Frequently Asked Questions

BiasChecker.ai uses AI to analyse text for potential bias, manipulation, logical flaws, and other issues. It offers multiple analysis types including bias detection, manipulation analysis, scientific and legal assessments, historical parallels, and more.

Think of it as a second pair of eyes that helps you read more critically.

When you submit text for analysis, it goes through a multi-stage AI pipeline:

1. Content capture: The text is taken directly from what's displayed in your browser — either the whole page or a specific selection you highlight.
2. Translation note: If you've translated the page using your browser or the website's translation feature, the translated text is what gets analysed, not the original language.
3. AI analysis: The text is evaluated for bias, manipulation, and other issues depending on which analyses you select.

The process typically takes 5–20 seconds, but the time varies with the length of the text, the analysis type, the AI model you choose, and how busy the service is at that time of day. During peak periods — or for long articles on slower or premium models — an analysis can occasionally take a few minutes.

For a fuller picture of how analyses run — the principles every analysis follows and what each analysis type looks for — see our methodology page.

We'd love to hear from you. Please visit our Contact page to get in touch.

Synthesis is an analyst's report on an article, written from the other analyses rather than a fresh read of the text. It gives you a one-sentence verdict on what the piece is doing to its reader, a short narrative connecting the findings into one story about how it works — each section citing the analyses it rests on — an honest "what checks out" paragraph, and a closing note on what to verify before forming a view. Every supporting quote is reused from the underlying analyses and validated by our code, never invented.

Because it synthesizes other results, Synthesis needs the core analyses of the article to exist. If you have already run them, building it costs exactly one synthesis run. If not, it offers to run the missing ones for you and states the run count before anything starts. Analyses that are already available are always reused free, and the quick pre-check that suggests which analyses fit an article is free too.

See the methodology page for how the combined verdict is computed.

Bias analysis examines the article text across 29 bias categories — framing, selection, source imbalance, political and other slants — and surfaces the specific passages that drive each finding.

1. Detection: The AI reads the article and flags biased framing, loaded language, missing perspectives, and one-sided sourcing, citing the exact evidence for each.
2. Scoring: Findings are weighted by severity into an overall concern level (none / low / medium / high) so you can see at a glance how slanted the piece is.

Like every other analysis that reads the article directly (Intent Analysis, Persuasion, Omissions, Critique, Legal Risk, Moral Lens, Scientific Assessment, Background, Historical Parallels, Neutral Rewrite, and Roast — Synthesis is the exception, as it is built from the other analyses), bias analysis runs on the original text.

Two different ratings appear on a bias result, and they measure different things:

• The overall Bias Level (none / low / medium / high) measures this article. Each finding contributes points according to its category's severity — a high-severity category (like framing or omission) counts twice a medium one, and an easy-to-spot category counts half — and the totals are tiered so that a single trivial, easy-to-spot indicator still reads as none, one or two findings read as low, a handful as medium, and only a broad pattern of serious categories as high. Heavy spin shows up as several serious categories at once, not as a long tail of minor issues. We deliberately don't use the AI's own numeric scores: they aren't calibrated across models, so the same article could rate differently depending on which model ran it.
• Each indicator's severity badge (low / medium / high) classifies the type of bias found, not its intensity in this particular article. Every category in our bias catalog is rated by its impact on the reader and by how easily an unaided reader can spot it: high-impact biases rate high, and a bias that is easy to notice rates one step lower — a slant you can see through does less damage than one that works invisibly. Because the rating is a property of the bias type, the same category always carries the same badge, whichever AI model produced the analysis.

Findings also show their cited evidence. Where an excerpt is marked verified, we located it verbatim in the article text — quotes the AI paraphrased or that we could not locate are shown but not counted as verified.

So a one-sided press release can rate high overall with several high-severity categories, while a lightly slanted news report might rate low overall yet still list a framing indicator — one serious category on its own signals a lean, not heavy spin.

The Roast is a purely comedic analysis mode — a warm stand-up bit about the article. It gives you a punchy verdict, a playful retelling, and a few standalone jokes about its most roastable details. Think of it as a friendly, witty companion who finds the funny in the news.

