And so Netflix has gone through several different algorithms over the years… They’re using Pragmatic Chaos now. Pragmatic Chaos is, like all of Netflix algorithms, trying to do the same thing. It’s trying to get a grasp on you, on the firmware inside the human skull, so that it can recommend what movie you might want to watch next — which is a very, very difficult problem. But the difficulty of the problem and the fact that we don’t really quite have it down, it doesn’t take away from the effects Pragmatic Chaos has. Pragmatic Chaos, like all Netflix algorithms, determines, in the end, 60 percent of what movies end up being rented. So one piece of code with one idea about you is responsible for 60 percent of those movies.
But what if you could rate those movies before they get made? Wouldn’t that be handy? Well, a few data scientists from the U.K. are in Hollywood, and they have “story algorithms” — a company called Epagogix. And you can run your script through there, and they can tell you, quantifiably, that that’s a 30 million dollar movie or a 200 million dollar movie. And the thing is, is that this isn’t Google. This isn’t information. These aren’t financial stats; this is culture. And what you see here, or what you don’t really see normally, is that these are the physics of culture. And if these algorithms, like the algorithms on Wall Street, just crashed one day and went awry, how would we know? What would it look like?
[Transcript of How algorithms shape our worldI]
When Pythagoras discovered that “things are numbers and numbers are things,” he forged a connection between the material world and mathematics. His insight “that there is something about the real world that is intelligible in mathematical terms, and perhaps only in mathematical terms,” was, according to Charles Van Doren, “one of the great advances in the history of human thought.” (p35) Are we at a similar precipice with culture and information, when algorithms shape our world and culture? When non-human actors can significantly impact upon the information we receive, and the choices we make? And if so, what does that mean for museums, for culture, for the way we understand our world?
This is a question I sometimes find myself grappling with, although I’m not sure I have any answers. The more I learn, the less it seems I know. But I’d like to take a couple of minutes to consider one aspect of the relationship between the algorithm and the museum, being the question of authority.
In 2009, Clay Shirky wrote a speculative post on the idea of algorithmic authority, in which he proposed that algorithms are increasingly treated as authoritative and, indeed, that the nature of authority itself is up for grabs. He writes:
Algorithmic authority is the decision to regard as authoritative an unmanaged process of extracting value from diverse, untrustworthy sources, without any human standing beside the result saying “Trust this because you trust me.” This model of authority differs from personal or institutional authority, and has, I think, three critical characteristics.
These characteristics are, firstly, that algorithmic authority “takes in material from multiple sources, which sources themselves are not universally vetted for their trustworthiness, and it combines those sources in a way that doesn’t rely on any human manager to sign off on the results before they are published”; that the algorithm “produces good results” which people consequently come to trust; and that, following these two processes, people learn that not only does the algorithm produce good results, the results are also trusted by others in their group. At that point, Shirky argues, the algorithm has transitioned to being authoritative.
Although I’ve previously touched on the idea of algorithmic curating, I’d never explicitly considered its relationship to authority and trust, so I decided to look a little deeper into these issues. Were there any commonalities between the type of authority and trust held by and in museums, and that held in algorithms?
Philosopher Judith Simon refers to Shirky’s post in an article considering trust and knowledge on the Web in relation to Wikipedia. She argues that people trust in Wikipedia’s openness and transparency, rather than in the individual authors. She writes “that the reason why people trust the content of Wikipedia is that they trust the processes of Wikipedia. It is a form of procedural trust, not a trust in persons.”
I think this procedural trust is also what we put in the algorithm. Blogger Adrian Chan puts it this way:
The algorithm generally may invoke the authority of data, information sourcing, math, and scientific technique. Those are claims on authority based in the faith we put in science (actually, math, and specifically, probabilities). That’s the authority of the algorithm — not of any one algorithmic suggestion in particular, but of the algorithmic operation in general.
We do not necessarily trust in the particularities; we trust the processes. Is the trust that people have in museums similarly procedural? Do we trust in the process of museum work, rather than in the individual results or in the people who work in museums?
