From Publishing Perspectives:
A potentially pivotal moment occurs this week in the closely watched development of the European Union’s “AI Act.”
Markets in many parts of the world, not just in Europe, are following along for clues and cues in terms of how artificial intelligence can be developed and applied “safely”—and even that term safely can be hotly debated, of course.
On Wednesday (December 6), the AI Act is to have its fifth “trilogue.” That’s the term for a negotiating session in which the European Parliament, the European Commission, and the Council of the European Union. Previous trilogue meetings on the Artificial Intelligence Act were held in June, July, September, and October. Originally, the idea was that this December trilogue would finalize the bill for the bloc this year, but there’s increasing concern that the timing of such progress will be take longer. This, on legislation that saw its first draft in 2021 and was first proposed in 2019.
What has happened in the interim—you won’t be surprised to read—is the rise of “foundation models.” Sometimes called “general purpose,” these are the systems designed as large-language models built for “deep learning” that can be adapted for a wide array of scenarios. This contrasts, of course, with the concept of a traditional program designed to handle a specific and narrow task set, maybe speeding up a bit of office drudge work. Such less ambitious programs require nothing like some foundation models’ contentious free-range feeding on information—often copyrighted content—to build their algorithmic-response structures.
A foundation model is a form of what’s called “generative artificial intelligence,” meaning that it can generate output from a broad base of ingested data.
At the highest intentional level, the over-arching core of discussion around this legislation has been, to quote the EU’s material, to handle “concerns especially with regard to safety, security, and fundamental rights protection.” But if the devil is usually in the details, a construct of digital details presents such a major chance for devilry that many observers now are worried about this important legislation’s progress.
Needless to say, the upheaval around OpenAI last month when its board fired and the rehired Sam Altman seemed to confirm fears that a major corporate player in the AI space could be thrown into turmoil by inscrutable internal governance issues. As Kevin Chan at the Associated Press is writing today, once the Altman fiasco had played out, European Commissioner Thierry Breton said at an AI conference, “‘At least things are now clear’ that companies like OpenAI defend their businesses and not the public interest.”
And yet, much discussed in coverage on the run-up to Wednesday’s trilogue is an unexpected resistance that’s been mounted by France, Spain, and Italy, which presented a pitch for self-regulation among AI players.
At The Guardian, John Naughton wrote up this “Franco-German-Italian volte face,” as he calls it, as the result of everyone’s worst fears: “the power of the corporate lobbying that has been brought to bear on everyone in Brussels and European capitals generally.” More broadly, the assumption is that in each EU member-state seeming to make that about-face and start talking of self-regulation as the way to go, something has been promised by industry advocates for the local national AI companies, a divide-and-conquer effort by lobbyists.
Link to the rest at Publishing Perspectives
PG notes that the reaction of the European publishers sounds a lot like that of American publishers.
As far as regulation is concerned, the current AI programs/services he has tried have their AI capabilities online, so geographical fences like the the European Union’s “AI Act” are unlikely to prevent individuals or organizations who wish to use AI services offered over the internet from a provider located anywhere in the world.
As one cyberlaw website put it, “Some believe that the internet should be operated as if it were a land all its own, independent of national policy.”