Stephen Downes

Readers will recognize this theme, and this page from Journal of Open, Distance, and Digital Education leads with my article, On ethical AI principles, and then follows up with criticisms from Jon Dron, Stella George, and another from Dagmar Monett. My target is work by Luciano Floridi (such as) suggesting there is a global consensus on ethical principles, and also the many policy statements, guidelines, and even laws, entrenching them into practice. A closer look reveals no such principles are anything like universal, and some (like explainability and accountability) aren't even ethical principles at all. The Dron and George response is (on my reading) essentially the assertion that, yes, principles aren;t universal; "principles are foundational guidelines, starting points, and orientations that are used to frame understanding and assist with decisions." Monett criticizes my characterization of AI: "when defining AI, introducing a new definition and not considering at least one of the many that already exist...  is a questionable omission.... "reviewing, summarizing, translating and composing" are overstatements of the capabilities of AI algorithms." Image: Cogent Infotech.

This is a teriffic article and I can't wait for the next few in the series (I found it via Data Science Weekly but I've now subscribed). It's a retrospective from the first day of the week-long Cosyne 2026 conference and breaks down into three main parts: a short discussion of open databases, introduction to a 4 hour (!) session on ways to compare neural signals, and a long discussion of the keynote from Chris Olah, co-founder of Anthropic. There are many gems in here, some things that reinforced my previously held views, and others than challenged them. One thing that matters to me: "we find that networks tend toward distributed representations and mechanisms, which make understanding both artificial and biological networks a pain, equally... , the most natural computational unit of the neural network – the neuron itself – turns out not to be a natural unit for human understanding. This is because many neurons are polysemantic: they respond to mixtures of seemingly unrelated inputs."

I think there is something like what the authors call 'the third mission' for universities, but I feel it is badly mischaracterized here (not that the authors are at fault; they are following a long tradition). Here is how they describe it: "translating the knowledge they generate into socioeconomic benefits." And here is what they consider to be the major problem: "Given that university research operates far upstream from practical implementation, connecting the lab to the market is no small task." The characterization is that universities are the source, the community is the recipient, and that it's an entirely one-sided relation. But of course that's not true. There are numerous actual and potential points of contact between the university and the community, ranging from the students that enrol, the priorities the community expresses, the data the community produces, the culture all of this inhabits. If we think of universities as producing nothing more than 'research outputs' that need to be 'translated' (or maybe 'mobilized') into benefits, we are seriously misguided.

In a nutshell, " This paper (32 page PDF) addresses the question of how artificial intelligence should be designed and used to support learning rather than merely improve immediate outputs." The authors draw a sharp distinction between AI for work and AI for learning (illustrated) which I think may overstate the case. Still, the point stands. From there, they develop a three-part framework, based on (a) pedagogical foundation "to determine how generative systems can provide the precise support necessary," (b) an adaptive foundation, "organising adaptivity into a continuous four-stage cycle:" capture, model, adapt and evove, and (c) a responsible design foundation that "addresses how AI companions can act with integrity and uphold human values," based on security, transparency, accountability and inclusion. Via Philippa Hardman.

Ian O'Byrne summarizes some recent stories, including Googles AI-search announcement and the Monet painting panned on Twitter and draws the conclusion that many others have drawn, that we are losing our connection to 'the source layer'. "For a long time, information came with visible signals attached. A citation, a publisher, a byline, or a recognizable human voice with its own perspective and flaws," he writes. "AI systems change that. They collapse the distance between asking a question and getting an answer." Well, maybe. But claims that the source layers is dead are wildly exaggerated. A record of who said what, and how they knew, is still important, even to AI. If AI does anything, it reinforces the need for an empirical basis that underlies our knowledge. You can see the links to sources in his own article. It wouldn't have been worth reading without them.

This is a decent article summarizing recent research from a variety of sources and discussing how AI will impact a teacher's day-to-day life. The major initial impact seems to be focused around assessment, with AI being used to create rubrics and other tools, as well as playing a role in scoring. Time savings here can be considerable. But the the major potential impact is described under the heading of 'teacher survival' as some report the impact of 80 hour workweeks. AI plays a role not necessarily in reducing this workload, but in changing its nature, allowing teachers to shed paperwork and bureaucratic tasks and to focus on the actual work.