AI History Games: How the Scenes Are Made, and What to Trust

A few years ago, a history guessing game could only show you a photograph, which meant it could only show you the last two centuries. Image models removed that limit, and a whole branch of the genre grew in the gap: games that generate a historical scene instead of finding one.

That's a real gain. It's also a category of content that deserves more skepticism than it usually gets, including from us — we make one of these games. This guide explains how the scenes are actually made, the specific ways they go wrong, and the questions worth asking before you trust what one of them shows you.

Everything here was checked in September 2026.

How an AI historical scene is actually made

The pipeline is shorter than most people assume:

1. Someone writes a description of a historical moment — the event, the place, the period, what should be visible.
2. An image model turns that description into a picture.
3. Someone decides whether the result is good enough to use.

The critical thing to understand is what happens in step two. The model is not consulting a historical record. It has no access to archives, no ability to check a date, no representation of what actually happened. It produces an image that matches the statistical shape of pictures associated with those words — what images of "a Roman harbor at dawn" tend to look like, given everything it was trained on.

That distinction is the whole basis for how much weight these images can bear. An AI historical scene is an illustration of a description, not evidence of an event. It belongs in the same category as a 19th-century history painting: an interpretation, made later, by someone who wasn't there — except that the someone is a model, and it can produce a thousand a day.

The five ways they go wrong

1. Anachronism that looks plausible

The classic failure. A tool, a garment, a building technique or a weapon from the wrong century, rendered as convincingly as everything around it. Models blend visual conventions across periods, and the seams don't show, because nothing in the image is flagged as less certain than anything else.

2. Invented specificity

A model asked for a particular named moment will happily produce a picture that looks like a record of that exact moment — specific faces, specific gestures, a specific arrangement of people. None of it is documented. The composition is invented, and the confidence of the rendering is what makes it misleading.

3. Training-data bias

Models reflect what they were trained on, and the visual record is wildly uneven. Western European and North American history has been painted, illustrated, photographed and filmed far more than most of the world, so that's what models are fluent in. Ask for an event outside that band and the output is thinner, more generic, and more likely to fall back on cliché.

This one deserves a note about our own game. Where&When's event pool leans heavily toward Europe and North America, and unlike a photo-based game, we can't blame the archives — we generate our images, so the balance of the pool is entirely our responsibility. It's a known problem we're working on, and it's fair to hold any game in this category to the same standard.

4. Photoreal portraits of real people

The sharpest line in the category. Generating a photorealistic image of an identifiable real person doing something they are only recorded as having done — or worse, something they didn't — manufactures a fake record of a real life. Scenes of events are one thing; synthetic portraiture of named individuals is another, and it's the place where this technology does the most damage for the least benefit.

5. Text, flags and insignia

Models mangle writing. Signage, banners, coats of arms and flags come out as convincing-looking nonsense, which is both a quality problem and an accuracy problem — a garbled flag in a scene can point a player toward the wrong country entirely.

What a responsible AI history game does

A checklist you can apply to any of them, ours included:

• Calls the images what they are. Reconstructions or illustrations, never "photographs" and never "archive footage."
• Puts the verifiable history somewhere. If the image is an interpretation, the reveal needs to carry the actual claim — the event, the place, the date, what happened — so there's something checkable behind the picture.
• Puts a human between the model and the player. Generation is cheap, which means the only thing standing between a hallucinated scene and a player is whether somebody looked.
• Doesn't fake real people. No photoreal portraits of identifiable individuals.
• Is honest about its coverage. Every pool in this genre is skewed. The ones worth trusting say so.

The games doing this

EraGuessr

EraGuessr builds AI-generated 360° panoramas of historical moments that you can look around before pinning the map and setting a year. It's the most developed version of the idea: a daily puzzle that's free and needs no signup, plus party rooms, ranked matches and a scene studio. EraGuessr+ costs $3.99 a month or $19.99 a year and unlocks the full scene library, filtering by era and place, the daily archive and scene generation.

WenWare

WenWare was the breakout of the category — a browser game built during a 2026 game jam that went viral in late April: one AI-generated historical panorama a day, 60 seconds, a pin and a year slider, and a shareable grid.

It's also a cautionary tale about the surrounding ecosystem. Its success produced a cluster of near-identical sites, several of which describe themselves as the official one. We're not linking any of them here, because we can't establish which is which — and that ambiguity is itself worth knowing about before you type a game's name into a search bar.

Where&When

Where&When is ours, so here is exactly what we do, stated so you can hold us to it.

Each round shows an AI-generated scene of a real, named historical event, with no caption. You pin where it happened and enter the year; the reveal then names the event, gives its place and date, and tells you what happened and why it mattered. The images are illustrated reconstructions — we describe them that way in the App Store listing, in this guide and everywhere else, and we never call them photographs.

Our practices: every image is reviewed by a person before it reaches a player; the verifiable history lives in the reveal text, not in the picture; we don't generate photoreal portraits of identifiable real people; and, as above, our pool's regional balance is skewed and is our own problem to fix.

What we get in exchange for all that is range. Photography begins in the 1830s. Generating scenes instead means the pool runs from the 1st century to the 2020s, so a round can come from antiquity, the middle ages, or last decade — which is the whole reason to build the game this way rather than licensing a photo archive.

Does the AI part make the game worse?

For the dating skill specifically, it changes what you're reading rather than removing it.

A photograph carries physical evidence: film characteristics, lens behavior, the precise look of a color process from a particular decade. Those are the tells that photo-dating games train, and an illustrated scene doesn't have them. What it has instead is the content — architecture, dress, tools, transport, the landscape — which is the older and more general skill, and the only one available for anything before the 1830s.

So a photo game trains a sharper instrument over a narrower range, and a generated-scene game trains a blunter one over all of recorded history. Both are real. If you want the full method for either, it's in how to guess the year of a historical scene, and the comparison of which games ask what is in GeoGuessr for history.

Frequently asked questions

Are AI-generated historical images accurate?

They're accurate the way an illustration is accurate: as good as the description behind them and the review in front of them, and never a record of the event itself. Treat the scene as an interpretation and the accompanying text as the claim to check.

What's the difference between an AI history game and a photo-based one?

Coverage and evidence. Photo games (TimeGuessr, WhenTaken, Chronophoto) use real photographs, so they're limited to roughly the 1830s onward but show you genuine documentary material. AI games (EraGuessr, WenWare, Where&When) can depict any period, but what they show is reconstructed.

Is it ethical to make history games with AI images?

It depends almost entirely on disclosure and restraint. Labeling reconstructions as reconstructions, keeping the verifiable history attached, reviewing output before publishing, and refusing to fake real individuals are what separate a reasonable version of this from a machine for producing convincing false memories.

Can AI history games be used in school?

With explicit framing, yes — and the framing is half the value. Telling students the images are reconstructions and asking them to separate what's documented from what's been filled in is a media-literacy lesson in its own right. We go through classroom use in history games for students.