Free AI Detectors Compared: Limits, Privacy and False Positives
October 19, 2026 · A practical comparison of free AI writing checks, subscription trials, input limits and the evidence needed to interpret a result.
Source checked October 19, 2026. Editorial research; no original accuracy benchmark or hands-on score is claimed.
Start with the limit you actually need
There is no accuracy winner in this guide: we checked official product information, not a shared test dataset. The useful first distinction is between a recurring free allowance, a public input limit and a subscription trial.
| Tool | What was verified | What remains to check |
|---|---|---|
| Pangram | Official help: up to 2,000 text words daily; 3 image checks daily | Confirm account allowance; paid trial is separate |
| GPTZero | Public field: 10,000 characters | Account quota, paid entitlements and retention |
| ZeroGPT | Public field: 15,000 characters | Scan frequency, paid entitlements and retention |
| Copyleaks | Paid credit plans; “Start free” route | Free-account or trial allowance before relying on it |
Sources: Pangram free-account help, GPTZero checker, ZeroGPT checker, Copyleaks pricing. All checked October 19, 2026; logged-in terms can differ.
What a false positive means
A false positive is human writing labeled as AI. A false negative is generated writing missed by the detector. A headline accuracy claim can hide an uneven tradeoff between the two. Scores from different services may use different definitions and should not be averaged into a verdict.
Turnitin’s AI Writing Report guidance explicitly cautions against making adverse decisions solely from its detector output. Our editorial recommendation is to require drafts, notes and a human discussion when another person’s work is disputed.
Free, trial and paid: a decision tree
First estimate the job. One article a week can fit a different route from hundreds of class submissions on one deadline. List typical document length, required languages, need for source matching, reporting requirements and whether the text may leave your organization. Eliminate options that cannot meet the privacy or input constraints before comparing price.
| If you need… | Start with… | Do not overlook… |
|---|---|---|
| A recurring small text allowance | Pangram’s documented free account | A daily cap may not fit a single large deadline. |
| A quick paste into a public checker | GPTZero versus ZeroGPT | A character counter is not a word or monthly quota. |
| Revision or writing-process context | GPTZero | Confirm the required document workflow and paid gates. |
| AI review plus source matching | Copyleaks | Check the shared allowance and the selected scan modes. |
| Private institutional submissions | An approved organizational agreement | A free consumer account is not automatically approved for those records. |
Four questions a free AI checker cannot answer alone
Who actually drafted the passage? The final text does not contain a complete record of how it was created. Use revision history, notes, intermediate files and a conversation with the author when provenance matters.
Was the assistance allowed? Rules about brainstorming, grammar correction, translation and generated prose differ between assignments and publishers. The classifier does not know which rule applies.
Is the material copied? Source matching and attribution review are separate. Human text can be plagiarized and generated text can have no indexed match.
Is the article useful and accurate? A human-origin label does not verify facts, quotes, permissions or reader value. For publishing, verify those directly instead of optimizing the detector score.
A false-positive example you can change
The probability that a flag is meaningful depends partly on how common generated documents are in the material you check. In a hypothetical set with very little generated writing, false alarms can make up a large share of the flagged set even when the false-positive rate looks small. The defaults below model 1% prevalence; no vendor has been assigned these performance values.
Expected true flags = N × prevalence × sensitivity. Expected false flags = N × (1 − prevalence) × false-positive rate. Fraction of flags that are true = true flags ÷ their sum. Change prevalence to 10% while keeping the other defaults: the expected fraction rises from about 47.6% to about 90.9%. That change comes from the sample population, not an improved detector.
Build an evaluation set without inventing ground truth
| Research to inspect | What it contributes | Limit on interpretation |
|---|---|---|
| Detection-tools study, 2023 | Historical evidence on reliability and text transformations | Older tools and test conditions are not current-version rankings. |
| Chicago research brief, 2025 | Four-detector comparison and threshold/error tradeoffs | ZeroGPT was not in that four-system comparison. |
| Pangram 4 technical report, 2026 | A model-specific report and benchmark claims | A model report is not a neutral all-product purchasing verdict. |
- Human originals: use material with credible drafting provenance. Do not label it “human” merely because another checker cleared it.
- Generated examples: save the model, date, prompt and raw output. Do not use just one model or topic to stand in for the whole category.
- Mixed drafts: record the editing steps and report them separately. A mixed label has a different meaning from pure generation.
- Realistic formats: include the same input lengths, languages and file types as your planned use. Compare extracted text before blaming a detector for an OCR problem.
- Uncertain cases: keep uncertainty visible. Do not force every borderline result into a binary outcome solely to make a table look complete.
The 2023 detection-tools study is useful for understanding test design and the effect of transformed text. Its results should not be copied into a table labeled as a current test of today’s models. A current benchmark needs its own versioned inputs and results.
Download our blank detector evaluation worksheet (CSV) and purchase checklist. They contain no detector results. Fill them with your own authorized samples.
Measure total review cost, not just the scan fee
A free scan can still be expensive if a reviewer spends ten minutes resolving each false flag. During a pilot, record minutes spent per report, repeat scans, upload failures and disputed cases. For an organization, add account administration and the cost of handling private records appropriately. “No subscription charge” is a narrower claim than “no cost.”
A useful purchasing comparison has three columns: allowed workload, review effort and unresolved risk. A product that is marginally cheaper but cannot preserve the original report may be a worse fit than one that supports a clear audit trail. Conversely, occasional low-stakes checks may not justify a team subscription at all.
What to do when your human writing is flagged
Preserve the submitted version, full report and date. Gather outlines, source notes and revision history. Ask which passages are at issue and explain how they were created. Do not repeatedly rewrite the text until a checker produces the desired number; that changes the evidence without resolving the original question. Where a decision is disputed, use the institution’s or publisher’s review process.
Frequently asked questions
What is the best free AI detector?
There is no universal winner established by this guide. Pangram has a documented recurring allowance, while GPTZero and ZeroGPT expose public checker routes with different visible input capacities. Choose by task, privacy and review workflow, then evaluate representative samples.
Are free AI detector results reliable enough to punish a student?
A detector-only decision is not supported by this guide. The result needs context, provenance and a fair review of the actual writing policy. Our comparison provides workflow information and hypothetical calculations, not an integrity verdict about any person.
Does a 0% AI score prove original work?
No. It does not establish authorship, absence of copied material or factual accuracy. Check citations, source evidence and writing history as appropriate to the task.
Can I convert characters to words for a fair quota comparison?
You can estimate for a known document, but a fixed conversion is unreliable across languages and writing styles. Count the actual text and keep the vendor’s original billing unit. Do not label a character limit as a word allowance.
Why are some prices marked unverified?
The public individual plan tables did not reliably expose current figures during the check, including a browser inspection. Publishing an older price as current would create false precision. We link the official route and state the evidence gap instead.