# The reading room | Eat. Shift. Or Die. Canonical: https://eatshiftdie.com/sources.html Sources, arguments & useful questions # Follow the thinking. A small, annotated collection. Start with a question, follow the original material and make up your own mind. Collection 01 ## Good at the task. Useful in the work? Studies that make a blanket claim about “AI productivity” harder to sustain. Read them together. Field experiment2023 working paper ### [Navigating the Jagged Technological Frontier ↗](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/) Fabrizio Dell’Acqua and colleagues Harvard Business School AI assistance improved performance on some consulting tasks and worsened it on another task beyond the tested capability frontier. Why it belongs here A useful reason to test the actual work, rather than assume that success on a neighbouring task will carry across. Read with care. A specific model, group of participants and task set. This is not a current ranking of AI tools. Which part of your workflow could look similar to a successful demonstration while demanding a different kind of judgement? [Open the original ↗](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/) Workplace study2023 · revised 2024 ### [Generative AI at Work ↗](https://arxiv.org/abs/2304.11771) Erik Brynjolfsson, Danielle Li and Lindsey Raymond Author manuscript · arXiv The introduction of an AI assistant in customer support was associated with productivity gains that differed across workers, with larger benefits for less experienced workers. Why it belongs here A concrete setting for thinking about how assistance can change access to expertise, rather than simply replace it. Read with care. Results from a particular workplace and task. A productivity effect there does not establish the effect in your organisation. Whose knowledge becomes easier to use, and how will the next generation learn to develop it? [Open the original ↗](https://arxiv.org/abs/2304.11771) Field reportJune 2025 ### [Microsoft 365 Copilot Experiment ↗](https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html) Checked 8 September 2026 Government Digital Service Cross-Government Findings Report Twenty thousand UK government employees, twelve organisations, three months. Read it alongside the Australian APS trial and compare where people found value. Why it belongs here A large public-sector trial offers a view of everyday use across different kinds of work. Read with care. Self-reported time savings, one product family, tools tested in late 2024. Participation is not the same as a measured productivity gain. What would this experiment look like at your organisation’s size, and who would read the result? [Open the original ↗](https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html) [Put the question to workWhere did the time go? ↗](https://eatshiftdie.com/tool-workflow.html) Collection 02 ## Who is doing the thinking? A closer look at the effort that moves around the answer, and the limits of what the evidence can tell us. Survey2025 ### [The Impact of Generative AI on Critical Thinking ↗](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/) Hao-Ping Lee and colleagues Microsoft Research · CHI 2025 A survey of knowledge workers examined how confidence and AI use relate to self-reported critical thinking effort. Why it belongs here It asks us to examine the human effort around checking, integrating and accepting an answer. Read with care. Self-reported survey evidence. It does not prove that AI causes a lasting loss of critical thinking ability. When an answer arrives looking finished, what makes you decide to inspect it? [Open the original ↗](https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/) Survey2026 ### [The Work AI Index 2026 ↗](https://www.glean.com/work-ai-institute/reports/work-ai-index) Checked 8 September 2026 Glean Work AI Institute and university co-authors Global report Six thousand digital workers across the US, UK and Australia, and a useful name for the work behind the output. Botsitting is the effort of making AI usable. Why it belongs here The checking, context and repair behind an output often disappear from the productivity story. Read with care. Vendor-published, with university co-authors. Self-reported experience from digital workers, not a productivity audit of the whole workforce. Who in your team is botsitting, and does anyone’s dashboard know? [Open the original ↗](https://www.glean.com/work-ai-institute/reports/work-ai-index) [Put the question to workWho gets to decide? ↗](https://eatshiftdie.com/tool-delegation.html) Collection 03 ## A machine answers back. A historical starting point, two books and practical guidance for what happens next. Original paper1966 ### [ELIZA ↗](https://doi.org/10.1145/365153.365168) Joseph Weizenbaum Communications of the ACM Weizenbaum described a program that used decomposition and reassembly rules to conduct text exchanges in natural language. Why it belongs here The historical starting point for the book’s question about what happens when a machine appears to answer back. Read with care. A historical language-processing program. The exercise here is inspired by simple pattern matching and is not the original implementation. What do you bring to a conversation that the program itself does not know? [Open the original ↗](https://doi.org/10.1145/365153.365168) Design research2019 ### [Guidelines for Human-AI Interaction ↗](https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/) Saleema Amershi and colleagues Microsoft Research · CHI 2019 Eighteen guidelines address how AI systems should behave initially, during interaction, when wrong and over time. Why it belongs here A practical way to think about the relationship around the model: setting expectations, supporting correction and adapting carefully. Read with care. Design guidance is a starting point. It does not remove the need to test a particular system with the people using it. What should a useful AI colleague do when it is wrong? [Open the original ↗](https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/) Book2024 ### [Co-Intelligence ↗](https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/) Checked 8 September 2026 Ethan Mollick Portfolio A practical account of working alongside AI, from one of the researchers behind the jagged frontier. Why it belongs here A companion for moving from impressive demonstrations to deliberate practice. Read with care. A fast-moving field. Treat specific tool advice as dated, the frames as durable. Which of your tasks sits just beyond the frontier, and how would you find out? [Open the original ↗](https://www.penguinrandomhouse.com/books/741805/co-intelligence-by-ethan-mollick/) Book2006 · second edition 2016 ### [The Box ↗](https://marclevinson.net/the-box/) Checked 8 September 2026 Marc Levinson Princeton University Press The container story the book retells. A simple box needed ports, standards and working practices to change before it could transform trade. Why it belongs here It puts the work around a new capability at the centre of the story. Read with care. Economic history. The parallel to AI is the reader’s to test. What is your industry’s equivalent of a port built for the old way of moving things? [Open the original ↗](https://marclevinson.net/the-box/) [Put the question to workTry the ELIZA-inspired exercise ↗](https://eatshiftdie.com/eliza.html) Collection updated 8 September 2026. The four additions are individually dated. Earlier selections were checked on 5 September 2026. These annotations connect sources to the site’s questions; they are not a complete bibliography of the book. Access conditions may vary by publisher.