There are a myriad of assumptions that we make about people working in museums; that they are well trained and professional; that they are experts in their particular domain. We implicitly trust the people, then, and the work that they do. However, in many cases, such as when we visit an exhibit, we don’t know who the specific people are who worked on the exhibition. We don’t necessarily know who the curator was, or who wrote the exhibition text. The lack of visibility inherent in many current museum processes obscures the individual and their work. The museum qua museum, therefore, acts as a mechanism for credibility because it purports to bring the best people together; because the people who work within are known to be trained professionals who use scientific methods, regardless of whether we know specifically who they are or what their particular training is. Ergo, the trust we have in the museum must also be a form of procedural trust. (Amy Whitaker concurs, “Institutional trust is founded on process, on the belief that there are proper channels and decision-making mechanisms and an absence of conflict of interest.” p32)
Shirky also speaks to the social element involved in authority. He explains:
Authority… performs a dual function; looking to authorities is a way of increasing the likelihood of being right, and of reducing the penalty for being wrong. An authoritative source isn’t just a source you trust; it’s a source you and other members of your reference group trust together. This is the non-lawyer’s version of “due diligence”; it’s impossible to be right all the time, but it’s much better to be wrong on good authority than otherwise, because if you’re wrong on good authority, it’s not your fault.
Authority isn’t just derived from whether we can trust a source of information, but additionally whether we can be confident in passing that information along and putting our name to the fact that we made a judgement on its trustworthiness. We shortcut the process of personal judgement using known systems that are likely to give us accurate and trustworthy results; results we can share in good faith. We trust museums because museums are perceived to be trustworthy.
Do the film companies that run their scripts through Epagogix’s algorithms do so because it helps them shortcut the process of personal judgement too? Can algorithms provide better insight, or just safer insight? Eli Pariser says this of Netflix’s algorithms:
The problem with [the algorithm] is that while it’s very good at predicting what movies you’ll like — generally it’s under one star off — it’s conservative. It would rather be right and show you a movie that you’ll rate a four, than show you a movie that has a 50% chance of being a five and a 50% chance of being a one. Human curators are often more likely to take these kinds of risks.
Right now, museums that do not embrace technology and technologically-driven solutions are often perceived to be risk averse, because doing so challenges existing practice. I wonder whether, with time, it will be those institutions that choose not to make choices driven by data that will become perceived as the risk-takers? This is a profession that is tied so strongly to notions of connoisseurship; what relationship will the museum have with the algorithm (internally, or external algorithms like those that drive Google and other sites)? I don’t have any answers yet, but I think it’s worth considering that museums no longer just share authority with the user-generated world; authority is also being shared with an algorithmically-shaped one.
What do you think?
6 thoughts on “In algorithms we trust.”
Hi Suzie, its an interesting curly and contemporary issue you raise. Perhaps more minefield than issue. I tried to think through the Netflix parallel for museums, and there is a direct one. Its quite conceivable, perhaps even desirable, that we have an algorithm that can predict, say, likely visitation to an exhibition on a given topic in a given city. Subject info combined with demographics and psychographics should make that reasonably achieveable. The “exhibition success predictor” app for museums. That’s just automating what we do now, and doesnt seem to relate to the authority of the museum per se.
Its gets more interesting if there was an algorithm that said how to pitch information/content in a way that was more likely to be believed.
Digressing momentarily, whenever I go to a museum I make a personal value judgement about the “authority” (believability, credibility, accuracy) of a museum, based on the society/culture/city its in, and the subject matter. Personally, I give less credibility to museums in some countries where the museum is clearly a propaganda vehicle. That doesnt necessarily make them less interesting, but maybe just less believable.
Now, wouldnt it be interesting if there was an algorithm “museum believability app” that you could use when you are visiting a museum in a city/country/context. Just use the gps location info to rate the building you are standing in front of, perhaps on the “grains of salt” index, or bovine fertiliser scale.
