Topic index
Every question about checking AI output, answered on its own page.
Each page takes one term people actually search for and answers it properly: what the check does, what it cannot do, and who signs off when the answer has to be attributable.
Take it apart before you sign it off.
Core concepts
- AI QATeams usually start with a scanner and discover six months later that nobody agreed what a failure means. QA for generated work is mostly agreement: which defects matter, who decides, and what happens when the answer is no.Read
- AI output checkerEvery checker produces a verdict. The useful question is whether the verdict can be inspected: which rule fired, how bad it is, what ran, and what the tool never looked at.Read
- AI quality gateDashboards measure what happened. A gate decides what is allowed to happen next. The difference is a threshold, an owner and a consequence, and without all three you have reporting rather than control.Read
- AI content verificationGenerated copy is not usually badly written. It is confidently unsourced, which is worse, because a fluent paragraph with an invented statistic reads exactly like a fluent paragraph with a real one.Read
- AI validation toolValidation tools are sold on outcomes and bought on trust. The gap between those two things is closed by a short, repeatable evaluation you can run before the pilot ends.Read
Human layer
- Human-in-the-loop AI reviewPutting a person in the loop does not automatically add safety. If the person has thirty seconds, no context and no authority, the loop is a signature that launders a machine's output.Read
- Human in the loop AIMost review processes fail at the same point: the person approving has no time, no evidence and no incentive to say no. The step exists on the diagram and nowhere else.Read
- Named human sign-offApprovals without names are organisational noise. A signature is useful because it attaches a specific person, a specific scope and a specific date to a claim about the work.Read
- Auditable AI reviewAudit is not a test of whether you had a process. It is a test of whether the process left traces detailed enough for a stranger to follow without you in the room.Read
Offline and private
- Offline AI QAOffline QA is not a privacy badge bolted onto a scanner. It is a different architecture: the rule engine ships to your browser once, and every check runs against the text in the tab, so what you paste is never in a request body.Read
- Zero-upload AI toolMost review tools ask you to upload the thing you least want to upload. A zero-upload checker does not, and that difference shows up in procurement, in the DPA and in how quickly you can look at a draft at all.Read
- Browser-based AI checkerThe difference between a checker that lives in a tab and one that lives on someone's server is not cosmetic. It decides who holds your draft, how fast the answer arrives, and whether the same input can be counted on to produce the same result.Read
- Client-side code checkHalf the code that gets reviewed now arrives from a prompt, a snippet in a chat window, or a prototype with no repository behind it. A check that needs repo access, a token and a pipeline cannot help with any of that.Read
- Privacy-first AI QAEvery quality process requires someone to look at the work. The privacy question is not whether that happens, but where the copy lives, who can read it, and how long it exists afterwards.Read
- Independent AI verificationWhen the tool that produced the work also grades it, agreement is the expected outcome. Independence is what turns a second look into a second opinion, and it is a property you can check rather than a claim you have to trust.Read
- Local AI verificationEvery vendor claims privacy. The difference between them is where the bytes actually go, and that is a question you can answer with a network tab, a contract and one test.Read
- Check writing without uploadingLegal drafts, unreleased pricing, a client's brand strategy: all of it goes through the same review process as a blog post, and none of it should enter a tool that stores what it receives.Read
Code checkers
- Leaked API key scannerEvery other finding in a code review can wait a sprint. A credential in source is a live incident with a clock on it, and generated code puts them in places nobody thinks to look.Read
- AI code checklistGenerated code reads well enough that review becomes a skim. A checklist puts the questions in a fixed order so a fluent diff cannot talk you past them.Read
- Static analysis for AI codeGenerated code usually looks tidy, so the defects worth catching are structural: a credential in source, an unescaped string, a swallowed error. All of those are visible in the text, before anything executes.Read
- AI bug detectionIt is not random. Assistant-written code fails in repeatable ways: the happy path only, the invented method, the swallowed error, the credential in the example that was supposed to be illustrative.Read
- Review AI codeGenerated code is fluent in a way human code often is not, and fluency is what defeats review. The method below replaces reading with a sequence of specific questions, asked in a fixed order.Read
Copy checkers
