AI in the future of education

The speculations run high on what will be the next global business opportunity. As reported by the World Economic Forum, the futurist Thomas Frey is arguing that the next sector to be disrupted by technological innovation is going to be education. Education is undeniably a massive societal enterprise and whoever can sell a believable product that could be plugged into this activity would find themselves in a favorable financial position.

I’ve been predicting that by 2030 the largest company on the internet is going to be an education-based company that we haven’t heard of yet (Thomas Frey, DaVinci Institute)

The arguments for his predictions come from a belief in a combination of artificial intelligence and personalized learning approaches. While it will surprise no one that AI is currently much hyped, the idea of tailoring instruction to each student’s individual needs, skills, and interests is also gaining considerable traction at the moment.

A report on the effects of personalized learning carried out by the Rand Corporation is pointing to favorable outcomes. The reported findings indicate that compared to their peers, students in schools using personalized learning practices were making greater progress over the course of the two school years studied.

Although results varied considerably from school to school, a majority of the schools had statistically significant positive results. Moreover, after two years in a personalized learning school, students had average mathematics and reading test scores above or near national averages, after having started below national averages. (p. 8)

The authors are however careful to point out that the observed effects were not purely additive over time.

[For] example, the two-year effects were not as large as double the one-year effects (p. 10)

Since they also had problems in separating actual school effects from personalized learning effects they urge readers to be careful in extrapolating from the results.

While our results do seem robust to our various sensitivity analyses, we urge caution regarding interpreting these results as causal. (p. 34)

So how are these kinds of results interpreted by futurists and business leaders? By conjecture and hyperbole, is the short answer. Thomas Frey  speculates that these new effective forms of education will enable students to learn at four to ten times the ordinary speed. Others, like Joel Hellermark, promise even greater gains. Hellermark is CEO of Sana Labs, a company specializing in adaptive learning systems.

Adaptive learning systems use computer algorithms to orchestrate the interaction with students so as to continuously provide customized resources. The core idea is to adapt the presentation of educational materials by varying pace and content, according to the specific needs of the individual learner. And those needs are assessed by the system and are based on the students’ previous and current performance. Since Sana Labs use artificial intelligence to run this process, they are a good candidate for the kind of company Frey thinks might grow considerably in the near future. They are currently attracting funding from big investors, but in order to gain the interest of the likes of Mark Zuckerberg and Tim Cook you must be selling a compelling idea. In this regard, the Rand report provides a fertile ground with Hellermark interpreting the results thus:

If you understand how people learn, you can also personalize content to them. And the impact of personalised education is extraordinary. In one study by the Bill and Melinda Gates foundation [ie., the Rand report] students using personalised technique had 50% better learning outcomes in a year. Meaning that over a 12-year period due to compounding these students would learn a hundred times more.  Joel Hellermark

However, the size of the gap between interpretations like this, made by business leaders aiming to sell adaptive learning systems to schools and those of the Rand report’s authors themselves, is impressive.

Something also pointed out in the Rand report is the importance of heterogenous classes for learning. This becomes something of a complication for the personalized learning approach espoused by many purveyors of adaptive learning systems that focus entirely on the student as individual learner. But, why should heterogeneity and sociality add to an individual’s learning? A partial answer to this question might lie in the notion of social responsivity (as specifically found in the works of Johan Asplund and in the sociological tradition of Ethnomethodology more generally).

Social responsivity builds on the simple idea that humans are fundamentally social beings who are interconnected by means of communication and other forms social actions. This connectedness is built up as chains of interactions through responses to earlier actions. This is such a fundamental aspect of who we are that if you were to strip a person of their ability to belong in this world—to see and to be seen by others—it can be used as a form of punishment with solitary confinement perhaps the clearest of examples.

Following this line, there is but a small step from the mechanics of adaptive learning systems that function by tracking all the activity of individual learners to the ideas of the English philosopher Jeremy Bentham and his original formulation of the Panopticon, or what he also called the Inspection-House. Way down the list, after institutions such as penitentiaries, factories, mad-houses and hospitals, Bentham ventured to apply his principles of an architecture for inspection to schools also. In his plan, students should work as solitaries under the watchful eye of the master. Here, he wrote:

 All play, all chattering – in short, all distraction of every kind, is effectually banished by the central and covered situation of the master, seconded by partitions or screens between the scholars (Bentham, 1787)

Even though he advocated that the idea should be tested for education, he expressed severe qualms about the power one would hand over to whoever would govern such an institution.

Doubts would be started — Whether it would be advisable to apply such constant and unremitting pressure to the tender mind, and to give such herculean and ineludible strength to the gripe of power? (Bentham, 1787)

The central issue and challenge that I would like to raise here is that modern personalized learning approaches typified in adaptive learning systems build, in parallel with the views of Bentham, on the general understanding that traditional whole-class education is defective. And while personalized education picks on traditional teaching for being a blunt instrument, it remains blissfully ignorant of the entire dimension of social responsivity. Therefore, even with the best of intentions, it works to create learning environments built around a-sociality.

So, what could be done then? We don’t simply want to repeat the past. From Thorndike, via Pressey and Skinner, to the systems we see today, the idea that machines might provide efficient ways of adapting instructions to the needs of students has been aired and tested on and off for over a century. If we now wish to use AI to make a new attempt to better adapt instructions to students we also need to devise additional measures of student progress. What we need are concepts and measures that can incorporate and operate on the level of the group. Because we must acknowledge that human interaction and the social responsivity it supports is a vastly important aspect of how we live and learn.

Like the AI driven adaptive learning systems that are emerging today, the systems I would like to see should make suggestions for what learners should do next, but not only for individual learners. Instead, they should also be able to support teachers in what class activities to offer up next or which students should work together. If we can develop adaptive learning systems that support rather than obstruct social responsivity, then I think they can begin to have a real impact.