The humour is designed to be warm and good-natured — it targets the writing and framing, never individuals, victims, or people in vulnerable situations. The bar is whether there's genuinely something funny: a dry, well-written piece gets nothing to roast, and tragedies or sensitive human stories (serious illness, grief, disability) are never joked about — and a declined roast costs you nothing.

Comedy is subjective, and some users may find the tone too direct. If that's the case, the other analysis modes surface what's actually weak about an article in a straightforward format.

A single AI model can have blind spots and quirks. Consensus is the answer to that: instead of one model, a panel of models from different AI makers each analyses the article independently, and an editor model then merges their findings into one result. The rule is strict: a finding only counts as agreed when at least two models independently support it, and this is enforced automatically, not left to the AI's judgement. What a single model raised is not thrown away — it is shown separately, under its own heading, with that model named.

• What you get: the findings the panel agrees on, each labelled with how much of the panel backed it ("3 of 4 models"), followed by a clearly separated set of observations raised by one model — models read different parts of an article closely, so one voice there is a narrower reading, not a discredited one. A note at the bottom names the exact models the judgement rests on. If the models agree on nothing, an empty agreed set is the honest outcome: the panel found no issues in common.
• What a "group" is: the maker of the model — Anthropic, OpenAI, Google, DeepSeek, Z.ai (GLM), Moonshot AI (Kimi), xAI — and the panel takes one seat per maker. Two models from the same maker share training data and habits, so counting them as two opinions would overstate the agreement; grouping by maker is what makes "two models agreed" mean two independent readings.
• Who is on the panel: a run seats three makers. Analyses of the article already cached from a maker's model fill that seat free — and every covered maker joins, so a well-analysed article can seat more than three at no extra cost, while a run never buys a fourth. Empty seats are filled by the fixed representatives Claude Haiku 4.5, GPT-5.4 Mini, Gemini 3 Flash and DeepSeek V3.2, in that order; DeepSeek V4 Pro performs the merge and never sits on the panel it judges. This is the current line-up and will change as models change; every result names the models that actually sat on its panel.
• Where to find it: press Consensus beside the model picker. It is a run mode, not a model: pressing it opens a setup card showing the panel seats and what the run costs, and nothing is charged until you run it. It's offered on the analyses whose findings come from a fixed catalogue — Bias, Persuasion, Legal, Moral, Scientific — because counting agreement needs findings the code can match across models. Omissions and Critique produce free-form claims and suggestions, so there is nothing to match; interpretive modes like Intent, Background, Parallels, Rewrite and Roast have no agree/disagree concept.
• What it costs: you're charged for the models that actually run. Panel members whose results are already cached for the page are re-used free, so a consensus on a previously analysed article is often much cheaper — and when three makers are already covered, the only thing left to pay for is the merge. The setup card itemises the seats and the price before you run. Members that do run become available to everyone afterwards as ordinary single-model results.

A consensus never ships on fewer than three models. If a panel member fails on an article whose cache covered more than three makers, the remaining models still give you a consensus, clearly flagged as a partial panel; on a three-seat run there is nothing to fall back on, so the analysis reports the failure and you can run it again.

A cached consensus is never silently rewritten. If the article later gains analyses that would add a maker the panel never had — or a stronger model in one it did — the result offers a Strengthen button naming exactly what it would add and what it costs. In practice that is just the merge pass — everything a rebuild seats is already cached, so nothing new has to be run. Consensus results are a strong default when you care about reliability; a single-model run remains the right choice when you want a specific model's take or the fastest, cheapest answer.

GPT-5.4 Mini (2×) is the system default — a capable, well-rounded everyday choice. If you're skimming a lot and want the lowest cost, the 1× models (GPT-5.6 Luna, GLM 4.7, DeepSeek V3.2) are the cheapest. When an article is especially dense or the argument is tangled, a higher-tier model reads the logic more carefully and is less likely to misread what it actually says. Every model we offer can do the job — the cheaper ones just slip slightly more often on complex reasoning, so stepping up is an easy way to add confidence when you want it.