- AI tell-list checkerThe tell-list is not a fingerprint. It is a list of habits that appear when text is produced by pattern rather than by someone deciding what to say, and habits are fixable in a way that authorship is not.Read
- AI phrasing detectorThe useful question about a generated draft is not whether a model wrote it. It is which sentences were assembled rather than written, because those are the sentences a reader skips.Read
- AI copy editorMost editing time goes to the checks that could be mechanical: placeholders left in, statistics with no source, sentences that all run the same length. Automate those and the human pass gets to do the part that needs taste.Read
- AI writing quality checkQuality is not one number. It is a small set of questions asked consistently: does this claim hold, is it structured for the reader, does it sound like us, and can anyone check it.Read
- AI readability auditReadability scores have a bad reputation because they were used as targets. Used as diagnostics, they point at exactly the paragraphs a reader will bail on.Read
- Unsourced statistics checkerNumbers survive review better than claims do, because a percentage reads like evidence. It is the cheapest thing for a model to produce and the most expensive thing for a brand to retract.Read
- AI content editing checklistOrder matters. Fixing sentences before facts wastes the work, because the best sentence on the page is often the claim you have to cut.Read
AI detection
- AI content detectorDetectors are asked who wrote a document and answer with a description of the prose. That mismatch is the source of almost every argument about them, and understanding it is how you use one without doing damage.Read
- False positive AI detectionA detector that is right 95% of the time will still be wrong about most of the people it accuses, when most of the people are innocent.Read
Humanising AI copy
- How to humanize AI textHumanizing is not a synonym pass. It is the work of turning sentences that describe a topic into sentences that state something a reader can check, and it is measurable.Read
- AI humanizerEvery humanizer on the market does some version of rewriting. The differences that matter are what it preserves, what it silently alters, and whether you would notice before publishing.Read
AI writing tells
- Overused AI wordsThe list is long, but the fix is not a thesaurus. Most overused words are placeholders for a fact the draft never had, so deleting them often beats replacing them.Read
- ChatGPT buzzwordsA reader who has spent a year with language models can spot a generated paragraph in a sentence. It is rarely one word; it is the arrangement, and every part of it is editable.Read
- Robotic AI writingSwap every stock word for a synonym and the page still sounds like a manual, because the tell is not in the vocabulary. It is in the metronome.Read
- Repetitive AI contentA model does not remember what it already told you. It reconstructs the argument each time, which is why the same point arrives three times in slightly different clothes.Read
- AI fluffNobody defends padding. They defend the sentence it sits in, and that is exactly how a page ends up 1,600 words long with 700 words of content.Read
- Generic AI writingIf your company name could be replaced with a competitor's and the page would still be true, you have written a page about your category rather than about your product.Read
- AI sounding textNo reader runs a classifier. They hear a formula they have seen a thousand times this year, and they stop trusting the page before they finish the sentence.Read
Hallucination and accuracy
- AI hallucinationsA hallucination is not a glitch. It is the normal output of a system that produces likely text, and once you accept that, the detection problem becomes an ordinary research problem with a routine answer.Read
- Stop AI making things upYou cannot instruct a model into accuracy. You can change the task so invention is harder, and make the remaining inventions cheap to find.Read
- AI factual errorsDrafts rarely fail by being entirely wrong. They fail by being right about the shape of a thing and wrong about the detail, which is why a reader who skims will not catch it.Read
- Outdated AI informationA model cannot tell you its knowledge has expired, so it writes about last year's prices and last release's API in the present tense, with the same certainty it uses for anything else.Read
Originality and attribution
- AI plagiarismThe phrase covers a legal question about ownership, an academic question about attribution and a craft question about originality. Answering the wrong one is how people end up confidently wrong.Read
- Lack of originalityMost pages are unoriginal because they contain nothing that required the writer to be there. That is fixable in an afternoon, without inventing anything.Read
Trust and search
- E-E-A-T and AI contentExperience, expertise, authoritativeness and trust are how quality raters describe a good page. None of them can be added to a page after the fact, and all of them are visible when they are missing.Read
- Does Google penalize AI contentThe question assumes authorship is the variable. In the published guidance it never has been: usefulness, originality and reliability are, and a model can be on either side of all three.Read
