Turning Point

Darrell M. West has spent his career at the Brookings Institution writing about how technology lands in politics, and Turning Point: Policymaking in the Era of Artificial Intelligence is his attempt, with retired Marine General John R. Allen, to pull together a decade of that work into a single policy map. Published by Brookings Institution Press in 2020, the book arrives at a moment the authors treat as a hinge — the point at which artificial intelligence stops being a laboratory curiosity and starts shaping credit scores, battlefield decisions, hiring, policing, medical diagnoses and national-security strategies. Allen’s presence is the signal of what the book is really about. He commanded coalition forces in Afghanistan and later led the Brookings Institution; he is not a technologist. The co-authorship announces that the book treats AI as a governance problem first and a technical problem second, and that the authors see the national-security dimension as inseparable from the civilian one.

The context the book joined was crowded but strangely thin. By 2020 there were plenty of popular books about AI — Kai-Fu Lee’s AI Superpowers, Max Tegmark’s Life 3.0, Nick Bostrom’s Superintelligence — but each tended to pick a lane. Lee wrote about the US-China race; Tegmark and Bostrom about existential risk; Cathy O’Neil about algorithmic harms in daily life. There were think-tank reports on autonomous weapons from CNAS and SIPRI, white papers from the OECD, and the first wave of national AI strategies out of Beijing, Paris and the Pentagon’s Joint Artificial Intelligence Center. What was missing was a book that sat in the middle of all those conversations at once — one that treated military AI, labor-market AI, criminal-justice AI and smart-city AI as a single policy portfolio belonging to the same elected officials and the same regulators. That is the slot West and Allen set out to fill.

Their central argument is that AI’s defining feature, from a policy standpoint, is not any particular capability but the breadth of its deployment. Because it will be inside almost every sector within a decade, no single regulator and no single committee can own it. The result is a governance vacuum: individual agencies do what they can inside their statutory lanes, Congress holds hearings, companies self-govern, and the actually consequential choices — about surveillance, about automated weapons, about what fraction of a loan decision a human must make — fall between the chairs. The authors want the United States in particular to recognise that the choice is not between regulating AI and leaving it alone, but between designing policy deliberately and letting a drift of unilateral corporate decisions and foreign-made norms settle the questions by default. They treat the possibility that the latter is already happening as the book’s central worry.

The structure of the book is sector by sector, which is both its strength and its clearest editorial choice. After an introduction that frames AI as a general-purpose technology on the scale of electricity or the internal combustion engine, the book works through chapters on finance, transportation, national security and defence, healthcare, criminal justice, education, business and the broader workforce, and smart cities. Each chapter follows the same pattern: a tour of the systems already deployed, the gains and the harms that have shown up in practice, the regulatory landscape as it stands, and a short set of recommendations. A later block of chapters pulls out of the sector view and takes up the cross-cutting questions: ethics, bias and accountability; data privacy; the US-China competition and the shape of international cooperation; and finally a prescriptive chapter on what a serious national AI strategy would look like. The design lets each chapter stand alone, which the authors clearly expected policy staff to use, and it allows the reader to see the common pathologies — opacity, bias in training data, concentration of expertise in a handful of firms — recur in different forms across every domain.

The examples and evidence are what give the book its weight, and this is the register in which the authors do their best work. In the finance chapter they walk through algorithmic trading, robo-advisors and the use of machine learning in credit decisions, with specific attention to the ways the 2010 flash crash exposed the fragility of a market in which the fastest actors are no longer human. The transportation chapter is the most dated now but was fresh at the time: it details the Waymo, Uber and Tesla deployments, the fatal Uber crash in Tempe, Arizona in 2018, and the regulatory patchwork across US states that let autonomous vehicles be tested on public roads with no uniform safety standard. The healthcare chapter walks through IBM Watson’s troubled oncology deployments at MD Anderson, the FDA’s early clearances of AI diagnostic tools in dermatology and ophthalmology, and the question of whether a radiologist can be held liable when they override a machine that turns out to have been right.

For a reader on the defence beat, the national-security chapter is the one that justifies Allen’s co-authorship. It covers Project Maven, the Pentagon’s effort to apply computer vision to drone video and the project whose cancellation under employee pressure drove Google out of defence AI; the formation of the Joint Artificial Intelligence Center in 2018 and the publication of the Department of Defense AI Strategy; the Chinese military-civil fusion doctrine and the role of firms such as iFlyTek and SenseTime in that effort; and the Russian lineage of autonomous and semi-autonomous ground vehicles that would, within a few years of the book’s publication, be tested on real battlefields. The authors devote serious space to lethal autonomous weapons systems — the LAWS debate at the UN Convention on Certain Conventional Weapons in Geneva, the Campaign to Stop Killer Robots, and the specific question of what “meaningful human control” means when a loitering munition can hold a target for hours before being told to strike. They treat the autonomy question as a continuum, not a binary, which is now the mainstream view but was still contested when the book went to press.