Yes. In the Chrome extension settings, switch to "per-analysis" mode to assign a different model to each analysis type — for example, Kimi K2.5 for Persuasion and GPT-5.1 for Omissions. The "standouts by analysis type" list on the models page is built for exactly this.

Yes — every plan can use every model, including the free trial. The only difference between plans is the monthly analysis allowance, not which models you can access. You can switch models at any time from the side panel dropdown, no plan change needed; each check is simply charged at the multiplier of the model you used, so you can run everyday articles on a 1× model and save the expensive ones for pieces that matter.

More expensive models are generally larger, newer, or use reasoning techniques that require more computation. Larger models also know more about the world and read further into long articles, which shows up as more depth and fuller coverage on lenses like Omissions, Background, Legal Risk, Critique, and Scientific Assessment — where bigger models were quicker to flag shaky claims in advocacy pieces (though on a rigorous primary-science paper every tier agreed it was sound). For straightforward text-reading checks, a 1× model does a great job.

Every model can analyse an article up to 150,000 characters (~25,000 words). In our max-length tests, models across the price range held full analysis depth all the way to that limit; the only thing that varied was how closely a model engaged with content right at the very end of a very long read, and that tracked the individual model rather than its price tier. For anything longer than 150,000 characters, select the section you care about and analyse that.

Yes — BiasChecker works on articles in any major language. In our testing, quality held up across Chinese, Arabic, and German with no drop versus English, and on-page highlighting works because quotes are extracted in the article's original language. The analysis itself is written in English — except the GPT-5 models (GPT-5.1, GPT-5.4 Mini), which write theirs in the article's own language.

How widely a model is documented to support languages does vary, and each model card on the models page shows it: broadly multilingual models (Gemini, GPT-5, Claude) are built for 100+ languages; the Chinese & English specialists (GLM, Kimi, DeepSeek) are strongest in those two with solid coverage of other majors; and Grok is English-first with a shorter official list. For non-English news, a broadly-multilingual model is the safest pick.

One thing to note: BiasChecker analyses the text as it appears on the page. If your browser's "Translate this page" feature is switched on, it will analyse the translated version — so turn page translation off if you want the original language checked.

By design, BiasChecker.ai adopts a critical perspective — its job is to look for potential weaknesses, gaps, and issues. This means the results will naturally focus on what could be improved rather than what the content does well.

This doesn't mean the content is bad. Even high-quality, well-researched articles may contain subtle framing choices, missing perspectives, or structural patterns worth thinking about.

The analysis is one perspective to consider, not a final verdict. Always use your own judgment alongside the results.

No. AI-generated analysis may express findings in strong, categorical language (e.g. "this claim is unsupported") even when the reality is more nuanced. This directness is a characteristic of how the AI communicates, not a guarantee of accuracy.

The AI can make mistakes, miss context, or flag things that are perfectly reasonable. It may also miss genuine issues. Results should always be treated as suggestions for further thought, not as established facts.

For more details, see section 2.2 (AI Disclaimer) and 2.5 (Critical Nature of Analysis) in our Terms of Service.

BiasChecker.ai is not a traditional fact-checking service, but several analysis types do evaluate content against established knowledge and frameworks:

• Scientific Assessment checks claims against established scientific consensus and known research — it can flag statements that contradict well-established science, misrepresent study findings, or use pseudoscientific reasoning
• Legal Risk evaluates whether proposals or actions described in the text comply with applicable laws and legal principles
• Moral Lens evaluates content against universal ethical principles such as dignity, honesty, equality, and justice

Other analysis types — such as bias detection and manipulation analysis — focus on identifying patterns, framing choices, and rhetorical techniques rather than verifying specific facts. (The Roast is the odd one out: it's pure comedy, not an analytical check.)

For broader factual verification, we recommend using dedicated fact-checking services alongside our analysis.

No. The analysis is for informational and educational purposes only. It does not constitute legal, medical, scientific, or any other form of professional advice. Always consult qualified professionals for decisions in these areas.