- Google penalty for AI contentThe word penalty implies a person read your pages and applied a rule. Most sites never get that far. They are simply ranked lower, because the pages stopped being the best answer.Read
- Thin AI contentA 2,000-word page can be thin and a 300-word page can be substantial. Length is the least useful measure of the five available to you.Read
- Scaled content abuseThe policy does not care whether a model, a freelancer or a script produced the pages. It cares that there are thousands of them and that each one exists to occupy a query.Read
- AI bias in writingGenerated prose is built from the most common things written about a subject. The most common things written are not the most representative, and the gap shows up as who a page assumes its reader is.Read
Document and web
- CSV validation toolMost broken imports are boring: one row with an extra comma, a header cell nobody named, a date written two ways. All of it is visible before the file reaches the loader.Read
- Markdown linterGenerated documentation usually parses. What it fails at is shape: two level-one headings, a section repeated under a synonym, a link that points at a hash, an image with no description.Read
- Broken anchor checkerExternal links get audited. In-page anchors do not, which is why a rewritten heading quietly breaks every deep link, table-of-contents entry and support answer that pointed at it.Read
- HTML markup checkerA generated page usually renders. What it fails at is the invisible contract: valid nesting, described images, a declared language, working links, and headings a screen reader can navigate.Read
- Missing alt text scannerAlt text is the accessibility failure with the widest reach and the simplest detection rule, which is why it is also the one most often fixed badly: a scanner passes, a reader is still lost.Read
- Website metadata auditThe title, the description, the social image and the canonical URL decide whether a page is ever read. They are also the fields most likely to be generated once and never revisited.Read
- Broken link checkerA 404, a fragment that resolves to nothing and an href of # all look identical in a report and need different fixes, which is why link audits produce so much noise and so little repair.Read
- Document QA workflowDocuments fail slowly. A missing heading breaks navigation, an unsourced figure survives three reviews, a superseded version keeps circulating. Each has a cheap pass that catches it, and the order is what keeps the pass cheap.Read
Image and provenance
- AI image quality checkMost image problems are not aesthetic. They are a two-megabyte hero image, a decode failure, a flattened contrast range or provenance metadata that was stripped somewhere between generation and upload.Read
- C2PA metadata checkerC2PA is a signed manifest embedded in a file, listing how an asset was made and edited. Most checkers can tell you one is present. Very few can tell you whether the signer is anyone you trust.Read
- Image provenance verificationProvenance is not a yes-or-no property of a file. It is a set of claims, each with its own strength, and the useful work is knowing which claim you actually have before you repeat it.Read
- Content Credentials validatorContent Credentials made provenance visible to normal people, and the visibility created a new misunderstanding: that a label on an image is a verified fact about the image.Read
- Check C2PA offlineProvenance checking is time-sensitive. Every hop through an editor, a CMS or a chat window can strip the evidence, so the useful check happens on the file as it arrives and again after each export.Read
- AI art handsFive fingers in an unusual pose, partly hidden, at three hundred pixels wide. Every part of that sentence is a reason a generator will get it wrong.Read
- AI image artifactsEvery generator leaves evidence of the parts it was guessing: the corner of a room, the far side of a table, the space behind a head. Learn where it guesses and you can inspect any image in a minute.Read
- AI text rendering issuesThe output looks like writing at a glance and like nothing at all when you read it. That gap is the most expensive defect in generated commercial imagery.Read
- Uncanny valley AI videoA single frame can be convincing. Two seconds of it moving shows whether the scene was simulated or merely described.Read
- Robotic AI voiceSynthetic speech fails less on timbre than on rhythm: every sentence shaped the same way, every word given equal care, nothing ever interrupted by being alive.Read
Workflows
- Verify before publishingMost publishing failures are not subtle. A number nobody can source, a placeholder that shipped, a claim that outran its evidence. A fixed sequence catches them before a reader does.Read
- Human review for AI contentThe difference between a useful human review and an expensive rubber stamp is decided before the file is sent: in the scope, the brief and the agreement about what comes back.Read
The services behind these pages
Topic index
The fastest way to use any of this is a check.
Paste code, copy, a document or an image into the console. The rules run in your browser and nothing is uploaded.