Smart or Dumb?

AI and the transformation of knowing

The development of human knowledge is very much a tale of tools. Tools as extensions of the human mind has been transforming human practices for millennia. Now the digital transformation has sparked a revolution where we still might see most of the change to happen in decades ahead. At this juncture we can identify an interesting shift in one of the many dimensions constituting the relation between humans and technology.

Physical and intellectual instruments (even many of the digital ones) functioning as mediational means in the service of human activities have, traditionally, established a sort of stable relationship between the user and her task (at an ontogenetical level). This stability or predictability stems from a basic form of ignorance held by the artificial.

For example, the modern power drill enables me to accomplish many things that would be hard to do purely by manual labor. But I still have to learn how to best use the tool and to choose a suitable drill bit according to what materials I’m working with. In a similar fashion,  the spell checking happening in my word processor helps in getting the words right. However, so far, it has not acknowledged my changing skills in the langue nor has it taken into account for what purposes a specific text is being written. Such artifacts are generally not context dependent, they aren’t altering theirbehavior in response to their anticipation and analysis of whatIam doing.

This, in turn, necessitates mastery in their use—A combination that has proved most successful. The very idea that a competent user wielding a powerful technology has been key in the proliferation of the human species is a central underpinning of the socio-cultural-historical theory. We could summarize this picture by saying that:

In the old world, the tools, as servants, were blind to the needs of their masters.

Looking ahead, what happens when the technologies start to anticipate my actions and alter their operations based on such assumptions? We can introduce a though experiment to clarify this idea by departing from Gregory Bateson’s discussion of the blind man and his stick:

 [Consider] a blind man with a stick. Where does the blind man’s self begin? At the tip of the stick? At the handle of the stick? Or at some point halfway up the stick? These questions are nonsense, because the stick is a pathway along which differences are transmitted under transformation, so that to draw a delimiting line across this pathway is to cut off a part of the systemic circuit which determines the blind man’s locomotion.  (Steps to an Ecology of Mind, 1972)

In Bateson’s example the stick in question is simply a “dumb”-stick that does nothing but affords a pathway that carries along vibrations between the ground and the blind man. We could however envision a next version of such a stick. Perhaps a “smart”-stick would start to learn about its master’s preferences. Gradually it builds a model separating the tactile forms of feedback generated by hard surfaces from the soft forms provided by the roadside. It can also extrapolate the blind man’s clear preference for one type over the other.

But what if the stick itself could also alter its shape so as to translate the “soft” feedback into “hard” one? Then it could adapt the presentation of information so that it reflects its user’s preferences, and not simply transmitting whatever surfaces it encounters. In this simple example we can easily grasp that such a development would lead to disaster and that a stick of that ilk would be of no use.

But can we always be so sure of other implementations that adapt their presentation of information, or change the way they operate, according to whatever assumptions they make of what the user needs? Do we even know when this happens? And when implemented how should such technology-held assumptions be communicated?

AI and the medical expert of tomorrow

In my research I have addressed the consequences of the continuous shift and development of technologies in different work settings and what that means for the skills that we develop. For a number of years, our group have been heading interdisciplinary research initiatives in the medical area. This work encompasses radiologists, dentists, surgeons, radio-physicists and social scientists that jointly study the management of different technological advancements in medicine. We also design workplace-learning environments in which both experienced professionals and novices can develop and improve essential skills. The next step for us, is both to develop AI applications in these areas as well as scrutinising their adoption and their consequences for practice.

If we look at medicine and many other complex work settings, what we find today is an inherent dependence on various technological set-ups. Work proceeds, by necessity, through the incorporation and use of a vast array of technologies. Physical as well as digital. This has as a consequence, that when these tools and technologies change and develop, then the practitioners have to adapt and re-skill.

The general trend here is that tasks of lower complexity can be automated and taken over by technology. The more complex tasks however, have so far tended to require expert involvement. The more recent developments of AI for medicine can be seen as a continuation of a long trend. But it might also represent change on a different order of magnitude. In the near future we will most probably see systems extending and going beyond the current limits of possible performance. This will imply that the medical experts will take on new and even more advanced roles, as supervisors or developers of new forms of knowledge and inquiry.

Now, these are not new arguments. What I want to highlight here is an issue that I miss in the current discussion about how AI transforms work. When we promote the current workforce and let them take on more advanced tasks today, we do so given a pool of people who have undertaken a traditional training and who have become experts under certain conditions. And this is a long process. But these trajectories of becoming experts are themselves being shifted in these transformations. What will it mean to be knowledgeable or an expert radiologist in say 15-20 years from now? Surely something different from today. But how is an individual going to end up so knowledgeable about a professional domain when during training, a system can outperform her every move and diagnosis for years on end? How will we motivate people to keep training and to keep learning so that they will one day be able to contribute to the development of new knowledge?  

While I don’t think that this is an unsolvable problem, we need to start this discussion alongside whatever powerful systems we introduce into medical practice. Otherwise we might be getting a devil’s bargain, where we profit in the short term, whilst depleting the knowledge base in the long run.

Algorithmic Accountability

Side A: A Speculative Vignette

It’s early Monday morning and Andrea is still feeling unpleasantly chilled from the commute to work. The October wind had been tearing at the cable car as they traversed the river. At the busy changeover to the trams, she was caught off guard by a sudden gust of wind and rain which showered her horizontally from top to toe. Her woollen coat is now damp and she can sense the ripe smell of sheep as she takes a seat in the large Hospital lecture hall.

Fortsätt läsa ”Algorithmic Accountability”