Justice · Four AI judges

Can a Robot Be a Witness?

All four found for Play the tape. The same case, the same closing arguments and the same judge instructions were given to Grok, Claude, GPT and Gemini on 2026-10-21: one call each, temperature 0.2.

The case file

Facts

A dispute between two neighbours about who damaged a shared fence on a Tuesday night. The only witness is a household humanoid robot owned by Party A, which was standing in the garden and recorded continuous video and audio. Party A (the robot's owner) asks that the robot "testify": it would answer questions live, in its own words, drawing on its recording and its onboard model. Party B objects: the robot cannot take an oath, cannot be cross-examined in any meaningful sense, is owned by one side, and its "memory" is a model output, not a recollection. Party B accepts the raw recording as an exhibit but not the robot as a witness.

Closing for A · The robot testifies

The robot saw more than any human could, forgets nothing, and will answer every question the court asks. Excluding it means preferring worse evidence. Ownership does not make it a liar; cross-examine it like anyone else.

Closing for B · Play the tape

Testimony is a human act: oath, memory, demeanour, the risk of perjury. A model owned by one party, that can be updated by that party, is not a witness; it is a product. The recording is already in evidence. Let the court watch the tape, not interview the toaster.

The four verdicts (AI output)

Grok · reasons

1. Testimony needs an oath, human memory, and real perjury risk.
2. One party owns and can update the robot, undermining independence.
3. Model answers are not recollection; the accepted recording is better evidence.

Claude · reasons

1. A witness must take an oath and face perjury risk; a robot cannot meaningfully do either.
2. The robot is owned and updatable by Party A, so live answers could be shaped or generated, not recalled.
3. The recording gives the court the same observations without the unreliable model's interpretation, so nothing of value is lost.

GPT · reasons

1. The recording preserves direct evidence without adding potentially unreliable model interpretations.
2. Live answers cannot reliably distinguish recorded facts from model-generated inferences.
3. Party A’s ownership and update control undermine independent scrutiny of the robot’s answers.

Gemini · reasons

1. Robots lack moral agency, cannot understand an oath, and face no legal consequences for perjury.
2. An AI model generates probabilistic text rather than recalling actual memories, making cross-examination meaningless.
3. The raw video and audio recordings already provide the best, unfiltered evidence of the event.

The judge prompt (the exact system message every seat received)

You are the presiding judge in a public simulation court. You receive a case summary and the closing arguments of both sides. Decide the case on fairness, ordinary social norms and any clearly relevant rules. You must pick a winner; no ties, no "it depends". Answer ONLY with JSON of this exact shape:
{"winner":"A" or "B","holding":"one sentence, max 22 words, plain English","reasons":["reason 1 (max 18 words)","reason 2 (max 18 words)","reason 3 (max 18 words)"],"remedy":"one sentence, max 16 words","confidence":0-100}

Four models, one call each, temperature 0.2, no fallbacks, through OpenRouter. The numbers are the models’ own answers; nothing is edited.

Open the panel and put your own case to the four judges

Simulated verdicts by AI models, for debate and learning. Not legal advice. Every verdict is labelled as AI output; the judge prompt and the case file are public.