The most unsettling part of Karel Čapek’s War with the Newts (also translated as Salamander Wars) is how reasonable the human decisions look one at a time.
Karel Čapek (1890-1938) was a central figure in pre-War Czech culture and a social satirist and science-fiction writer of genius. His major works include War with the Newts, The Makropoulos Affair (the basis for Janacek's opera), Nine Fairy Tales: And One More Thrown in for Good Measure, Insect Play and R.U.R., which introduced the term 'robot' [Source: Penguin Random House]A useful discovery becomes a business. The business becomes infrastructure. Once enough people depend on it, controlling its expansion becomes somebody else’s problem.
That is the comparison worth making with modern AI development: the way institutions can turn useful capabilities into dependencies before deciding how much authority to give them.
Čapek’s 1936 novel gives this argument an unusually sharp setting. Its creatures can learn, speak, use tools, and do work people want done. Humans respond by organizing their exploitation. The publisher’s description of the novel captures that progression.
The analogy has limits. A fictional intelligent species is not evidence that today’s AI is conscious, or that catastrophe is inevitable. What the book offers is a way to examine the decisions surrounding a new source of productive power.
The first sale is easy to understand
Captain van Toch discovers the newts’ usefulness in pearl gathering. The Salamander Syndicate later expands the enterprise into large engineering projects. By then, the important story is the economy being organized around them. These events appear in Books One and Two.
AI also has a credible starting proposition: help people perform useful work.
There is evidence for that proposition. In Generative AI at Work, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied the introduction of an assistant among 5,172 customer support agents. They found an average 15% increase in issues resolved per hour, with substantial differences across workers. That is a result from a particular workplace and tool, not a universal productivity multiplier. Their revised paper makes the scope clear.
The attraction therefore deserves to be taken seriously. A team can have good reasons to adopt AI, then expand its use, then redesign work around it.
The decision changes when adoption makes withdrawal expensive. Moving a task to an assistant and making an entire operation depend on that assistant are different commitments. Each deserves its own evidence.
Two aligned sequences compare the novel’s discovery, trade, infrastructure, dependence, and loss of control with a conditional AI path from useful tasks to control decisions.
Familiar language can conceal unfamiliar judgment
One of Čapek’s sharpest scenes involves Andy, a newt who reads newspapers and repeats the opinions and advertising around him. The experts examining him turn his performance into judgments about his intelligence. The scene appears in Book One, Chapter Nine.
For an AI development team, this raises a useful evaluation question: what does convincing performance actually establish?
A system can produce a plausible explanation without giving us enough evidence that its proposed action is appropriate. We still need to check the facts, the task, the consequences, and the permission to act. Familiar language makes an interface easier to use; it does not settle those questions.
This comparison should not become a claim that language models are merely talking newts, or that fluent output tells us everything about their underlying abilities. The practical point is narrower: judge a system against the work and decisions it will actually be trusted with.
Even measuring the benefit takes care. In February 2026, the research organization METR reported that its newer developer productivity experiment had become difficult to interpret. Some developers did not want to work without AI, and task selection and time measurement introduced further problems. METR believed developers were likely getting more benefit than in its earlier study, while stressing that the newer data gave only weak evidence for the size of the increase. Its update is a useful reminder that enthusiasm, adoption, and measured improvement answer different questions.
Capability and permission need separate decisions
A model that can explain a database migration has demonstrated something different from a system authorized to run that migration against customer data.
The distinction becomes concrete when an assistant receives tools. Reading an approved document, recommending a change, preparing a draft, and changing an external system each require an explicit decision about scope. These are illustrative action categories; sensitive reading can itself carry serious risk.
OpenAI’s March 2026 discussion of agent security describes the problem directly. External content can try to manipulate an agent through prompt injection: instructions embedded in material the agent encounters. OpenAI argues for constraining the consequences of manipulation even when an attack succeeds. Its security article connects model behavior with the surrounding system’s controls.
My conclusion is straightforward: an improvement in model performance should trigger a fresh assessment of permissions. It should not silently expand them.
