Now they're at the center of a landmark legal case that ultimately has the power to completely change how we live online. On February 21, the Supreme Court will hear arguments in Gonzalez v. Google, which deals with allegations that Google violated the Anti-Terrorism Act when YouTube's recommendations promoted ISIS content. It's the first time the court will consider a legal provision called Section 230. Section 230 is the legal foundation that, for decades, all the big internet companies with any user generated stuff--Google, Facebook, Wikimedia, AOL, even Craigslist--built their policies and often businesses upon. As I wrote last week, it has "long protected social platforms from lawsuits over harmful user-generated content while giving them leeway to remove posts at their discretion."

Chatbots have been around since the 1960s and coders have been trying to pass the Turing test ever since, creating increasingly sophisticated iterations of natural language processing (NLP) software. A recent episode, where a Google engineer was sacked for claiming that the search engine's chatbot generator software known as LaMDA was sentient, perhaps demonstrates the leaps and bounds that NLP has made over the past few years. However, it's only with the public release of a new chatbot called ChatGPT that the potential of NLP has been taken seriously by the wider public. ChatGPT is a conversational piece of software released by OpenAI, designed to answer questions posed in natural language and even have a dialogue with users. It has been trained on a multitude of online data from Wikipedia to Reddit, although the information is only correct up until 2021. As well as answering general queries and therefore being a potential threat to Google, it also has the ability to write bespoke articles on any topic which is sparking off existential debates amongst academics and professional writers.

French DJ and music producer David Guetta has discovered AI tools, and thought it would be fun to use an unauthorized deepfake of Eminem's voice to rev up a huge crowd at a live show. It looks like it worked, but it raises legal and ethical questions. In a tweet last week, Guetta shows the live performance moment in question, then explains how he did it, presumably using something like ChatGPT to write the lyrics, and then another service like Uberduck or FakeYou to turn the lyrics into a soundbite. "There's something I made as a joke, and it worked so good I could not believe it! I discovered those websites that are about AI. Basically, you can write lyrics in the style of any artist you like. So I typed'write a verse in the style of Eminem about Future Rave,' and I went to another AI website that can recreate the voice. I put the text in that, and I played the record, and people went nuts."

Shocker: people aren't quite sure that they trust artificial intelligence to operate in their best interests, per a new poll. In a press release, the think tank MITRE released the results of a new poll, conducted in tandem with the marketing research firm Harris, that asked people their opinions about AI. Spoiler alert: they lowkey hate it! "Most Americans express reservations about AI for high-value applications such as autonomous vehicles, accessing government benefits, or healthcare," the press release reads. "Moreover, only 48 percent believe AI is safe and secure, and 78 percent are very or somewhat concerned that AI can be used for malicious intent."

With an increased focus on incorporating fairness in machine learning models, it becomes imperative not only to assess and mitigate bias at each stage of the machine learning pipeline but also to understand the downstream impacts of bias across stages. Here we consider a general, but realistic, scenario in which a predictive model is learned from (potentially biased) training data, and model predictions are assessed post-hoc for fairness by some auditing method. We provide a theoretical analysis of how a specific form of data bias, differential sampling bias, propagates from the data stage to the prediction stage. Unlike prior work, we evaluate the downstream impacts of data biases quantitatively rather than qualitatively and prove theoretical guarantees for detection. Under reasonable assumptions, we quantify how the amount of bias in the model predictions varies as a function of the amount of differential sampling bias in the data, and at what point this bias becomes provably detectable by the auditor. Through experiments on two criminal justice datasets -- the well-known COMPAS dataset and historical data from NYPD's stop and frisk policy -- we demonstrate that the theoretical results hold in practice even when our assumptions are relaxed.

Segmentation and Rhetorical Role Labeling of legal judgements play a crucial role in retrieval and adjacent tasks, including case summarization, semantic search, argument mining etc. Previous approaches have formulated this task either as independent classification or sequence labeling of sentences. In this work, we reformulate the task at span level as identifying spans of multiple consecutive sentences that share the same rhetorical role label to be assigned via classification. We employ semi-Markov Conditional Random Fields (CRF) to jointly learn span segmentation and span label assignment. We further explore three data augmentation strategies to mitigate the data scarcity in the specialized domain of law where individual documents tend to be very long and annotation cost is high. Our experiments demonstrate improvement of span-level prediction metrics with a semi-Markov CRF model over a CRF baseline. This benefit is contingent on the presence of multi sentence spans in the document.

Holzenberger et al. 2020 has modeled statutory Legal Judgment Prediction (LJP) has recently reasoning by classifying US tax law provisions gained considerable attention in the mainstream concatenated with textual case descriptions. We NLP community (e.g., Aletras et al. 2016; build on this prior work in two ways. First, we Chalkidis et al. 2019, 2021, 2022b; Santosh et al. develop and evaluate our model on a public dataset 2022, 2023). In LJP, the outcome of a case should (Chalkidis et al., 2022b) of cases by the European be classified/predicted based on a textual description Court of Human Rights (ECtHR), which hears complaints of case facts. In actual legal reasoning, legal by individuals about possible infringements practitioners (e.g., advocates, judges) determine relevant of their rights enshrined in the European Convention rules from the sources of law (e.g., statutes, on Human Rights (ECHR) by states. To the regulations, precedent) that are relevant to the case best of our knowledge, this is the first work applying at hand. They then carry out an analysis to determine article-aware case outcome prediction setting which rules apply to the case at hand, and to human rights adjudication.

People break up, miscarry, and lose loved ones. Their online streaming and shopping recommendations, however, do not necessarily update, and may serve as unhappy reminders of their loss. When users want to renege on their past actions, they expect the recommender platforms to erase selective data at the model level. Ideally, given any specified user history, the recommender can unwind or "forget", as if the record was not part of training. To that end, this paper focuses on simple but widely deployed bi-linear models for recommendations based on matrix completion. Without incurring the cost of re-training, and without degrading the model unnecessarily, we develop Unlearn-ALS by making a few key modifications to the fine-tuning procedure under Alternating Least Squares optimisation, thus applicable to any bi-linear models regardless of the training procedure. We show that Unlearn-ALS is consistent with retraining without \emph{any} model degradation and exhibits rapid convergence, making it suitable for a large class of existing recommenders.

We report on an experiment in case outcome classification on European Court of Human Rights cases where our model first learns to identify the convention articles allegedly violated by the state from case facts descriptions, and subsequently uses that information to classify whether the court finds a violation of those articles. We assess the dependency between these two tasks at the feature and outcome level. Furthermore, we leverage a hierarchical contrastive loss to pull together article-specific representations of cases at the higher level, leading to distinctive article clusters. The cases in each article cluster are further pulled closer based on their outcome, leading to sub-clusters of cases with similar outcomes. Our experiment results demonstrate that, given a static pre-trained encoder, our models produce a small but consistent improvement in classification performance over single-task and joint models without contrastive loss.

Unlearning algorithms aim to remove deleted data's influence from trained models at a cost lower than full retraining. However, prior guarantees of unlearning in literature are flawed and don't protect the privacy of deleted records. We show that when users delete their data as a function of published models, records in a database become interdependent. So, even retraining a fresh model after deletion of a record doesn't ensure its privacy. Secondly, unlearning algorithms that cache partial computations to speed up the processing can leak deleted information over a series of releases, violating the privacy of deleted records in the long run. To address these, we propose a sound deletion guarantee and show that the privacy of existing records is necessary for the privacy of deleted records. Under this notion, we propose an accurate, computationally efficient, and secure machine unlearning algorithm based on noisy gradient descent.