A tad more seriously, it would be a courageous museum director indeed who ignored the developments in algorithms that relate to museums.
Fascinating comment, Frank. The “exhibition success predictor” app sounds both useful and gruesome simultaneously; a paint-by-numbers exhibition-judge that ensures that only exhibitions with a great chance of success in their particular geographical location are put on. You’d really want to hope that the algorithm is correct in that case. And what of the role of human judgement? Lots of interesting questions there.
But the idea of finding a content-pitching “formula” is an intriguing concept. My intuitive response is that it cannot or (more accurately) should not be possible, but I know that is because I want to believe that there must be something “human” about great story-telling; and maybe that’s just not true. If you could come up with the perfect formula for a great exhibition, would you use it? And if so, what becomes the role of the curator?
A Nieman Labs discussion titled “There’s no such thing as an objective filter: Why designing algorithms that tell us the news is hard” offers some interesting thoughts on the different approaches and thought processes of technologists and humanists when it comes to news, and designing ‘good’ filtering algorithms. It concludes that “we don’t yet really know how to combine the pragmatic demands of technology with the social aspirations of the humanities.” This might be at the crux of the problem for museums, too?
I’ve found that the Netflix algorithm has gotten worse. I’m less satisfied with it now than I was when I first started using the service years ago. I think Eli Pariser’s statements about risk get to the heart of it. Heuristics like these are designed to minimize the chance of error, not maximize the chance of success. They’re risk-averse by design. I’m tired of mediocre movies that I can’t love because they’re not bold enough for someone to hate. I don’t have enough time to sit around and do nothing while watching a movie that isn’t worth five stars, even if it’s good enough to be worth four. And I’d rather take the chance of seeing a couple one-star movies in the middle of all of those five-star ones if it means not having to sit through one more three-star movie.
I imagine that as people become more accustomed to trusting these algorithms, they’ll simply come to the conclusion that movies suck nowadays. There just aren’t enough good ones being made anymore. It might never occur to them that the recommendation engine simply failed to find the real gems for them.
Machine-enabled curation has to be risky to be really successful. Netflix is probably not the model we should looking at. Most commercial ventures in this field should probably be best considered case-studies in how not to use algorithmic heuristics for our field. Businesses are risk-averse by nature. We need to look for risk-taking algorithms, the ones that we know sometimes fail but usually succeed and do so spectacularly. There’s not as much research into those as I wish there was. The history of AI research is actually pretty boring from that perspective.
I think perhaps better than the algorithms used to “curate” movies or music we should be looking at the algorithms used to model cultural behaviors in simulations. The kinds of things that end up as crowd modeling for emergency-preparedness simulations, military simulations and epidemiology research. When these algorithms fail (and they do) it’s obvious and often times produces ridiculous or hilarious results. But when they work, they’re the most convincing and useful simulations around.
Yeah, I have to say, I am not necessarily all that interested in the netflix algorithm per se (it just happened that two of the quotes that I felt provoked interesting perspectives both referred to it). But this idea that machine-enabled curation has to be risky to be successful is interesting, particularly given that one reason that I think people are drawn to formulaic solutions would be to minimise risk (right?).
We are moving into a different space from what I was originally thinking about, which was the realisation that people often trust the invisible processes behind the information they access (be it in museums, or Google). As you say, if people only see links for mediocre movies, they might think that movies have generally gone down in quality, which might not be true. Unless something erodes that trust, unless (like you) the algorithm does not seem to produce good results, then people will likely trust that what they are seeing, that the options they are being shown are the best ones, whether or not that’s true. But that’s also true in museums, right? If I go to an art gallery that purports to show me the best abstract art from 1960 but what they have is really only “B grade” art, and I don’t have enough personal information or knowledge to know better, then I am going to really believe that the abstraction from that time was terrible. So we trust the unseen processes, but sometimes the processes might not work, or might not work well.
But the question of how else museums might use algorithms within their practices (curatorial or otherwise) is interesting. What other roles/uses do you think they could have within the museum context?