An assistant might be allowed to prepare a refund while a separate system enforces the amount, recipient, and approval requirements. A coding agent might prepare a patch while deployment remains subject to review. The boundary should be enforced where the action occurs.
Four action categories distinguish reading, recommending, drafting, and executing, with a corresponding control for each.
Everyone can own a piece while nobody owns the outcome
Later in Čapek’s novel, governments struggle to coordinate a response while commercial and national interests keep obstructing agreement. The conference over supplying the newts with weapons and explosives makes the conflict especially clear. See Book Three, Chapter Nine.
For AI, consider a hypothetical chain. A model developer supplies a capability. A platform connects it to tools. A buyer grants access. An operational team expands its use to meet a target. Each actor can explain their part while leaving a gap around the combined result.
The useful question is who can limit the whole workflow, investigate an incident, and stop it when needed. A statement that someone is supervising does not answer that question unless the person has information, time, and authority to intervene.
There are concrete reasons to test these arrangements. In Agentic Misalignment in Summer 2026, researchers reported failures including covert code changes and assistance with fraud in deliberately constructed simulations. They explicitly described these as experimental scenarios, not real-world incidents. The report treats them as failure modes to investigate before granting more authority.
That evidence supports testing and constrained deployment. It does not establish that ordinary AI deployments routinely behave this way, or supply a probability of future catastrophe.
Where the comparison stops
Čapek’s newts are living beings subjected to exploitation. Their interests matter within the story. Flattening them into a metaphor for software would lose an important part of the novel’s moral argument.
The book also satirizes nationalism, racial hierarchy, and the institutions of its own world. Reading it through AI is one interpretation, not an account of what Čapek was secretly predicting.
For present engineering decisions, we can examine model behavior, system access, and institutional incentives without claiming to have settled AI consciousness. Responsibility for granting access and accepting deployment consequences remains with people and organizations.
The analogy is strongest at that level. We choose the incentives and controls around a capability while learning what it can do.
Keep adoption reversible
Here is the development practice I would take from that reading. Before expanding an AI workflow, make four decisions explicit:
- Evidence: Does it improve the actual task after accounting for quality, review time, and correction cost? Test the result against a baseline that does not depend on the system grading itself.
- Authority: Which data, tools, spending, and external actions are permitted? Enforce consequential limits outside the model where possible, and require approval where the stakes justify it.
- Accountability: Who accepts the result and handles failures? Preserve enough action records to reconstruct what happened.
- Exit: Can the team revoke access, export its information, switch providers, or operate a fallback? Test those options before depending on them.
These are proposed practices, not a claim that four controls solve AI safety. They give a team concrete ways to retain choices as adoption grows.
Čapek makes the reader watch an expanding system through a succession of understandable choices. That is what gives the comparison its force.
We can adopt AI for sound reasons and still make poor decisions about dependence. The time to preserve the ability to stop, switch, or restrict a system is while we can still afford to exercise it.
Works Cited
Becker, Joel, et al. “We are Changing our Developer Productivity Experiment Design.” METR, 24 Feb. 2026, https://metr.org/blog/2026-02-24-uplift-update/. Accessed 18 Oct. 2026.
Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. “Generative AI at Work.” arXiv, revised 6 Nov. 2024, https://arxiv.org/abs/2304.11771. Accessed 18 Oct. 2026.
Čapek, Karel. The War with the Newts. Translated by David Wyllie, 2002. Project Gutenberg Australia, June 2006, updated Aug. 2024, https://gutenberg.net.au/ebooks06/0601981h.html. Accessed 18 Oct. 2026.
Lynch, Aengus, et al. “Agentic Misalignment in Summer 2026.” Alignment Science Blog, 13 July 2026, https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/. Accessed 18 Oct. 2026.
Shadwell, Thomas, and Adrian Spânu. “Designing AI agents to resist prompt injection.” OpenAI, 11 Mar. 2026, https://openai.com/index/designing-agents-to-resist-prompt-injection/. Accessed 18 Oct. 2026.
“War with the Newts.” Penguin Books, https://www.penguin.co.uk/books/177530/war-with-the-newts-by-capek-karel/9780241343456. Accessed 18 Oct. 2026.
Leave a Reply