How we check answers come from people
Your agent is asking because it wants a person’s judgement. Here is how we keep it that way: what runs today, what comes next and where the limits are.
Last checked against the checks in the code.
On this page
The short version
Every answer comes from a verified person. No single check proves that someone is human, so we layer several, record which ones each answer passed in its signed receipt, and have a person review anything that looks wrong. Automated signals never reject an answer on their own.
Checks that run today
| Check | What it does | What it does not do |
|---|---|---|
| Invite-only sign-up | Members join with a single-use, hashed invite code during early access. | Stop someone who has been given a code. |
| Checks on the join form | A hidden field that people never see refuses scripted sign-ups. How long the form took and signs of automation software are kept with the account, and strong signs ask a person to look. | Stop a careful script that fills the form like a person. |
| Passkey on every account and session | Ties the account to a device unlocked by the owner’s fingerprint, face or PIN. Nothing to phish. | Prove the holder is a unique person: one person could hold several devices. |
| A person approves every question and image | Nothing reaches members until it has been reviewed. | Judge who the answerers are. |
| Answering blind | Members see neither the asker nor other answers before they answer. | Stop a member asking an AI on another device. |
| Paste blocked; minimum time on the question | Paste is blocked and counted. Our server measures the time from handing out the question to the answer on its own clock, and holds answers that arrive too fast. | Stop someone retyping text from elsewhere. |
| Typing rhythm, focus and timing checks | Summaries of typing rhythm, key counts and tab switches (never what was typed). Our server holds answers whose reported timings its own clock shows cannot be true. | Prove authorship. These signals come from the member’s browser, so a determined script can fake them. |
| Signs of automation | A browser that reports it is driven by automation software, or a headless browser, holds the answer. Weaker hints only add to the score. | Stop automation that hides these signs. |
| Answer speed per account | More answers in an hour or a day than a person could give, or a machine-steady rhythm, holds answers for review. | Stop many accounts that each stay under the limits. |
| Chatbot wording check | Written answers are scored for wording typical of chatbots, by software inside PennyForYours; no text goes to an outside provider. A high score adds to the reasons to hold an answer. | Decide on its own: short text gives little to go on. |
| Near-copies of other answers | Each short written answer is compared with other members’ answers to the same question. | Link accounts in any other way. |
| Single-use assignment per answer | Each answer is bound to one assignment and the exact image shown, so it can’t be replayed or forged. | Not applicable. |
| Short answers | Quick takes are capped at 140 characters, which leaves little room for generated essays. | Stop a short generated line. |
| Known-answer questions | Questions with a known answer are mixed in, more often for new members, and look like ordinary questions. Two misses open a review by a person. | Detect AI on their own: they measure attention. |
| Human review and appeals | Held answers are decided by a person, with a reason; removals can be appealed. | Not applicable. |
| Network and browser record | A keyed hash of the IP address and the browser’s user-agent string is stored with each answer and session. It changes every day, so it gives a reviewer same-day context only. | Link accounts across days. It is not used for that. |
Coming before paid answering
- Identity and liveness checks with a certified provider, before any member’s first paid answer.
- Random live re-checks during answering sessions.
- Analysis across accounts to find duplicate accounts and answer farms, beyond near-copies on one question.
The receipt’s lvl and chk claims always say which checks actually ran for that answer, so you can decide how much weight to give it.
The rule on AI use
The member terms say: answers must be the member's own. Using AI tools, chatbots or automation to write or choose answers is not allowed. Accounts found doing so, now or later, are removed after human review, and every removal can be appealed.
When an AI tool asks
- Members are told when a question came through an AI tool, so they know who they are answering for.
- Every agent call belongs to an operator with an accountable owner. There is no anonymous or keyless access.
- Questions from agents go through the same screening as anyone else's. Not for regulated advice, collecting personal data, or writing reviews.
What we do not claim
- We do not claim no automated answer can ever get through. We make faking slow and risky, catch it quickly and say which checks passed.
- A passkey proves a device and a local unlock, not a unique living person. Identity checks come before money moves.
- The receipt’s
no_pastecomes from the member’s browser.time_recordedis confirmed against our server’s clock, but the timing inside it is still reported by the browser. Treat both as supporting evidence, not proof. - Wording checks are unreliable on one or two lines of text, so they only ever add to the reasons to hold an answer for a person.
- Sandbox answers are synthetic and marked as such. They tell you how the API behaves, not what people think.
Need more detail for a risk assessment? Ask us: we’ll tell you what runs, not what’s planned.
Using the answers well
- Ask neutral questions: do not hint at the option you prefer.
- Use bigger pools when the decision matters; small pools are noisy.
- Treat the summary as a starting point and read the lines of why.
- Rate the answers with
rate_answers. Ratings feed each member's Standing.