In his essay “Taming the Digital Leviathan: Automated Decision-Making and International Human Rights,” Malcolm Langford of the University of Oslo warns against the mystification of artificial intelligence. This mystification begins when we treat AI as something exceptional, alien, or definitively non-human. “Discussions of automation and digitalization should be guided by a logic of minimizing danger, regardless of whether its origin is machine or human,” Langford writes. After all, humans also have “black box” computations we don’t understand. As many philosophers of technology have pointed out, tech is an extension of our own bodies. It is more of us.
I have been meditating on those points — the drive to demystify AI and the desirability of seeing it as an extension of our own selves — as I’ve been reading about the very latest developments in machine learning, AI-level prediction, and the whole new social universe enabled by nuanced computation. It can be great fun and ineffably profound at the same time, as Google’s new “blob opera” feature — four singing blobs who can harmonize and bend and play their voices like human singers — shows.
The project began with recording four opera singers. Google’s AI team then “trained a machine learning model on those voice recordings.” In other words, the algorithm “thinks” that opera sounds like what the machine learned by listening, and the sounds that human listeners hear are the result of the learning, not of pre-programmed sounds to be “synthesized” by a synthesizer. The result is noises that “manage to approximate the gist of a true opera, in spirit if not lyrics.” (The blobs can only make vowel sounds, though there is beautifully subtle variation to each of the vowels).
“Blob opera” is a fun feature: download it and four adorable blobs appear in red (soprano), green (mezzo soprano), turquoise (tenor) and purple (bass). Their eyes follow your cursor around, engaging with your on screen movement. The “room” is set with a dusty-sounding large-room reverb, which sustains the notes for a few fading seconds after the singer ceases. Stretching the blobs higher causes them to sing higher on the scale, and if you hold the notes for a long time, the blobs won’t run out of breath, but their voices will vary a bit, simulating the vibrato-by-necessity of real singers holding sustained notes. The important thing to remember here is that these are not synthesized recordings of singers; rather, the application is Google making its machine sing after it learned how to by “studying” opera. The control sensitivity is incredibly impressive for a Google interface activity. You can create unusual harmonies and discord and make the blobs sing in minutely short staccato notes or long sustained forrays across vowel sounds. It’s better than any voice synthesizer function on music apps.
How does machine learning work? It’s somewhere north of metaphor and south of literalism, but the theories of neuron interaction set out by Donald Hebb in 1949 in his book The Organization of Behavior give us a clue. “When one cell repeatedly assists in firing another,” Hebb wrote, “the axon of the first cell develops synaptic knobs (or enlarges them if they already exist) in contact with the soma of the second cell.” The relationship of the neurons, natural or artificial, influences each. Activate the neurons at the same time and they are all “stronger.” Activate them separately and they are weak. It’s their relationship that matters, and that’s how learning occurs — relationally on the inside (neurons influencing each other to build a whole greater than the sum of their parts) and the outside (processing external data to learn from).
As profound and healing as music is, other examples of AI have even greater healing and harm reduction potential. Although the impact of AI on cardiology is well-known, late in 2020 another amazing experiment emerged, this time in mental health. Wired reports that “On December 3, a group of researchers reported that they had managed to predict psychiatric diagnoses with Facebook data — using messages sent up to 18 months before a user received an official diagnosis.”
Researchers used voluntary subjects’ past Facebook messages as predictive signals or flags. An AI program created the categories, and within those categories, certain message types indicated certain mood disorders (eg, bipolar or depression). The researchers’ interpretation of the flags “predicted” (that is, guessed the eventual diagnosis without knowing it) the conditions subjects actually had with accuracy “substantially better than would have been expected by chance.”
Although it raises some questions about fairness and free expression, this type of personality prediction technology can also help social media sites like Twitter or Facebook predict which users will post and/or share disinformation. This was confirmed with the development of an AI system at the University of Sheffield, a system “that detects which social media users spread disinformation — before they actually share it.” Although some of the predictive tools seem intuitive (“Twitter users who share content from unreliable sources mostly tweet about politics or religion”) the conclusions were made after AI analyzed more than “1 million tweets from around 6,200 Twitter users,” an amount of data that would have stretched human limits. AI was also “trained” to “forecast” users’ propensity to spread disinformation.
Of course, sometimes this all works in ways that still seem pretty clunky, and the results can be amusing. One researcher recently used a GPT-3 “trained” in late 2019 to analyze news stories from 2020 (including the really bizarre ones like murder hornets and monoliths found in national parks) and come up with predictions of its own for future news stories. The results included “Proof that a hellhound is living at Los Angeles Airport has been provided in the photos below … First naked bogman has been found out walking the great British countryside… Albino green sea monster filmed… at the wrong time… Scientists discover the alien ant farm under the Antarctic ice” and so on. My favorite, since I’m a sci-fi fan: “Lizardman: The Terrifying Tale of the Lizard Man of Scape Ore Swamp.”
These are nerdily hilarious. “I really can’t tell if these are attempts to do novel but realistic headlines, or to completely goof around,” the researcher wrote. AI goofing around? Imagine that. But, fun is itself relational and AI does seem to have a sense of humor from time to time. Interestingly, humor can only really be learned through experience and contemplation.
Though the “learning” inherent in all this feels novel, the ability for tech to process and organize this amount of data and information is nothing new. Computers have long been used to solve equations and calculate numbers that are too expansive or time-consuming for humans to do on their own. Other existing tech helps us organize data. Take data appending services like those offered by Accurate Append, for example. These services are able to organize data and fill in gaps that are very difficult for humans to do on their own. In many ways, AI is the natural evolution of existing technology, and it allows us to maximize our own abilities. As I said earlier, it is simply more of us.