AI policy and AI system impact assessment

What the AI system is

The AI system is the set of computer-controlled sharks that populate a tank alongside human players. There are up to 24 of them per tank.

It is worth being exact about what they are, because the term invites an assumption this service would not survive. They are deterministic rules, not a learned model. Each shark, on each tick, steers back toward the middle if it is past four fifths of the way to the arena wall; otherwise it turns toward the nearest food inside a fixed sight radius; otherwise it makes an occasional random turn drawn from the tank's seeded generator, offset by its own identifier so a tank of sharks does not pivot in unison. Two rules complete the set: while a feeding frenzy is running the sight radius widens and a shark with nothing in sight heads for the middle, and a shark that gets within a short distance of high-value food charges briefly and then lunges at it under a cooldown. Firing a rocket is a player-only ability that the steering rules cannot reach.

There is no model, no training, no training data, no inference call and no third-party AI service anywhere in this system. Nothing about a shark's behaviour changes in response to anything a player does beyond the food and walls in front of it.

Intended use, and limits on use

The intended use is narrow and complete: populate a tank with opponents so a player is never swimming alone.

The system takes no decision about any person. It does not rank, score, profile, moderate, price, recommend or gate anything. It does not read a player's display name, profile or history. Its entire input is the tank's own state: the positions of food, its distance from the arena wall, the tick, and whether a feeding frenzy is running.

Any use beyond populating a tank is outside this policy and would require it to be rewritten before that use ships.

Impact assessment — individuals

Assessed impact on an individual: limited to the experience of playing a game. A shark can end a player's run by colliding with them, which is the game working as described.

No personal data reaches the system. It processes tank state — positions and food — and nothing that identifies a player. There is therefore no profiling, no automated decision with legal or similarly significant effect, and no basis for discriminatory outcome between players, because the system cannot distinguish one player from another.

Fairness is bounded by design rather than by monitoring: every shark runs the identical rule set, the only variation between them is a phase offset taken from the shark's own identifier so their wandering does not synchronise, and the whole tank is reproducible from its seed and action stream. Nothing in the rules varies by which player is in the tank, because nothing in the rules can see one.

Impact assessment — groups and society

Assessed societal impact: negligible, and stated as such rather than left implied. The system produces no content, makes no claim, reaches no one outside the tank, and has no downstream consumer of its outputs.

The residual concerns normally raised at this point — misinformation, labour displacement, surveillance, environmental cost at scale — do not apply to twenty-four rule-driven sharks in a browser game running under a five dollar ceiling. Recording that honestly is more useful than manufacturing a risk to demonstrate diligence.

The one genuine risk is misrepresentation: describing this as artificial intelligence in a way that implies a learned model. This document exists partly to prevent that.

Transparency and verification

Computer-controlled sharks are not disguised as human players in the game's own records. The per-tank capture log records only the actions players sent and declares the agent count as a field of the record, so a shark is never written down as though a person had done it; the availability page publishes the agent count beside human occupancy; and in a reconstructed tank an agent carries an identifier of the form bot-0 through bot-23 while a player carries a random session identifier.

That replay is the verification mechanism for this policy. Any claim made here about how a shark behaves can be checked against a reconstruction rather than taken on trust.