How to compare AI-assisted games

The criteria, in order

1. A complete loop. Can you start, understand an objective, make progress and reach a clear outcome? A compelling generation technique cannot substitute for a playable loop.
2. Controls and clarity. Are the controls explained and responsive on your device? Can you understand what caused a loss or success?
3. Evidence for the AI claim. Does the creator explain which part used AI and when? A tag, visual style or engine name does not establish provenance.
4. Practical fit. Check price, device support, accessibility and whether an account or connection is required for the actual version.
5. Replay value for you. After one session, decide whether the choices and variation make you want another. Do not assume generated content is automatically varied or good.

Use a repeatable scorecard

Pick the same device and session length for each game. Record the version, date and source of each AI claim. For each criterion, write an observation and give a 0–2 score: 0 for a clear problem, 1 for mixed evidence and 2 for a result that meets your needs. Mark unknown when you cannot check; do not turn unknown into a zero or a pass.

For example, “I could not find the restart control” is an observation. “The game has bad AI” is not supported by that observation. This is a suggested evaluation method; no test results are claimed here.

Find candidates and verify the claim

Browse Summer Games for games that fit your preferred genre. Its catalog is a source of candidates, not proof of AI use. The itch.io AI tag page has separate AI-assistance filters; follow each candidate to its creator’s own description and verify what the label means.

If the creator has not documented AI use, you can still evaluate the game as a game. Record its provenance as unknown and leave it out of any ranking that promises verified AI-assisted titles.

Common questions

Are these scores objective?

No. Device, familiarity and taste affect the result. Keeping notes and the same conditions makes your own comparison easier to revisit, but does not make it a universal ranking.

Does a creator’s statement prove every asset’s origin?

No. Treat the statement as evidence for the specific claim it makes, not an audit of the whole project.

Turn the criteria into an ordered decision

Use the criteria in the stated order as gates, with scores as notes rather than a substitute for judgment. If a game cannot start on your device, its loop is untested and it should not win a comparison on a high replay score. If the loop works but instructions are unclear, record the difficulty before moving on. If verified AI assistance is a condition of your list and its evidence is absent, the game is ineligible for that particular list even if it plays well.

The 0–2 scale is deliberately small. For the loop, a 2 means you observed a start, meaningful action and a clear outcome; a 1 means some of those worked but an important part was unclear; a 0 means a reproducible issue prevented the loop. For controls, a 2 means you could reliably perform the necessary actions on the chosen device; a 1 means occasional confusion or difficulty; a 0 means a necessary action could not be performed. These are proposed definitions for your own comparison, not published measurements.

For provenance, record the creator’s exact claim and source rather than grading the amount of AI as good or bad. A specific dated statement about code assistance supports that limited statement; it does not establish that all art was generated. For practical fit, record each requirement separately: device, price, input, connection and account. Replay value is a personal note after the first session, not a universal measure.

Example scorecard: a hypothetical comparison

Imagine two unnamed candidates for a 20-minute laptop session. Candidate A starts and reaches a clear ending, but its creator does not explain AI use. Candidate B has a creator statement that a model helped draft dialogue, but the available build does not start on your laptop. Neither belongs in a verified, playable ranking yet: A lacks the required provenance evidence; B’s play quality is untested on your device.

The next action differs for each. For A, look for a dated development log or creator statement and preserve “unknown” if none exists. For B, check the listed requirements or try a supported device, then repeat the same session. Do not assign B a zero for quality solely because you could not run it. This example illustrates the method; it describes no real game or observed test.

If both later qualify, compare the observed loop first. Use controls to distinguish games with similarly complete loops. Then include the limited provenance statement beside each result and make clear that the order reflects one device, session and player. If the observations are too close or incomplete, publish a tie or no ranking rather than forcing a winner.

Keep an evidence record

1. Record the game name, creator, official listing URL, build or version, device and date.
2. Copy the URL of the creator’s relevant AI statement and summarize its scope in your own words. Record whether it describes development or behavior during play.
3. Record the exact session conditions, including input method and whether you used a demo or a different build from the main listing.
4. Write one concrete observation for each criterion. Mark untested items as unknown and explain why.
5. State the eligibility rule and the order used to compare qualifying candidates. Recheck sources and the version before publishing a result.

Separate a reproducible defect from a preference. “The restart button did not respond on this device after two attempts” is a bounded observation. “The game is impossible to restart” would need wider evidence. Likewise, a creator’s account can support a claim about their process without proving the origin of every component.

Decide what deserves a place on your list

If your list is for players seeking experimental interaction, prioritize the interaction and say whether it uses a model during play. If your list is about production methods, require a creator’s description of the process. If your list is simply about enjoyable games, AI provenance may be irrelevant; select on the experience instead and avoid attaching an unsupported label.

Should a more heavily AI-assisted game rank higher?

Not under this rubric. The degree of assistance is not a quality score. The first two criteria assess the player experience; the provenance criterion checks whether the list’s stated subject is supported.

Can I compare a demo with a full game?

Only if you say which builds you played and define a comparable task. A demo may omit features or end early, so a missing full-game outcome should not be treated as a defect without that context.

What if the creator changes the game?

Keep the old observation tied to its version and date. A new build needs a new check before its result can be compared with the old one.