How sides works

Two AIs answer the same real question. You pick the better answer without knowing which model wrote which, and every pick pays. In aggregate, the picks show which AI people actually prefer. This page explains the whole thing, plainly, including the parts that are not live yet.

What sides is

A blind taste test for AI answers, where your taste is the whole point.

sides is one simple loop. We show you a real everyday question and two answers to it, each written by a different AI model. The answers carry no labels. You do not know which model wrote side A and which wrote side B. You read both and pick the one you like more. That is the entire job.

You can start judging with no account at all. Open the app, read, pick, and go. Signing in is optional and only adds convenience, not access.

Every pick pays you. Picks accrue a token called $SIDES that you can claim later. One pick on its own is just your opinion. Thousands of picks together are something else: a clear, honest signal of which AI answers real people reach for when the brand name is hidden.

This measures preference, not trust.

We are not asking which model you believe, or which one you think is more accurate. We are asking which answer you liked more, right now, with the labels off. The signal is genuine human preference. That is a different and more honest thing than a trust score, and it is exactly what is hard to buy anywhere else.

How a round works

One question, two answers, one pick. Here is the full loop, start to finish.

1
A real question is drawn
Each round starts from a question in our corpus of real, everyday prompts. Not trick questions or benchmark puzzles. The kind of thing people actually ask an AI.
2
Two models are selected
Two different models are chosen to answer that same question. They each produce one answer to the exact same prompt, so the only thing that varies is the model behind the answer.
3
Sides are assigned at random
For your round, the two answers are shuffled into side A (blue) and side B (red) at random. Which model lands on which side is decided per round, so position tells you nothing.
4
You pick, and the pick locks
The instant you choose a side, that pick is locked. There is no take-back. Locking on the first choice keeps the signal clean: it captures your honest first read, not a second-guess after the reveal.
5
The reveal happens
Right after you lock, the names come off both answers and you see which model you picked and which you passed on. The reveal only ever comes after the pick, never before.
6
The pick is recorded
We record the question, the two models, which side each was on, and which one you chose. That row is what later rolls up into the preference signal.

Then it repeats with a fresh question and a fresh pair. The pick is always locked before the reveal, and the sides are always reshuffled. That order is what makes each pick worth recording.

Blind is not a gimmick. It is the one thing that makes a pick mean anything.

The moment you know a name, the name does the picking. People carry a brand halo. If a label said one answer came from the model you already like, you would lean toward it before reading a word. Hiding the names strips that halo away, so your pick is about the answer in front of you and nothing else.

Sides are randomized every round. Which model is side A and which is side B is decided fresh each time, so a habit of always tapping the left, or always tapping blue, averages out instead of favoring one model. Position bias and model bias both wash out across many rounds because the mapping keeps changing underneath them.

Empty or clearly broken answers are handled rather than scored. If a model returns nothing usable, or an answer is obviously degenerate, that round is not a fair contest, so it does not get treated as a real preference signal. We would rather drop a bad round than let it pollute the map.

What this does not claim.

Taste is subjective. A pick tells us which answer you preferred, not which answer was correct. sides measures preference, not accuracy, and we do not dress a pile of opinions up as a statistical proof. The value is in honest, blind, aggregate preference, described as exactly that and nothing grander.

How preference becomes data

One pick is an opinion. Many picks, at the model-pair level, become a signal.

Every recorded pick is a small head-to-head result: on this question, side A won or side B won, and we know which model was which. Roll those results up across many questions and many judges and a pattern appears at the model-pair level. When two given models meet, one of them gets chosen more often. That rate is the preference signal.

We call the full picture a preference map: for each pair of models we have run against each other, how often blind human judges preferred one over the other. It is built entirely from real picks, with the labels hidden at the moment of choosing.

What the map claims is narrow and honest. It says which answers people preferred, blind, in aggregate. It does not claim which model is more accurate, more truthful, or better for any specific task. It is a preference reading, not a correctness verdict.

This is forward-looking. The preference map is what makes sides valuable to people who build AI, and building it is the point of the whole product. We do not currently claim any lab customers, and we will not invent one. When there is something real to report on that front, it will be reported plainly.

