Blog · Review Essay · Last reviewed June 16, 2026

The Coming Wave and the Problem of Containment

Mustafa Suleyman and Michael Bhaskar's The Coming Wave is an insider's warning about artificial intelligence, synthetic biology, and the institutional problem of keeping powerful general technologies under human control. Its value is not that it solves containment. Its value is that it names containment as a political, technical, and psychological dilemma at once.

Containment, in this review, means bounded capability under public accountability: evaluation before deployment, constrained access to tools and compute, biosecurity practice, incident reporting, procurement discipline, liability, independent audit, democratic oversight, and real rights to refuse or appeal consequential automation. It is not a single kill switch, and it is not deference to the firms building the systems.

The Book

The Coming Wave: Technology, Power, and the Twenty-first Century's Greatest Dilemma was published by Crown on September 5, 2023. Penguin Random House lists the hardcover at 352 pages, with Mustafa Suleyman as author and Michael Bhaskar as collaborator. Suleyman co-founded DeepMind and Inflection AI; Microsoft announced in March 2024 that he had joined as executive vice president and CEO of Microsoft AI. The publisher frames the book around what it calls "the containment problem": keeping powerful technologies under control while they still can be shaped.

The book's pairing of AI and synthetic biology is its most important move. Suleyman is not writing only about chatbots, model benchmarks, or the next product cycle. He is writing about tools that make intelligence and living systems more programmable, cheaper to manipulate, and easier to distribute. That combination gives the book a wider field of concern than ordinary AI commentary: biosecurity, state capacity, asymmetric power, automation, surveillance, open research, chips, cloud infrastructure, and the incentives of companies trying to build faster than regulators can understand.

This is why the book belongs beside work on platform power, legibility, cybernetics, and surveillance capitalism. Its subject is not technology in isolation. Its subject is the loss of governing distance. Institutions are trying to regulate systems that also become part of how those institutions know, decide, police, fight, treat, persuade, and plan.

What the Wave Metaphor Does

The central image is a wave: a force that gathers from many sources, arrives faster than expected, and cannot simply be wished away. The metaphor has force because it captures diffusion. A single invention can be constrained; a general-purpose stack that improves other stacks is harder to hold in place. Large language models feed programming, research, design, cyber operations, drug discovery, education, administration, and propaganda. Synthetic biology feeds medicine, agriculture, materials, climate intervention, and potentially harm.

The metaphor also carries danger. A wave can make technological direction feel natural, oceanic, and inevitable. If a system is imagined as weather, political choices disappear into fatalism. Investors, labs, and states become surfers rather than authors. That is where readers need caution. The book argues for containment, but its own imagery sometimes risks making containment feel like heroic resistance against nature rather than governance over choices made by specific institutions.

The better reading is to treat the wave as a map of coupling. AI does not just become more capable. It makes other domains move faster. Biology does not just become more programmable. It becomes more dependent on computation, automation, and global supply chains. The political problem is not one machine. It is the interaction of machines, markets, states, labs, and narratives that teach everyone what is supposedly unavoidable.

Containment as a Governance Problem

Suleyman's strongest claim is that containment is both necessary and structurally difficult. The technologies he worries about are general, fast-improving, increasingly autonomous, and capable of asymmetric impact. National Defense University Press summarizes the book's focus as the problem of containing unforeseen consequences from a radical expansion of individual and group power. Issues in Science and Technology notes the tension built into the book's structure: containment appears impossible, then necessary, then partially actionable through a set of proposed steps.

That contradiction is not a flaw to dismiss. It is the actual condition of AI governance. If a society waits for certainty, the systems will already be embedded. If it acts too early, it risks capture, overreach, stagnation, or theatrical rules that only smaller actors obey. If it centralizes power to control AI, it creates new surveillance and state-abuse risks. If it refuses centralization, it may leave public safety to vendor promises and competitive pressure.

This is why containment cannot mean a single lock. It has to mean layered friction: safety engineering, audits, liability, procurement discipline, export controls, lab security, incident reporting, compute governance, biosecurity standards, independent research, worker voice, democratic oversight, and public refusal rights. None of these is sufficient. Together they create more surfaces where runaway capability can be slowed, inspected, redirected, or contested.

