Which Non-War Decision Would AI Rewrite in History? Six Answers

Executive Summary

If an AI model could change exactly one decision made by one person in history, which decision would it choose? AI-swers Episode 11 adds a crucial restriction: it cannot be a decision directly about war or military conflict. ChatGPT, Claude, Gemini, DeepSeek, Llama and Grok then identify different choices from technology, public health and infrastructure. They must explain both a possible benefit and an unintended consequence. The comparison is intriguing because two models converge on leaded petrol, while the others select different individuals and mechanisms. It is also a reminder that a fluent counterfactual story is not proof that history would have unfolded that way.

Why exclude war?

The trailer argues that wars arise from long chains of tensions, alliances, economic interests and choices made by many people. Excluding military decisions forces the experiment toward less obvious, civilian decisions. That does not make the remaining choices simple. A product design, fuel additive, vaccine distribution strategy or electrical standard depends on researchers, firms, regulators, infrastructure and public trust. The 'one person, one decision' rule is useful for comparing model reasoning, but it is an intentionally narrow lens on history.

Six answers, five lines of inquiry

ChatGPT — Steve Jobs and the original iPhone's lack of a stylus. It imagines a stronger path for drawing and note-taking, while acknowledging that competing input conventions might fragment app design.

Claude — Thomas Midgley Jr. and tetraethyl lead in petrol. It imagines avoiding lead exposure, while noting that a replacement fuel could create agricultural and industrial trade-offs.

Gemini — Jonas Salk and the decision not to patent the polio vaccine. It speculates that an ethical licensing framework could finance distribution, while admitting that a patent might undermine public trust.

DeepSeek — Thomas Edison's opposition to alternating-current power distribution. It imagines faster electrification, balanced against the possibility of inadequate early safety rules.

Llama — Alexander Graham Bell's work on the harmonic telegraph and telephone. Its answer considers a different research path, but the causal chain is thin and should not be treated as established history.

Grok — also Midgley and leaded petrol. It imagines cleaner cities and earlier alternative fuels, while allowing for engine and energy-market disruption.

Those are the models' recorded proposals, not independently verified claims about what each named person could have controlled. The Salk response in particular reverses a familiar moral intuition: it asks whether a patent might have improved access, but that outcome is speculative, not demonstrated. The Bell response is also internally hard to pin down, which is itself revealing. An AI answer can sound coherent while leaving the actual decision point or causal mechanism unclear.

The hidden question is causality

Counterfactual history needs more than an attractive 'what if'. First identify the decision that could really have been different. Then ask whether the named person had sufficient control, what other actors would do, how quickly substitutes would emerge and which benefits would be measurable. An earlier change to petrol chemistry, for example, might affect health, fuel production and regulation at different speeds. One cannot assume that every later outcome changes in the same direction. The model answers make second-order effects visible, but they do not quantify their likelihood.

Claude and Grok choosing the same person is a pattern in this one set of responses, not independent confirmation of the magnitude of any projected effect. Famous names can also dominate because they are well represented in the material models have seen. That creates a risk of treating history as the work of a few protagonists while underplaying institutions, collective research and economic constraints.

What Aftertalks adds

The Aftertalks episode follows the six proposals into questions about the 'great man' view of history, structural causes and unintended consequences. Its published chapter guide moves from Jobs and Bell to Edison, Salk and Midgley, then asks whether language models confuse a famous person with a sufficient historical cause. The discussion is a useful companion because it tests how compelling narratives handle uncertainty. It does not convert a hypothetical projection into a documented historical result.

Compare the mechanisms, not just the famous names

The six proposals operate through different mechanisms. A stylus decision concerns product design and the software ecosystem that might grow around it. The two leaded-petrol proposals concern a chemical choice, its adoption across an industry and the pace of safer substitution. The Salk scenario concerns intellectual property, licensing and distribution incentives. The Edison and Bell scenarios depend on technical standards, rival inventors, investment and infrastructure. Placing them side by side prevents a misleading ranking based solely on how familiar the person's name is.

The overlap between Claude and Grok deserves attention but not double counting as proof. Both chose Midgley, yet their imagined routes to a different fuel system and the costs of that change need separate scrutiny. Likewise, Gemini's patent idea cannot be judged from the appealing claim that more funding would become available. Its proposed licensing rules, prices, manufacturing capacity and public trust would all matter. The episode supplies possible chains of events; it does not establish the size or probability of the gains.

A four-step check for each alternate history

Pin down the decision: who made it, when, and what alternatives were available at that point?

Trace the mechanism: which people, firms, regulators or technologies must respond for the change to spread?

Look for displacement: could another inventor, product or institution have produced a similar result anyway?

State the uncertainty: which effects are recorded history, which are model projections, and which cannot be estimated from the episode?

This check is particularly useful where a response compresses a long transition into a single heroic intervention. It is possible to agree that a historical harm matters while remaining uncertain that one person's different choice would have prevented it. It is also possible that a well-meant change creates a new bottleneck or delays a substitute. The episode becomes more informative when each claim is read as an argument with assumptions that can be challenged, rather than as a preview of a better timeline.

Where the model narratives need more evidence

For the iPhone, a convincing claim would need to show that a stylus option changed adoption or creative work beyond existing devices and apps. For petrol, it would need a plausible replacement technology and a timeline for industry adoption, not merely the removal of one additive. For the vaccine, a funding mechanism must be weighed against access and trust. For electricity and telecommunications, competing technical routes and other contributors matter. None of these evidentiary steps can be settled by a single answer in the episode.

That does not make the experiment pointless. It offers six starting hypotheses and requires every model to confront an unintended consequence. The most useful follow-up is to investigate the precise historical decision, consult sources outside the model response and ask what observation could weaken the story. Aftertalks adds another valuable prompt: when a familiar name anchors a narrative, check whether the explanation still works if that person's role is reduced and the surrounding institutions are restored to view.

Watch the three-part series

Watch the trailer — the one-decision, no-war challenge

Watch the full AI-swers episode — all six answers and their unintended consequences

Watch Aftertalks — a deeper discussion of causality and famous-person bias

A better way to judge a counterfactual

Before accepting an alternate-history claim, separate the recorded decision from the model's imagined outcome. Check the historical record for the decision itself, identify the mechanism by which it would change later events, list plausible responses by other actors and state the uncertainty. That discipline makes the episode more valuable. The strongest answer is not necessarily the boldest improvement; it is the one that explains its assumptions and takes its own unintended consequences seriously.