The criminal-justice chapter is where the authors are least willing to split the difference. They walk through the COMPAS recidivism-prediction tool and the ProPublica investigation that found it mispredicted along racial lines; the use of facial recognition by police departments in cities including Detroit and Baltimore; and the predictive-policing pilots in Los Angeles and Chicago. The chapter functions as the book’s clearest demonstration of what it means for an AI system to inherit the bias in its training data: a model trained on arrest records does not predict crime, it predicts arrests, and arrests are concentrated where police are already looking. The authors use that chapter to argue for a federal baseline on algorithmic accountability in criminal justice — notice, audit, appeal — which they treat as a template that could extend to employment, housing and credit.

A thread that runs across every chapter is the US-China competition. The authors are clear-eyed about the asymmetry: a political system that can direct capital, data and talent into chosen national priorities, that treats military-civil fusion as doctrine, and that is willing to deploy AI inside its own population without the frictions of a privacy regime, will produce results in some directions faster than a system that must debate every deployment. They do not conclude from that that the United States should imitate the Chinese model. They conclude that it needs its own, deliberately democratic answer — more federal R&D spending, far more serious investment in AI talent pipelines in universities and community colleges, a reform of immigration policy to retain foreign-trained graduates, a coherent federal data-protection framework, and sustained investment in the defence-industrial base’s ability to adopt commercial AI quickly. The specific recommendation that recurs is the one most policy books of this era were converging on: a national AI strategy with real budget authority, not a coordinating council.

On ethics the book is methodical rather than philosophical. The authors run through the familiar list — fairness, transparency, accountability, privacy, safety — and anchor each to institutional mechanisms that already exist. Transparency becomes model-card and dataset-sheet requirements and the right to an explanation; accountability becomes audit regimes modelled on financial-services supervision; safety becomes the certification regimes already used for medical devices and aviation. The move is deliberate: rather than inventing a parallel regulatory structure for AI, they want to graft AI oversight onto the regulators who already understand the sectors in which AI is being deployed. It is a conservative institutional choice, and probably the most durable argument in the book.

Reception was broadly favourable, within the limits imposed by the book’s design. Policy readers treated it as a usable reference and a decent entry point for staffers who needed to be briefed on a sector they had not previously followed. Reviewers in Foreign Affairs and the specialist technology-policy press praised the breadth of coverage and the fact that the authors took the national-security chapter as seriously as the civilian ones, which was not the convention at the time. The pushback came from two directions. Researchers working on long-horizon AI safety felt the book did not engage seriously with the possibility of systems that outperform humans at general tasks, which the authors treat as a secondary concern well behind the near-term governance questions. Civil-liberties critics argued the book’s reliance on existing sectoral regulators understates the degree to which those regulators have been captured or outpaced in the domains that matter — financial supervision, policing, employment. The authors would probably accept both points as fair descriptions of trade-offs they made consciously.

Since publication the policy conversation has moved in some directions the book anticipated and in others it did not. The push for a federal privacy law has gone nowhere, exactly as the authors feared it might. The Pentagon’s AI organisation has grown, acquired a budget line that would have been inconceivable in 2018, and been reorganised at least twice. The arrival of large language models — the public launch of ChatGPT in late 2022, GPT-4 in 2023, and the proliferation of comparable systems since — has shifted the risk conversation toward model capabilities and compute, which the book treats only glancingly. The EU AI Act was finalised in a form that broadly resembles the risk-tiered approach West and Allen recommend, with provisions for high-risk use cases and a specific regime for general-purpose models, while the United States has moved toward a lighter executive-order approach that leans on existing agencies. The Ukraine war has forced the autonomous-weapons debate into contact with ground truth, and the pattern of military adaptation on both sides — commercial quadcopters, FPV strike drones, counter-UAV electronic warfare, battlefield data pipelines — looks a great deal like the civil-military blur the book predicts.

Where Turning Point sits in the field is as a textbook, not a monograph. Someone who already works in defence AI will not find the national-security chapter revelatory, and someone who already works on predictive policing will find the criminal-justice chapter familiar. Where the book earns its place is for the reader who needs to see the whole landscape at once — a committee staffer, a journalist taking the beat for the first time, an industry executive moving between sectors, a European official trying to work out how to answer the authors’ particular American preoccupation with Chinese competition. It pairs naturally with Kai-Fu Lee’s AI Superpowers for the geopolitical half of the picture and with Paul Scharre’s Army of None for a deeper treatment of autonomous weapons, which Turning Point touches but does not dwell on. It does not substitute for either. Readers who want a defence-first survey should start with Scharre and come back to West and Allen for the civilian context.

What is likely to age well is the structural argument — that AI governance will not be solved by a single AI regulator, that it will be solved, if at all, by threading accountability through dozens of existing sectoral regimes. That argument now looks closer to the mainstream of policy practice than it did in 2020. What is already dated is some of the specific technology coverage, inevitably: autonomous vehicles are discussed in a register of imminent deployment that reality has not matched, and the entire post-2022 wave of generative models is absent. The book remains useful for its organising frame and for the quality of its sector-by-sector reporting; it has stopped being useful as a current state-of-the-art document, which the authors would almost certainly concede. For a reader coming to AI policy today, it is a map of how the serious institutional thinking was arranged just before the current wave broke, and that is a specific and valuable thing for a book to be.