Cipher checks
An in-depth look at the checks Cipher runs and what each one detects.
Cipher runs about fifty individual checks across three levels: Basic, Pro and Max. Each check looks at one kind of evidence and contributes to the overall risk score. Together they cover bots and AI-written answers, and also careless, ill-intent and wrong-fit respondents. This page explains what every check detects and, at a high level, the signal it looks at.
The exact thresholds, weightings, and models behind these checks are part of how Cipher stays hard to game, so they are not published here. What follows is the intent of each check.
How levels stack
A level includes every check from the levels below it. Basic is fast, free and runs on every response on every plan. Pro and Max add deeper behavioral, writing, network, and cross-response analysis, and cost credits per response. Every flag comes with plain-language reasoning at every level.
Basic: speed, care, attention and automation
Fast checks on every plan, including Free, at no charge. They catch the most common careless, low-effort and automated responses.
- Rapid completion. Detects responses finished faster than a person could plausibly read and answer.
- Uniform timing. Flags timing across questions that doesn't look like a person reading.
- Low interaction. Catches responses with too little interaction behind them to have been filled in by hand.
- Straight-line answers. Detects grids and scales answered without reading the items.
- Careless rating patterns. Flags long runs, speeded ratings and answer patterns that break from how other respondents answer.
- Impossibly fast. A stricter speed check for input faster than a person can produce.
- Minimal effort. Flags written answers that put no real thought in.
- Attention check. Grades instructed-response questions against their expected answer.
- Hidden trap field. Flags answers to a field people cannot see.
- Excessive paste. Detects answers brought in from elsewhere rather than written in the survey.
- Pointer velocity spikes. Flags pointer movement a hand does not produce.
- WebDriver detected. Detects browsers driven by automation frameworks.
- Automation detected. A broader check for scripted environments.
- Missing plugins. Detects browsers that don't look like ones a person uses.
- Suspicious user agent. Flags browsers that identify themselves as known bots or in ways real ones don't.
- Device mismatch. Detects devices whose reported details don't fit together, a sign of a spoofed device.
- Screen anomaly. Flags screens no real device has.
- Suspicious pauses. Detects pauses that look automated rather than human.
- Behavioral model. Weighs many interaction features together when raw pointer and keystroke data is available.
Pro: how each answer was written
On the Hive plan. Pro adds how each written answer was produced and how the person moved through the survey. Each response uses Cipher credits.
- Writing process. Reads how each written answer was produced: typed, revised, pasted or copied out, how fast, and when.
- Text forensics. Finds hidden characters and assistant-style habits in written answers.
- Robotic typing. Detects typing that a machine produced.
- No corrections. Detects long written answers entered in a way people don't write.
- Mouse teleporting. Detects pointer movement that no hand could make.
- Excessive tab switching. Flags repeatedly leaving the survey when the survey didn't ask for it.
- Window focus loss. Detects long stretches away from the survey when the survey didn't ask for it.
- Hover behavior. Checks that the pointer behaves around answers the way a reading person's does.
- Scroll patterns. Detects scrolling that doesn't look like a person reading.
- Mouse acceleration. Separates human pointer movement from synthetic movement.
- Pointer targeting. Judges each click and tap against the size and position of what was pressed, so the result holds for any layout or theme.
- Sentence rhythm. How much sentence length varies in a written answer. One supporting signal, never a verdict on its own.
- Contradiction (light). Catches answers that directly conflict with each other within the same response.
- Copied text. Spots written answers that read as lifted from somewhere else.
- AI-written answers (light). Combines how an answer arrived with a light AI review of its wording.
- Quality assessment. A light AI review of whether open answers are relevant, specific, and on topic.
- Semantic analysis. Looks at whether answers are meaningful and coherent rather than filler that only looks valid.
Max: across responses, network and the web
On the Swarm plan, and on Enterprise. Max adds checks that compare responses with each other, network checks, a thorough AI review and a web search.
- Fraud ring detection. Identifies clusters of responses that appear to come from the same coordinated source.
- Answer sharing. Detects responses with near-identical answers, a sign of shared answer keys.
- Coordinated timing. Flags groups of responses submitted in synchronized bursts.
- Device sharing. Detects many responses tied to the same underlying device.
- Baseline deviation. Compares a response against the normal pattern for the survey to spot outliers.
- VPN detection. Flags connections coming through common VPN services.
- Datacenter IP. Flags traffic originating from servers and hosting providers rather than residential networks.
- Tor detection. Flags connections arriving over the Tor network.
- Proxy detection. Flags anonymizing proxies used to mask origin.
- Timezone validation. Checks whether the device timezone is consistent with the network location.
- AI-written answers (thorough). A thorough AI review on top of the writing-process evidence.
- Contradiction (thorough). Cross-checks the whole response for subtle logical inconsistencies.
- Web source check. Searches the web for long written answers that match text already published online. Questions that ask people to quote, cite sources, or paste text are skipped.
On hosted surveys, Max also weighs a respondent's reputation across your own surveys.
How Cipher spots AI-written answers
Cipher catches AI-written answers from human behavior. It looks at how each answer originated and where it came from, whether it was typed, revised or pasted, how quickly it arrived, and when. Wording is read on top of that evidence, never on its own.
Mobile and touch checks
On phones and tablets the mouse and keyboard checks above cannot fire, so Cipher runs an equivalent set against touch-native signals. See Cipher on phones and tablets for the full picture.
- Rigid device. Detects emulators and devices that aren't held by a person.
- Synthetic touch. Flags touch input produced by software rather than a finger.
- No touch on a touch device. Catches a device that presents as a phone or tablet yet produces no touch input while answering.
- Uniform taps. Detects taps that no finger would produce.
- Impossible swipe. Flags swipes a hand could not make.
- Paste without a keyboard. Detects long answers brought in from elsewhere, including on platforms that hide pasting.
These checks carry a coverage profile so that a response with thin touch evidence is scored conservatively rather than waved through.
Survey-aware scoring
On surveys hosted by Surbee, Cipher reads the survey before scoring a response. Leaving the page on a question that sends people elsewhere, and pasting into a question that asks for a paste, are expected behavior there, so the tab-switching, focus, pause and paste checks set them aside. Creators can set these expectations per question, and Cipher also learns them from earlier respondents to the same survey.
The SDK doesn't send this context yet, and its tracker doesn't record pointer targeting, so these two features only apply to hosted surveys for now.
Beyond the named checks
Underneath these checks, Cipher extracts many more features from each response, covering behavior, device, network, content and the survey's own quality questions. The named checks above combine those features into the signals you see in the dashboard and in the validation API response.