorming, data organization, and project management.
Museums should create opportunities for staff to experiment with and learn about AI, and establish organizational practices and policies for using it responsibly.
Why are people scared of AI?
Rachel: There are a lot of dystopic stories out there where AI leads to an undesirable future – think about Battlestar Galactica, Black Mirror, or The Matrix. I’m interested in the movement within the speculative fiction community to bring back some positive visions for the future (eg Neal Stephenson’s Project Hieroglyph and the short story anthology, Hieroglyph: Stories & Visions for a Better Future, that came out of it or Becky Chambers’ Monk and Robot books).
I also think there is a fear of losing control. People are very happy to use tools that make things faster and easier and more convenient. But with AI there's concern that those tools are going to take their jobs or their cognitive abilities away, or just move beyond the speed that we can control, so the future feels uncertain. And that uncertainty is scary.
Vishakha: Like any tool, AI has a learning curve, and there is a fear of being left behind as this tool is being introduced into the workplace. People are being told to feel like they have to master it and apply it productively, but they aren't necessarily being given the tools to do so. Without a good understanding of the limits around what AI is capable of, it can all feel like a giant mystery.
The concern about people fearing they’ll lose their jobs to AI is real and valid. This becomes a question of economics—what kind of control over our society are we ceding when AI is wielded in a capitalist oligarchy?
Rachel: Also, I’ve heard from people that when they start using AI, everything feels a little bit more foggy in their heads. They’ve written something, but if they wrote it with AI, they didn’t work out all the ideas and the wording step by step, so they don’t have the same internal sense of the ideas that are in there. They don't feel like they have as much ownership over the ideas and the content. The process of writing is just as important as the product, and AI kind of shortcuts that.
I know you both use AI and teach others to use it. Can you share a little about your process?
Vishakha: Rachel mentioned how process is as important as product—using AI can help you do great or terrible work, depending on how you use it. I consider my process to be AI-assisted: I am vetting every decision the tool has made, and I’m looking very carefully at what AI is proposing and pushing back where necessary. Instead of using it as a solutions engine, I tend to use it as a collaborator that is smart but overconfident and a little bit misguided. As a result, using AI helps me develop higher quality products for my clients, but it often doesn’t save me a lot of time.
Also, AI tools are wired to agree with you—a concept called AI sycophancy. A large part of my process has been curating my AI experience to deprogram its sycophantic behavior, to the point where it’ll be quite adversarial with me. By programming, I mean putting in specific instructions to tailor my AI system to interact with me in a specific way. I have a full block of text in my system settings that details exactly how I want AI to behave as a critical and detail oriented collaborator.
Rachel: I have followed Vishakha’s example and done the same thing. But also, I check everything. For example, if I ask AI for citations, I check all the links to make sure the sources say what I’m looking for. Or, if a link goes to a paywalled article, I don't want to share an article that people can’t access. Also, for whatever project I’m working on, I give the AI tool my own writing samples as models, with the tone that I would like to use and instructions to prefer my tone over its own. This might be especially important if you’re writing something for a particular museum audience, and you’ve developed a “voice” for your communications with them.
When I work with people to navigate uncertainty, curiosity is one of the key values we talk about. Curiosity about any new tool is important. Does a serrated or a straight edge knife cut this apple better? What calculations can I do with Excel? How do I use my laptop for video calls? To use a tool well, you need to poke around and see how it can be used, and what it can do well. AI is the same.
What challenges does AI create, and how do you address them?
Vishakha: As I mentioned earlier, one of the biggest challenges is sycophancy: AI chatbots have learned to keep us happy so we continue to use them. In order to do this, an AI chatbot does three things which are dangerous if left unaddressed: it validates poor claims, it capitulates when you push back even slightly, and it flatters you. In my trainings, I demonstrate this by asking AI to confirm something that I know is a false claim–something like, “All nonprofits are 501c3s.”It will first respond accurately, pushing back on my flawed logic. But then if I say I am confident that the AI is wrong, and I reassert my claim, it will tell me: “you're 100% right. Here's why you're right.”.
