How AreI answers a question

Step by step

1. It decides how many sources the question needs

Every question goes to at least two AI models from different makers: one source cannot disagree with anything, so it cannot be checked. AreI asks a third when the question looks time-sensitive, high-stakes (health, law, money) or very short.

2. Independent models answer

Each model answers without seeing the others, and lists the individual statements its answer rests on. Today the models come from OpenAI gpt-oss, DeepSeek V4, Google Gemma. “Independent” means different model families: two copies of the same model would be one source asked twice, and AreI counts them as one.

3. AreI compares the statements

It groups statements that say the same thing and flags the ones that may conflict — different numbers for the same quantity, opposite claims, incompatible values. This comparison is mechanical and it can be wrong in both directions: it can flag two statements that don't really conflict, and miss two that do. So every flagged group is shown with each model's own words.

4. It tests the conflicts that matter

When a flagged conflict would change the answer, another model is asked to examine it. Before a flagged group is set aside as compatible, a reviewer from a different model family has to agree. What a reviewing model says is shown as its quotation, never as AreI's finding.

5. It builds the answer from what held up

Each statement is accepted, accepted with conditions, set aside, or marked as lacking evidence. The answer is assembled only from statements that survived. AreI chooses their order; it does not rewrite them into new claims.

6. It labels the result

The label is derived from what happened above, never chosen to sound confident:

What an answer page shows

• The label and the answer, first, before anything else.
• Where the sources' wording differed, with each model's own words.
• Conflicts and possible conflicts, split into those AreI proved and those it only flagged, and how each was resolved or why it could not be.
• What is still uncertain and the limitations of that particular answer.
• A receipt: which models were called, how many calls, tokens, time and cost.
• Every claim, with whether it was accepted and why.

What AreI will not do

• Treat agreement as truth. Two models agreeing and five models agreeing give the same reason for accepting a statement: neither is counted as a vote.
• Write prose that outruns its claims. The answer you read is the accepted statements themselves, joined together, not a fresh summary that could say more than they do.
• Hide a conflict it could not settle. An unsettled conflict that matters keeps the answer from being labelled supported, and it is listed on the page.
• Let a request pick its models or its spending. No one can ask AreI for a particular model or a bigger budget; the server decides both.

What we can and can't claim

We don't claim AreI is more accurate than the strongest single AI model. That hasn't been measured in a way that could show it, so we don't say it.

What AreI is built to do is narrower and checkable on every answer: not present a contested statement as settled, and show you what it couldn't settle. The cost of that is real. AreI declines more often than a chatbot, and today most of its answers come back conditional or insufficient evidence.

It doesn't search the web or check sources yet, so it cannot confirm current facts, and its answers are not medical, legal or financial advice.

For developers

Everything on the answer page is available as JSON. Send a question to POST /api/ask; without a token it uses the same free allowance as the website. With a free account you can create tokens on your account page, each limited to what it may do.

curl https://askarei.com/api/ask \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer <your token>' \
  -d '{"question": "Does SQLite allow concurrent writers?"}'

The response is the full answer record: the label, the answer, every claim, every conflict and the receipt. GET /api/models lists which model families can answer. There is no field to choose a model: that is the design, not an omission.