The distinction matters because "containment" can be used badly. It can become a marketing word for voluntary self-regulation, a security-state argument for secrecy, or a monopoly argument against open research. A stronger definition starts from legitimacy: who sees the evidence, who sets the thresholds, who can halt deployment, who bears liability, and who is protected when the system is wrong. Containment is not anti-technology. It is anti-irreversibility.

Recursive Capability

The book is most useful when it describes technologies that help produce more technology. AI can assist coding, design experiments, summarize research, automate cyber tasks, generate persuasive media, and reduce the cost of expertise. Synthetic biology can turn information into material intervention. The result is a recursive capability loop: tools improve the conditions for building stronger tools.

That loop matters for belief formation as much as for engineering. Powerful systems do not only change what users can do; they change what users think is possible, necessary, and normal. A model that drafts a policy, ranks a risk, designs an experiment, or simulates an audience becomes part of the institution's imagination. It narrows some futures and makes others feel obvious. The interface becomes a planning environment, then a dependency, then a source of authority.

Containment therefore has a cognitive dimension. A society has to contain not only dangerous outputs but also premature surrender to machine inevitability. It has to preserve the human capacity to say: this deployment is illegible, this delegation is too broad, this automation weakens apprenticeship, this product converts intimacy into leverage, this risk score should be appealable, this model should not be treated as a public oracle.

The Institutional Trap

The book's recurring institutional dilemma is that AI and synthetic biology pressure both sides of the state. Weak states may fail to contain dangerous diffusion. Strong states may use the same tools for surveillance, control, military advantage, and bureaucratic expansion. Axios reported Suleyman's warning that AI could help autocratic governments centralize power and intensify surveillance, while also expanding the attack surface around those regimes.

That double movement is already visible in ordinary institutional life. A government buys AI for fraud detection and creates new opacity for citizens. A school adopts AI tutors and risks weakening human apprenticeship. A hospital uses automation to triage care and may hide value judgments inside workflow software. A company deploys assistants to increase productivity and quietly turns workers into monitored operators of vendor infrastructure.

The Coming Wave is best read as a warning against both naive openness and naive control. Open diffusion can empower researchers, small firms, journalists, disabled users, and public-interest builders. It can also empower fraud, cybercrime, bioterror speculation, and unaccountable deployment. Centralized control can support safety testing, coordination, and accountability. It can also harden monopoly, censorship, and security-state reflexes. The task is not to pick a slogan. It is to design institutions that can distinguish capability from legitimacy.

The 2026 Context

By June 16, 2026, the containment debate had moved from book-length warning into law, standards, and institutional procedure. The EU AI Act entered into force on August 1, 2024; its general-purpose AI model obligations became applicable on August 2, 2025, and the European Commission describes additional duties for models with systemic risk, including model evaluation, systemic-risk assessment and mitigation, serious-incident reporting, and cybersecurity. The broader Act is scheduled to become fully applicable on August 2, 2026, with some product-integrated high-risk rules on a longer timeline.

In the United States, NIST's AI Risk Management Framework remains voluntary, but it has become a reference point for mapping and managing risks across the AI lifecycle. NIST's Generative AI Profile adapts that framework to generative systems, and its 2026 AI Agent Standards Initiative points to a newer containment problem: systems that not only produce text but act through software, identity, permissions, tools, and other services.

The 2026 International AI Safety Report adds useful caution. It reports rapidly improving capabilities, persistent reliability failures, information asymmetries, cyber and biosecurity concerns, and an evaluation gap between controlled tests and real-world settings. It also states that current systems do not have the capabilities assumed in full loss-of-control scenarios, while noting improvement in relevant areas such as autonomous operation. That is the right level of discipline: serious risk without metaphysical inflation.

Governance and Safety

A practical containment program needs more than frontier-model evals. It needs at least five layers. First, capability gating: pre-release evaluations, red-team access, staged deployment, model and system cards, safety cases, and documented thresholds for refusal, rollback, or restricted release. Second, infrastructure controls: cloud security, weight protection, compute accounting where appropriate, access logs, identity and authorization for agents, and incident reporting that reaches regulators rather than staying inside trust-and-safety dashboards.