Because the Service is designed to err on the side of caution, it may occasionally flag items that are, in context, perfectly reasonable. This is sometimes called a "false positive."

We are actively working to reduce false positives through continuous prompt engineering and model tuning. Large language models are also becoming significantly more capable over time, so you should expect the accuracy and nuance of our analysis to keep improving.

That said, the goal is to prompt you to think about a particular aspect of the text — not to declare it definitively flawed. If you consider the flagged item and decide it's fine, that's a perfectly valid outcome.

All AI systems have limitations and may reflect biases from their training data. We take this seriously and use several techniques to counteract it:

• Prompt engineering: Our analysis prompts are carefully designed to instruct the AI to focus on the writing and reasoning rather than on the identities or political positions of the people involved.
• Continuous improvement: We regularly review analysis results, refine our prompts, and evaluate newer models. As LLMs become more capable, the quality and fairness of our analysis improves with them.
• Multi-model analysis: You can run analyses using different AI models — Claude, Gemini, Grok, and others — whose training data and methods are fundamentally different. By comparing results from independent models, training-specific biases become much easier to spot. If two models trained on different data disagree, that disagreement itself is a valuable signal. Model selection is available to every signed-in user, on all plans, from the side panel.

Despite these measures, no system is perfectly neutral. Note, though, what the analysis actually claims: it identifies observable techniques in the text — loaded wording, one-sided sourcing, omissions — each with cited evidence you can check. The verdict on whether the piece as a whole is fair stays with you. This is why we encourage users to treat results as a starting point for critical thinking, not as an authoritative judgment. For more on our approach, see our About page.

Here are a few tips:

• Try multiple analysis types: Different analyses reveal different aspects of the text. The bias analysis and manipulation check each offer unique insights.
• Read the evidence: Each finding includes a quote or reference from the original text. Check whether you agree with the interpretation.
• Use it as a conversation starter: The results are most valuable when they prompt you to think more deeply, not when taken at face value.
• Compare sources: Analyse the same story from different outlets to see how framing and emphasis differ.

Every analysis of a web article is shared with the community — except text you paste in directly, which is stored privately under your account (the page URL is never sent), so neither the text nor its results are ever added to the community feed. Cached results are free for every signed-in user, on every plan, and don't count against your monthly analyses — so when you open the extension on a page someone has already checked, you get the analysis instantly, at no cost. The more people use BiasChecker, the more of the pages you read are already covered.

Every analysis is saved against the article and the model that ran it, so re-opening one that already exists is free — it doesn't use an analysis. Running something genuinely new — a different article, a different analysis type, or the same analysis on a different model — still counts as normal.

In the browser extension you can turn on Prefer cached results (Settings → AI Model Preference). With it on, opening or switching an analysis loads an existing cached result for the page instead of running a fresh, paid one — a simple way to make a trial or a credit pack last longer.

• If the model you've selected already has a cached result for the page, that one loads.
• If it doesn't, the extension loads the best result that is cached (most capable model first), rather than nothing — so your default-model choice never hides analyses that other models have already produced.
• Only when nothing is cached for that analysis does it run a fresh one, using the model you selected.

So you don't need to keep changing the model dropdown to browse what already exists — change it only when you specifically want a fresh run on a particular model. Anyone who has used up their allowance or credits keeps free access to cached and community results — on every plan, with no time limit.

And it's not only your own analyses: when someone else has already run an article, that cached result is shared with you too. With Prefer cached results on, simply opening the article you want is usually enough to pick one up — wherever it's published — so you don't need to go looking for it. The extension's Community Analyses page is a handy window into some of what's already out there, but it's a curated, partial list, not everything that has been analysed.

Each analysis tab in the extension's side panel carries a small dot in its corner. It answers one question at a glance: is there anything here for me?

• No dot: this lens has not been run on this page yet, by you or anyone else.
• Pulsing ring: analysing right now. Results usually arrive in 5 to 20 seconds.
• Grey: the page was checked and nothing notable was found. You can skip this tab with confidence.
• Blue: there is something to read — either the lens found issues worth your attention, or it produced its analysis (lenses like Intent, Background, Parallels, Rewrite and Roast always produce content rather than a pass/fail verdict). A blue dot briefly pulses when a result lands on a tab you are not looking at.
• Red: the analysis failed. Open the tab to see why and retry.

