When Curation Becomes Creation

Algorithms, Microcontent, and the Vanishing Distinction between Platforms and Creators

Authors: Liu Leqi, Dylan Hadfield-Menell, and Zachary C. Lipton

To appear in Communications of the ACM (CACM) and available on arXiv.org.

Ever since social activity on the Internet began migrating from the wilds of the open web to the walled gardens erected by so-called platforms (think Myspace, Facebook, Twitter, YouTube, or TikTok), debates have raged about the responsibilities that these platforms ought to bear. And yet, despite intense scrutiny from the news media and grassroots movements of outraged users, platforms continue to operate, from a legal standpoint, on the friendliest terms. 

You might say that today’s platforms enjoy a “have your cake, eat it too, and here’s a side of ice cream” deal. They simultaneously benefit from: (1) broad discretion to organize (and censor) content however they choose; (2) powerful algorithms for curating a practically limitless supply of user-posted microcontent according to whatever ends they wish; and (3) absolution from almost any liability associated with that content.

Today’s platforms play an increasingly active role in shaping what people see, arguably creating derivative media products of their own.

This favorable regulatory environment results from the current legal framework, which distinguishes between intermediaries (e.g., platforms) and content providers. This distinction is ill-adapted to the modern social media landscape, where platforms deploy powerful data-driven algorithms (so-called AI) to play an increasingly active role in shaping what people see and where users supply disconnected bits of raw content (tweets, photos, etc.) as fodder. 

Continue reading “When Curation Becomes Creation”

OpenAI Trains Language Model, Mass Hysteria Ensues

On Thursday, OpenAI announced that they had trained a language model. They used a large training dataset and showed that the resulting model was useful for downstream tasks where training data is scarce. They announced the new model with a puffy press release, complete with this animation (below) featuring dancing text. They demonstrated that their model could produce realistic-looking text and warned that they would be keeping the dataset, code, and model weights private. The world promptly lost its mind.

For reference, language models assign probabilities to sequences of words. Typically, they express this probability via the chain rule as the product of probabilities of each word, conditioned on that word’s antecedents p(w_1,...,w_n) = p(w_1)\cdot p(w_2|w_1) \cdot p(w_n|w_1,...,w_{n-1}). Alternatively, one could train a language model backwards, predicting each previous word given its successors. After training a language model, one typically either 1) uses it to generate text by iteratively decoding from left to right, or 2) fine-tunes it to some downstream supervised learning task.

Training large neural network language models and subsequently applying them to downstream tasks has become an all-consuming pursuit that describes a devouring share of the research in contemporary natural language processing.

Continue reading “OpenAI Trains Language Model, Mass Hysteria Ensues”