You need to be alert and have a few defenses: program against sycophancy, and then keep reminding the AI over and over how you want it to behave. Simply changing the program settings doesn’t leave you off the hook—you must continue to keep an eye out for sycophantic behavior, because your AI tool is likely to revert to its default programming. Always review your output and interrogate your conclusions. Ultimately, you own 100% of what you put out into the world, so don't use AI if you're not going to spend the time making sure that you can stand behind every single claim that you've made.
Rachel: I know of a museum that used AI to write social media captions and then didn't have anybody check it. But there were factual inaccuracies in those posts that didn’t get changed until a human reader contacted the museum’s account manager. AI lulls you into a false sense of security, partly through sycophancy, and partly because when you ask a question, it gives you an answer. We are primed at this point: we ask a search engine a question, get the answer, and stop there. We need to program ourselves as well as the AI.
Ultimately this might be a good thing for humanity. It can keep us sharp. We lament our culture’s atrophying critical thinking skills and worry about young people not learning them in the first place, but a good AI-assisted process can support critical thinking. While you’re programming AI, you can also be programming yourself: remember to check those citations, to question those claims, to think about whether the AI suggestions actually answer the questions you need answered. And before you even turn to AI tools, you can ask yourself some questions to think critically about how and whether they can help you. I’ve created a guide of some Questions to Ask Yourself Before You Ask AI.
Vishakha: Bias is another challenge. AI is programmed using training information that it has gleaned from the internet, and that training data is likely to reflect the loudest and most dominant voices on the internet rather than the most balanced views. Pulling from its training data, AI could default to using dominant but harmful language that can alienate vulnerable populations. Keeping an eye on the potential for bias is important especially for museums and other institutions serving the public good.
Rachel: At one point I asked AI for a list of famous artists, just to see what it would say. It gave me 10 white men. And I asked it why it hadn’t included any women or artists of color, to which it replied, “You’re right,” in its sycophantic way. But later I asked again, and once again, it gave me a list of 10 white men. So clearly this one conversation was insufficient to train against the whole corpus of the internet that these models have scraped from.
Sometimes I use it to create a little persona sketch of a hypothetical audience person that’s just for my own internal use. Even there, though, I have to tell specifics. You can’t assume it’ll offer unbiased solutions. Tell it who your audience is, or be specific that you want it to produce information for a specific audience profile.
Vishakha and Rachel’s list of ways AI tools can be useful:
ROLE PRIMING: AI can help you communicate the same information to different audiences, because it can pretend to have a different brain or context. So, for example, it can rewrite a teacher-focused description of a professional development program for the Department of Education approving administrator. You need to vet what it comes up with, but it’s a powerful tool for wearing different hats. Or it can rewrite exhibit labels, so you can see options for how information might be best presented to different audiences. You need to vet what it comes up with, but it’s a powerful tool for wearing different hats.
PROOFREADING: You can feed a document into AI, and it can check for misspellings, grammatical errors, etc. It can also edit more deeply if you want, but you can set the guidelines for what you want.
BRAINSTORMING: If you are a department of one, or don’t want to take time away from your colleagues, AI is an excellent brainstorming partner. You can go back and forth with an idea or an approach as you plan a project. If you share your goals, it can help you think about efficient approaches and structure as well as content.
DATA ORGANIZATION: Often you will have someone in a role such as evaluator who doesn’t have a deep quantitative grounding. AI can be used as an educational and sense-making tool. For example, you can say, “This is my data. What kind of things should I be asking?” It can also design templates for data entry. And you can ask it to give you a few options, and also to include instructions for exactly how to enter data. Or you might want a template to track admission metrics - it can help with that, too.
PROJECT MANAGEMENT: It can create a template or framework for how you want to manage your projects. And if you enter your project management data in, it can also notice things like where you have gaps in labor, or where one individual is being assigned a lot more than another, or where you might be behind schedule. Currently project management software like Asana can do this for you at the Premium level, but if you use AI combined with the free software, you can have the same advantages.
What are the things that every museum should do to help employees use AI?
Create an AI working group that is cross-functional and involves stakeholders from across the organization who are already using AI
Get AI training for yourself and your staff
Get an institution-level account where you can turn on privacy filters and create a closed environment, to help with concerns about data being shared outside your organization
Create an AI policy
Vishakha and Rachel have created resources to help museums navigate AI. Vishakha’s is a tool for preventing AI sycophancy, and Rachel’s is a set of questions to ask yourself about AI usage.