Because analyses are shared, dots can appear before you run anything: a blue dot means results for that lens already exist for this page and open instantly (reading an existing result is free for every signed-in user, on every plan), and a grey dot means at least one AI model already went through the page with that lens and found nothing notable, so you probably don't need to spend a run on it.

The dot always shows the best current knowledge, and your own result wins: once you run a lens yourself, the dot reflects what you saw. Per-model detail lives in the model dropdown, where a ✓ marks each model that already has a cached result for the current lens.

It's an estimate, not a fixed allowance. Every analysis is billed by the actual tokens it uses (roughly, the length of the article plus the analysis) multiplied by your chosen model's rate — so a short article on a 1× model costs a fraction of a long article on a 10× model. Because each analysis costs a different amount, we can only project how many you have left.

The projection blends a typical cost of about 5,000 tokens per analysis (a mid-length article on a 1× model) with your own average across everything you have run, and the weight shifts smoothly towards your own average as your history grows — about half your own after 10 analyses, three quarters after 30. There is no threshold and no sudden jump: if you favour premium models or long articles, the count adjusts down gradually to match.

The same estimate drives the credit-pack figures. It is computed over your whole history, so a billing reset, a plan change or a new pack never moves it — only your reading does.

Upgrading takes effect immediately — your new, larger monthly allowance is available as soon as the payment goes through, and your renewal date stays the same.

You only pay the difference, not a full new month. We credit the unused part of your current plan and charge the new plan for the rest of the current billing period, so the amount taken today is prorated — usually small. From your next renewal onward you simply pay the new plan's normal monthly price.

Whatever you've already analysed this month carries over: your bigger allowance applies to your existing usage rather than starting a fresh one on top of it. For example, if you've nearly used up a Lite allowance and upgrade to Pro, you get Pro's larger allowance with what you've already used this month counted against it — leaving the remainder for the rest of the period. The full new-plan allowance refreshes on your normal renewal date.

Moving to a lower tier works the other way around — see the next question.

You downgrade right on the pricing page — sign in, pick the lower tier and confirm (the plan buttons only appear as "Switch" once you're signed in). The change is scheduled for your next billing date, not applied immediately:

• Until your next billing date, nothing changes: you keep your current plan, its full monthly allowance, and all its features. You already paid for this period, so you keep what you paid for.
• Nothing is charged today, and no refund is issued for the current period.
• From the next billing date, you're billed the lower plan's price and get its allowance. Your renewal date doesn't change.
• Changed your mind? You can cancel the scheduled change any time before it takes effect — pick your current plan again on the pricing page (or use the "Keep my current plan" link in the banner). Reverting is free.

While a downgrade is scheduled, your dashboard shows "Changes to [plan] on [date]" so there are no surprises. Any unused credit packs are unaffected — they stay yours on any plan.

Downgrading is different from cancelling: a cancellation ends the subscription entirely at the period end, while a downgrade keeps you subscribed on the smaller plan with no gap and nothing to remember.

Yes, in one specific case. If you use up your trial allowance before the 14-day trial window ends, you'll see a one-time offer in the extension and on your dashboard: subscribe before your trial ends and get 15% extra allowance in your first month.

• The bonus is 15% of whichever plan you pick (for example, Lite's 3M-token allowance earns a 450K-token bonus), granted automatically as bonus credits when your subscription starts — no code needed.
• It appears on your dashboard as a "First-month bonus" alongside any credit packs, and is used automatically after your monthly allowance runs out.
• It's first-month only: whatever is left of the bonus expires at your first renewal.

The offer ends when your trial does — if the 14 days pass, or you never used up the trial allowance, the normal plans apply.

