Who is accountable when the model is wrong.
10 topics · 15 curated works
No prior grounding assumed.
Losing Humanity: The Case Against Killer Robots
Human Rights Watch & Harvard IHRC · 2012
Argues fully autonomous weapons that select and engage targets without a human in the loop cannot reliably comply with the laws of war's distinction…
+1 more at this level
Assumes you know the vocabulary.
The Future of Employment: How Susceptible Are Jobs to Computerisation?
Frey & Osborne · 2013
Estimates the probability of computerisation for 702 US occupations from the engineering bottlenecks that resist automation, concluding about 47% of…
+6 more at this level
Primary sources and full treatments.
The Algorithmic Foundations of Differential Privacy
Dwork & Roth · 2014
Builds differential privacy up from first principles as a mathematically precise guarantee — bounding what an attacker can learn about any one…
+5 more at this level
12 of 15 works
Human Rights Watch & Harvard IHRC
Argues fully autonomous weapons that select and engage targets without a human in the loop cannot reliably comply with the laws of war's distinction and proportionality requirements, and calls for a preemptive international ban.
50 pageslink checked 17 Sept 2026Hannah Fry (The Royal Institution)
Argues that algorithms fail in systematic rather than random ways, so the useful question is not whether they are accurate but whether the people relying on them can tell when they are wrong.
link checked 17 Sept 2026Frey & Osborne
Estimates the probability of computerisation for 702 US occupations from the engineering bottlenecks that resist automation, concluding about 47% of US employment is in occupations at high risk over the following two decades.
72 pageslink checked 17 Sept 2026Ribeiro, Singh & Guestrin
Introduces LIME, which explains any classifier's individual predictions by fitting a simple, locally faithful model around the specific instance being explained, treating the underlying model as a black box.
10 pageslink checked 17 Sept 2026Joy Buolamwini & Timnit Gebru
Auditing commercial face classifiers by skin tone and gender together, rather than one at a time, exposes error rates an order of magnitude worse for darker-skinned women.
link checked 17 Sept 2026Eloundou, Manning, Mishkin & Rock
Rates occupational tasks by exposure to large language models directly and via LLM-powered tools, finding roughly 80% of the US workforce has at least 10% of their tasks affected, concentrated in higher-wage, higher-education occupations rather than the routine manual work earlier automation research expected.
35 pageslink checked 17 Sept 2026Irene Solaiman
Replaces the binary of 'open' versus 'closed' model release with a gradient of options — from fully closed to fully open — and lays out the tradeoffs and safety considerations at each point on that gradient.
13 pageslink checked 17 Sept 2026European Parliament and Council of the European Union
Establishes the first comprehensive AI-specific legal framework of its kind, sorting AI systems into risk tiers from banned practices down to minimal risk and attaching graduated obligations — the most stringent on general-purpose and high-risk systems — to each tier.
144 pageslink checked 17 Sept 2026Sastry et al.
Argues compute is a uniquely detectable, excludable and quantifiable input to AI development compared to data or algorithms, which makes it a practical lever for AI governance, and surveys the specific policy mechanisms that lever could support.
56 pageslink checked 17 Sept 2026Dwork & Roth
Builds differential privacy up from first principles as a mathematically precise guarantee — bounding what an attacker can learn about any one individual from a query's output — and derives the mechanisms and composition rules that make it usable in practice.
281 pageslink checked 17 Sept 2026Lundberg & Lee
Shows that several existing feature-attribution methods are special cases of one game-theoretic quantity, Shapley values, and unifies them into SHAP, with a provably unique solution satisfying local accuracy and consistency.
10 pageslink checked 17 Sept 2026Shokri, Stronati, Song & Shmatikov
Shows an attacker who can only query a trained model's outputs can often determine whether a specific record was in its training set at all, exposing a privacy leak that exists even when the training data itself is never released.
16 pageslink checked 17 Sept 2026This subject genuinely sits in more than one domain. These fields approach the same ground with different methods.