AI Guides › Playbooks

The Demo-Reading Card: Judging an AI-Built Game Before You Believe It

By Nigel Guy · 7 min read

Someone shares a clip of an AI generating a playable game in real time, and you do one of two things: declare that games, films and software are finished, or roll your eyes and scroll on. Both feel like judgement and neither is. A demo is a claim about a capability, made by the people who built it, in the best conditions they could arrange.

The rule: never judge a demo by what it shows; judge it by the five things it had to hold together, and by what the people who built it admit it cannot do.

You will probably never make a game. That is fine. The same habit applies to every "AI did X live" clip you will see about spreadsheets, video, code or voice, and it keeps you from buying the wrong thing or ignoring the right one.

What is the capability underneath the demo?

A caveat first. There is no single "AI-built game", so I could not confirm which clip you saw. Two well-documented examples sit behind most of them, and this guide uses both.

The shared idea is that no game engine runs underneath. The model predicts what the next picture should look like given your last input. Microsoft's own wording is worth keeping: think of it as playing the model, as opposed to playing the game. That is why it matters beyond games: it is a system that simulates how a world responds to actions, and that is useful for prototyping, training robots and testing ideas. It is also why it is not yet a way to ship a game.

The Demo-Reading Card

Keep this card next to any demo. Fill it in before forming an opinion.

Line Question Genie 3 / Project Genie WHAMM (Quake II)
1. Who can use it Can you try it, or only watch? Google AI Ultra subscribers, 18+, where Ultra is sold; the launch was US-first A public Copilot Labs experience at launch
2. How long How long does it hold together? 60 seconds per world Short sessions; trained on one part of one level
3. What breaks What do the makers admit? Less realistic worlds, control latency, characters sometimes uncontrollable Enemy interactions are weak
4. What is missing What was shown that you cannot do? Promptable world events from the August preview were not in the prototype Anything outside the trained level
5. What it costs Plan, credits, time? Needs an Ultra plan; check the £ price at checkout Free to try at launch; check whether still live

Line 4 is the one people skip. Google's launch post says some capabilities shown in August, such as events that change the world as you explore, were not in the prototype. A demo can be real and still not be the product.

Step by step: using the card

  1. Find the primary source. Search for the lab's own post or help page, not a reposted clip. For Genie that is Google's Project Genie post and its Labs help page; for Quake II it is the Microsoft Research article.
  2. Fill lines 1 and 2 first. Access and duration are plain facts and settle most arguments. Sixty seconds per world is a very different thing from a game.
  3. Hunt for the limitations section. Labs list them. Copy them onto line 3 in their words.
  4. Compare clip to product (line 4). Note anything in the clip the prototype does not offer.
  5. Write one sentence. "This shows [capability]; it does not yet show [missing thing]." If you cannot finish the sentence, you have not read enough.

Worked example (hypothetical)

Say you run a small design studio, and a colleague posts a clip captioned "AI makes games now, we're doomed." You fill in the card for Project Genie. Line 1: it needs an Ultra subscription, which you do not have, so you cannot test it. Line 2: 60 seconds per world. Line 3: the help page lists latency, a character that sometimes runs backwards, and streams that darken over time. Line 4: the world looks like your prompt only loosely. Your sentence: "This shows a model can generate a navigable world frame by frame; it does not show a repeatable, editable, shippable game." Your response to the colleague is then calm and specific, and your decision for the studio is "watch, do not buy". Compare the cost of an Ultra plan only if prototyping ideas would be genuinely useful to you.

What are the three ways people misread demos?

A prompt to fill in the card

You are a careful research assistant helping a non-specialist judge an AI demo.

Demo: [DEMO_NAME_OR_LINK]
My goal: [WHY_I_CARE, e.g. "decide whether to try it for my business"]
Sources I can show you: [PASTE_OFFICIAL_POST_OR_HELP_TEXT]

Task: fill in a five-line card: (1) who can use it, (2) how long it holds together, (3) limitations the makers state, (4) anything shown in the demo that is not in the product, (5) cost or plan needed.

Rules:
- Use only the text I pasted. If a line cannot be answered from it, write "not in the source" and list what I should look for.
- Do not guess prices, dates or availability.
- Quote the maker's limitation wording briefly; do not soften it.
- End with one sentence: "This shows [X]; it does not yet show [Y]."

Before answering, check each line cites something I pasted.

Fill in the demo name, your reason and the official text.

What to skip

How do you know it's working?

After a month you should be able to read a new demo and, within five minutes, state its access, its duration and one admitted limit without watching the clip twice. If you are still arguing about "what AI can do" in general terms, you are not using the card.

Guardrails

Sources

All 751 AI guides · JulieMango plans from £17/mo