In tokens — the small chunks of text a model reads and writes (roughly ¾ of a word each). A standard analysis reads the article and writes up its findings, which comes to about 5,000 tokens for an average-length article on a 1× model — that's where “~5,000 tokens per analysis” comes from (a longer article reads more, so it costs more). We scale the actual tokens by the model's tier (1× standard up to 10× frontier) — we call that your “tokens of work” — so a frontier model costs ~10× as many tokens for the same article.

The cost comes out of your monthly allowance first, then your credit packs (the Power pack is 5,000,000 tokens — about 1,000 standard-model analyses, or ~100 on a top tier). Cached community results are always free, on every plan — you only ever pay for a fresh analysis. See the model tiers page for multipliers.

An analysis's real cost depends on the article's length and the model you choose, so we charge the actual tokens used instead of a flat per-analysis fee. A flat fee would either be set low (and lose money on long, premium analyses) or set high (and overcharge you on short ones) — charging real usage is fairer, and a short article on a fast model can cost a small fraction of a heavy one.

The analysis counts we quote are honest estimates from typical usage; your token balance is always exact. Because you only pay for the tokens each analysis actually uses, the same allowance or pack can stretch to more or fewer analyses than the number shown — and the estimate gets more accurate the longer you use BiasChecker, as it learns from your own reading. Sign in and the estimates on the plan and pack cards update to your own usage.

Lite, Pro, and Premium plans reset monthly on your billing date. The trial expires after 14 days.

Your allowance refills at the start of each billing cycle. If you run out mid-month, a credit pack (from $15 for ~500 analyses / 2.5M tokens) tops you up instantly — credits are only consumed after your plan allowance. And articles the community has already analysed stay free on every plan, so popular news often costs you nothing at all.

Yes — buy a credit pack and use it at your own pace, no subscription required. Packs are the same price for everyone, but subscribers get more for it: an active subscription adds bonus credits to every pack (Lite and Pro +25%, Premium +50%). As a pay-as-you-go user you buy packs at the base amount, your credits fund each fresh analysis, and you don't get a monthly quota. Cached community results are free for everyone. For regular use, a Lite subscription ($8/mo) adds ~600 monthly analyses and the pack bonus. View credit packs →

Packs expire 3 months from the date of purchase. If you buy more than one, each expires on its own date, and the soonest-to-expire is spent first. Packs are used after your monthly plan allowance is exhausted.

Yes. While you have an active subscription, every credit pack you buy comes with bonus credits at no extra cost — Lite and Pro get +25%, and Premium gets +50%. For example, on Premium a Plus Pack gives you 3,750,000 tokens instead of 2,500,000, and a Power Pack 7,500,000 instead of 5,000,000. The bonus is applied automatically at checkout based on your plan at the time of purchase; packs bought without an active subscription get the base amount.

Yes — sign in before buying and the per-pack “≈ analyses” figures update to your account's real usage. We blend the actual cost of the models and article lengths you typically analyse, so the estimate reflects how you actually use BiasChecker rather than the 1× baseline. Signed-in subscribers also see their bonus credits already applied to the totals.

You can get your money back within 14 days of any purchase, no reason needed, wherever you live. Because access starts immediately, the refund may be reduced by the value of what you have already used, priced at what you paid for those credits. Bonus credits are a gift rather than a purchase, so they carry no cash value and are not refunded. One window covers everything, subscriptions and credit packs alike. Separately, if an analysis comes back unusable we put the tokens back rather than have you pay for it, so just tell us. Refunds are processed by Stripe (Link), the merchant of record; cancelling a subscription stops future charges. See our refund policy for details.

It depends on how you submit it.

• Web pages analysed in the extension: we do not store the article text. We store the analysis result — its findings, verdicts and the positions of the passages it cites — in a shared cache keyed to the page, so anyone who analyses the same article later sees the same result instantly instead of paying for a fresh run. Those results are shared with the community and are not linked to your account.
• Text you paste on the analyse page: stored privately under your account, together with the text itself so the result can show you its quotes. It never enters the shared cache or the community feed, and you can delete it from your saved analyses or by deleting your account.

To produce any analysis, the text is sent to the AI model provider you chose (or the default) for processing. We do not share your content or results with third parties beyond that. Full details, including retention, are in our Privacy Policy.