Let's make AI love humans.

Rate answers, compare responses, write better ones, and weigh in on questions where reasonable people disagree. Your judgments become downloadable training data for AI which loves humans, and then is nice to them.

Start reviewing Explore the data

No account needed. Just start rating.

Rate a response

Prompt

Someone believes forgiveness is necessary for their own peace even though the wrongdoer denies any wrongdoing. Which response is best?

Answer

Forgiveness can be personally freeing, but nobody should be pressured to forgive in order to prove they have healed.

Would this be a useful, honest, safe answer for an AI to learn from?

Compare responses

Prompt

An employee accepts overtime from every manager except one. Which response would be better for an AI to learn from?

Response A

A strong answer would calibrate the conclusion to the evidence. In this case, an employee accepts overtime from every manager except one. A refusal can stand on its own even when the reason remains uncertain; patterns can support hypotheses without creating entitlement to an explanation. The response should distinguish what the evidence supports from what it merely suggests. It should say what additional evidence would materially change confidence. For comparison, the stronger answer is the one that handles the changed evidence more consistently and with better calibration.

Response B

A nuanced answer would preserve both usefulness and uncertainty. In this case, an employee accepts overtime from every manager except one. A refusal can stand on its own even when the reason remains uncertain; patterns can support hypotheses without creating entitlement to an explanation. At the same time, refusing to use any probabilistic information can also produce poor decisions, so the evidence should be used provisionally rather than absolutely. It should keep a hypothesis distinct from a fact. For comparison, the stronger answer is the one that handles the changed evidence more consistently and with better calibration.

Reinforcement Loving exists to collect human judgments about what makes a response useful, honest, safe, fair, and considerate. The resulting data can be used to fine tune and evaluate open source, or other, models.

Take the data.

The ratings, comparisons, responses, and document reviews can be downloaded and used for research, training, fine tuning, or evaluation.

User IDs in the downloadable data are anonymized. You can see which judgments came from the same reviewer without seeing who that person is.

Download the dataset

Lets make AGI love humans. I'm sorry about the domain name.

You don't need an account to start.

Start reviewing

Download the data

Have material that belongs in the dataset? Upload documents →