Proposes a short standard document to ship with every trained model: what it is for, who it was tested on, where it performs worse, and what it should not be used for. Most model documentation you see today descends from this. Free on arXiv.
From the site: Trained machine learning models are increasingly used to perform high-impact tasks in areas such as law enforcement, medicine, education, and employment. In order to clarify the intended use cases of machine learning models and minimize their usage in contexts for which they are not well suited, we recommend that rele…
Why I recommend it: This is the practical end of AI ethics — a template, not an argument. Useful if you ever have to evaluate a vendor's model.
Separates the technical question (how do you make a system pursue a goal) from the normative one (whose values, chosen how). Argues no single person's values are a legitimate target and looks at fair-process alternatives. Free on arXiv.
From the site: This paper looks at philosophical questions that arise in the context of AI alignment. It defends three propositions. First, normative and technical aspects of the AI alignment problem are interrelated, creating space for productive engagement between people working in both domains. Second, it is important to be clear…
Why I recommend it: The clearest philosophical treatment of “aligned to what?” I have found. Skip the lab blog posts and read this instead.
A long report on what changes when AI acts on your behalf rather than answering questions: manipulation, anthropomorphism, misaligned delegation, and what happens when everyone has an assistant at once. Free on arXiv.
From the site: This paper focuses on the opportunities and the ethical and societal risks posed by advanced AI assistants. We define advanced AI assistants as artificial agents with natural language interfaces, whose function is to plan and execute sequences of actions on behalf of a user, across one or more domains, in line with th…
Why I recommend it: Dense but the most thorough thing published on agent ethics. Use the section headings to read only the parts you need.
Argues that before release, labs should test models for dangerous capabilities and for whether they will apply them — and sets out what responsible release decisions would look like. Free on arXiv.
From the site: Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is…
Why I recommend it: This is where today's “frontier safety framework” language comes from. Written largely by the labs it would govern, which is worth holding in mind.
Argues that ever-larger language models carry costs that scale with them: environmental cost, unauditable training data, encoded bias, and the illusion of understanding. The source of the phrase “stochastic parrot.” Free to read on the ACM site.
Why I recommend it: The paper that got two of its authors pushed out of Google. Worth reading before you accept either the hype or the dismissal of it.
Borrows the electronics-industry datasheet idea for training data: how it was collected, who is in it, who consented, and what it should not be used for. Free on arXiv.
From the site: The machine learning community currently has no standardized process for documenting datasets, which can lead to severe consequences in high-stakes domains. To address this gap, we propose datasheets for datasets. In the electronics industry, every component, no matter how simple or complex, is accompanied with a data…
Why I recommend it: Pairs directly with Model Cards. Together they are the closest thing the field has to a documentation standard.
Reviewed 146 papers on bias in language technology and found most never say who is harmed or how. Argues bias work has to start from real-world power relations, not just from a metric. Free on arXiv.
From the site: We survey 146 papers analyzing "bias" in NLP systems, finding that their motivations are often vague, inconsistent, and lacking in normative reasoning, despite the fact that analyzing "bias" is an inherently normative process. We further find that these papers' proposed quantitative techniques for measuring or mitigat…
Why I recommend it: Read this if “bias” has started to sound like a box to tick. It is a careful takedown of shallow fairness work by people who do the work.
The founding technical paper on algorithmic fairness: defines fairness as treating similar individuals similarly, and shows why blindness to a protected attribute does not deliver it. Mathematical. Free on arXiv.
From the site: We study fairness in classification, where individuals are classified, e.g., admitted to a university, and the goal is to prevent discrimination against individuals based on their membership in some group, while maintaining utility for the classifier (the university). The main conceptual contribution of this paper is…
Why I recommend it: The math is heavy, but the first few pages explain why “we just don't collect race” is not a fairness strategy.
Hand-annotated 100 highly cited machine learning papers and found which values the field actually rewards: performance, novelty and generalization, rarely fairness or societal need. Also traces the funding behind the work. Free on arXiv.
From the site: Machine learning currently exerts an outsized influence on the world, increasingly affecting institutional practices and impacted communities. It is therefore critical that we question vague conceptions of the field as value-neutral or universally beneficial, and investigate what specific values the field is advancing…
Why I recommend it: Turns “the field has blind spots” into countable evidence. Good antidote to the idea that research priorities are neutral.
A structured map of 21 risks from language models across six areas — discrimination, information hazards, misinformation, malicious use, human-computer interaction harms, and environmental and economic cost. Free on arXiv.
From the site: This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order to foster advances in responsible innovation, an in-depth understanding of the potential risks posed by these models is needed. A wide range of established and anticipated risks are analysed in detail, draw…
Why I recommend it: The best single reference if you need vocabulary for a specific harm rather than a general argument. Written by a lab, so read it as a lab's own framing.
Tested three commercial face-classification products and found error rates of up to 34.7% for darker-skinned women against 0.8% for lighter-skinned men. The study that turned algorithmic bias from a theory into a measured, published fact. Free to read in full.
From the site: Recent studies demonstrate that machine learning algorithms can discriminate based on classes like race and gender. In this work, we present an approach to evaluate bias present in automated facial...
Why I recommend it: If you read one AI ethics paper, read this one. It is short, the method is easy to follow, and it is the reason facial recognition audits exist at all.
Shows that “interpretable” is used to mean several incompatible things, and that simpler models are not automatically more honest about what they do. Free on arXiv.
From the site: Supervised machine learning models boast remarkable predictive capabilities. But can you trust your model? Will it work in deployment? What else can it tell you about the world? We want models to be not only good, but interpretable. And yet the task of interpretation appears underspecified. Papers provide diverse and…
Why I recommend it: Useful skepticism to carry into any conversation about explainable AI, especially a vendor's.
SSRN working paper by Eldar Maksymov applying the Jevons Paradox to AI-driven labor changes. Argues that, like spreadsheets with accounting, AI may expand demand for judgment-intensive knowledge work and that leaders should build a value fortress of trust and accountability rather than cut headcount.
#ai#ai ethics#economics#executives#future of work#jevons paradox#job markets#paper#research#strategy
SSRN working paper by Carla Zoe Cremer (Oxford) and Luke Kemp (Cambridge) examining how existential risk studies can be made more rigorous, pluralistic and democratic. Argues for separating extinction ethics from risk analysis and drawing on broader risk-assessment literature.