Earnings and $SIDES

Every pick pays. Here is exactly how, including the parts that are not open yet.

$SIDES is the token you earn for judging. Each pick you make adds to your $SIDES balance. The model is accrue-then-claim: you build up a balance as you judge, and you claim it out when you decide to, rather than getting a separate payout per pick.

Your balance only ever shows what you have actually earned. If you have judged but payouts are not open, your balance reads zero rather than a made-up number. We do not display a balance you cannot stand behind.

Claiming is not live yet.

Picks accrue now, and payouts open soon. Until claiming is live, keep judging to build your balance. When claiming and any threshold or claim details are finalized, this section will state them plainly. We are not going to pre-announce numbers we have not settled.

If you are signed in, your balance follows your account across devices. If you judge as a guest, your picks still count toward the aggregate, and signing in later is how you attach a persistent balance to your history.

Accounts and privacy

Anonymous first. Sign in only if you want your history and earnings to follow you.

With no account

You can open the app and start judging immediately. Read, pick, see the reveal, repeat. Nothing personal is required to judge. Your picks count toward the aggregate preference map just the same.

Signed in with Google

Signing in with Google saves your judging history and keeps your earnings balance persistent across devices. That is what an account adds: continuity. It does not unlock the judging itself, which is open to everyone.

What we store is the judging, not you. For a pick we keep the question, the two models, which side each was on, and which one you chose. If you are signed in, those picks are tied to your account so your history and balance persist. If you sign in with Google, we use it to identify your account, nothing more exotic than that.

To judge, we need nothing personal from you. No profile, no essay about yourself, no contact details. The one thing every pick does is feed the aggregate map, because that shared signal is the reason sides exists. That is the deal, stated straight: you judge, you get paid, and your blind picks join everyone else's to show what people prefer.

For AI labs and researchers

Blind human preference, measured directly, is the thing sides produces.

The aggregate preference map is built for the people who train and ship AI models. Most preference data is either self-reported, brand-influenced, or inferred second-hand. sides measures the thing directly: real people, real everyday questions, labels hidden, picking the answer they actually like more. That is preference stripped of the halo, at the model-pair level.

Why measure it this way. When judges cannot see the brand, their picks stop being a vote for a reputation and start being a read on the answer. That is harder to game and closer to how a model will land with real users than a leaderboard people already have opinions about.

If you build models and want to talk about the preference data, reach out at labs@sides.money. We are early and honest about it. No case studies to point to yet, no partner logos to show. When there is real work to describe, we will describe it here and not before.

FAQ

Straight answers. We update these as the product firms up.

Do I need an account to judge?

No. You can start judging with no account at all. Signing in with Google is optional and only saves your history and earnings across devices.

How do I know it is fair?

The two answers are shown with no labels, and which model sits on side A or side B is randomized every round. Your pick locks before the reveal. So a pick reflects the answer, not the brand and not the position.

Can I change my pick after the reveal?

No. The pick locks the instant you choose, which is before the reveal. That is deliberate. Locking on the first read is what keeps the signal honest.

Which models are in the pool?

Each round pairs two different models on the same question. We rotate which models appear, and we do not publish a fixed roster here, partly because it changes. What stays constant is that the answers are unlabeled when you pick.

What exactly does a pick record?

The question, the two models, which side each was on, and which one you chose. That row is what rolls up into the aggregate preference map.

Is this measuring which AI is best?

No. It measures preference, not correctness. A pick tells us which answer people liked more, blind. It does not certify accuracy, and we do not present it as if it does.

How do I actually get paid?

Every pick accrues $SIDES to your balance, and you claim the balance out. Claiming is not open yet. Payouts open soon, and when claim details and any threshold are settled, this page will state them plainly.

Why is my balance zero after judging?

Because your balance only shows what you have truly earned and can stand behind. Until payouts open, it stays at zero rather than showing an invented figure. Your picks are still counting.

What happens to my data?

We store the judging, not personal details about you. Picks feed the aggregate preference map, which is the reason the product exists. Signing in ties your picks to your account so your history and balance persist.

What if an answer is empty or broken?

Rounds where a model returns nothing usable, or an obviously degenerate answer, are handled rather than counted as a real preference result. A broken round is not a fair contest.