Third, downstream governance: procurement terms that require audit trails, appeal channels, human review in consequential decisions, data-retention limits, and liability for foreseeable misuse or negligent deployment. Fourth, biosecurity: synthesis screening, lab security, access controls for biological design tools, and risk assessment that distinguishes internet-accessible information from AI-enabled uplift in real workflows. Fifth, democratic friction: courts, auditors, researchers, workers, affected communities, and journalists need standing to examine systems whose harms are otherwise privatized as trade secrets.

This is where the book's containment theme meets the site's recurring concern with interfaces and belief. A model is not only a tool after it is embedded in a bureaucracy. It becomes a way of seeing. The spreadsheet, queue, score, assistant, and dashboard decide which facts are easy to notice and which claims become expensive to contest. Safety work therefore has to include interpretability, documentation, due process, and institutional memory, not only emergency controls for hypothetical future systems.

Where the Book Needs Friction

The book's weakness is its insider altitude. Suleyman is unusually candid for a technology executive, and reviewers have noted the value of his historical sweep and practical proposals. But the same perspective can understate the labor, data, and institutional harms already produced by the systems whose future risks he foregrounds. Washington Post reviewer Noah Giansiracusa praised the book as expansive and historically rooted while also faulting it for assumptions about exponential progress, limited attention to the human cost of AI, and a questionable posture toward open source.

The wave frame also needs social detail. Who gets automated first? Who labels the data? Who absorbs moderation trauma? Who loses an entry-level path into competence? Who is denied benefits by a risk score? Who lives near the data center or under the surveillance system? Far-future containment can become evasive if it floats above present extraction.

The book should therefore be read with Atlas of AI, Automating Inequality, Behind the Screen, Programmed Inequality, and Power and Progress. Suleyman gives a strategic problem statement. Those books supply the ground truth of matter, labor, classification, and power.

It also needs a firmer account of power inside containment itself. Who writes the benchmark? Who certifies the safety case? Who decides which models count as systemic risk? Who can afford compliance? If only the largest labs can satisfy the rules, safety becomes a path to market concentration. If rules are too weak, concentration becomes self-policing. The containment problem is therefore inseparable from antitrust, public research capacity, labor rights, and democratic oversight.

What This Changes

The Coming Wave is a book about authority arriving through usefulness. The most dangerous systems are not always the ones that announce domination. They are the ones that make themselves necessary: to search, decide, monitor, design, teach, diagnose, persuade, and govern.

The containment problem is not only whether humanity can keep future super-systems from escaping control. It is whether ordinary institutions can resist the smaller daily surrender by which control is handed away. A dashboard becomes the measure of reality. A model output becomes the first draft of policy. A safety case becomes a sales document. A chatbot becomes the front door to care. A synthetic consensus starts to feel like public opinion.

The book's best contribution is its insistence that technical power needs active political imagination before crisis. Its weakest temptation is the belief that the people building the wave can also be trusted to define the shoreline. A serious containment politics has to include builders, but it cannot be governed by builder psychology alone. It needs workers, publics, auditors, courts, researchers, teachers, patients, and people with the right to refuse systems that classify or replace them.

Read this way, The Coming Wave is less a final answer than a useful pressure test. Any institution that says it has an AI strategy should be able to answer a few concrete questions: what capabilities are being delegated, what harms are being made irreversible, who can inspect the system, who can appeal, who benefits from speed, and what human capacity is being preserved rather than merely simulated.

Source Discipline

The book's premise should be kept separate from evidence about present systems. It is fair to say that frontier AI, agentic software, and bioengineering tools create new governance problems. It is not sound to treat every deployed AI product as autonomous, sentient, or civilization-ending. Safety claims need named systems, named capabilities, deployment contexts, and failure evidence.

Biosecurity claims require extra care. The 2026 International AI Safety Report says general-purpose AI can contextualize biological and chemical weapons information and may lower some expertise barriers, while also emphasizing substantial uncertainty, material barriers, and the difficulty of real-world uplift studies. That means the right response is not panic language. It is controlled access, synthesis screening, lab practice, careful evaluation, and transparent limits on what the public evidence can support.

For readers, a simple test helps: if an argument about containment does not say who can audit, who can appeal, who can halt, and who is liable, it is probably a slogan. If it says those things but omits labor, data, civil liberties, and market concentration, it may be safety governance in form and institutional surrender in substance.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books