The most accurate AI detector in 2026
Most accurate: Pangram. Its readings in our run: Test in progress
Tested: In progress since 9 November 2026. Page updated: .
Short answer
Sorted only on accuracy, Pangram leads the published evidence: it states a false positive rate of 1 in 10,000, and in a 2025 University of Chicago working paper it alone met a 0.5% false positive cap. Our own run is still scoring it.
Vendor accuracy figures above 99% describe the vendor's own test sets. Ours is one fixed sample set, so read both, and read the base behind every number.
Run RUN-2026-11-A started on 9 November 2026 and is in progress: GPTZero and Originality.ai are scored on the first 120 samples, and every other tool shows a test in progress tag until its readings are in.
The pick table for teachers
Scored on student work only: essays written without AI, essays by students writing in a second language, and AI essays.
| Place | Tool | Score for teachers | False alarms on student essays | Reading | Detection on AI essays | Free plan | Best for | Main limitation |
|---|---|---|---|---|---|---|---|---|
| 1 | Test in progress | Test in progress | Test in progress | 20 checks a day | A second reading you can defend | Free plan caps daily checks | ||
| 2 | Score pending | 0 of 100 | 20 of 20 | Trial credits | Editors checking batches | No permanent free plan | ||
| 3 | Score pending | 0 of 100 | 20 of 20 | Yes | Teachers marking essays | Short texts score unreliably | ||
| 4 | Test in progress | Test in progress | Test in progress | Limited free credits | Departments choosing one tool | Reports are hard to read | ||
| 5 | Test in progress | Test in progress | Test in progress | No | Schools with a licence | Only through your institution |
One student essay, two readings
Sample H-002 · student work, written without AI · run RUN-2026-11-A
GZGPTZero
OAOriginality.aiWhile the idea of having a fully-autonomous is very real, scientists are still very far from it. Research is being conducted every day to improve the efficeincy and safefty of these new semi-autonomous vehicles. In my opinion, driverless cars are inevitable and I would not doubt that I will end up seeing one in my lifetime. But as of right now, cars can't do as much as scientists are hoping for in the future. With that being said, I will there will be a more positive outcome of driverless cars than negative.
Ch. 10% AICh. 22% AIMany skeptics say dirverless cars will never be possible. But, what they don't realize is that we are closer than we ever have been before regardless of how much research is still necessary to succesfully make a fully autonomous car. Google's concept car has been able to drive over half a million miles without a crash. Their car still requires a driver, but it is only for more advanced tasks such as traffic issues or roadwork. Other companies have also attempted to create a driverless concept car and the concepts have showed potential. They are far from driverless but they are going in the right direction. BMW announced that their car could handle driving speeds of up 25 mph but still require the focus of a human being behind the wheel.
Ch. 10% AICh. 22% AIWhile I am very supportive of driverless cars, I think there will always be flaws, no matter the size. I think there will always be malfunctions but that does not mean that those can be fixed either. The research that has been put into these vehicles have been very strenuous and have proven that there will never not be flaws in creating an autonomous car. They still may crash but the probability of a fatal crash in a driverless is much lower than the probability of a fatal crash due to human error. Humans get distracted thus creating more chances for an accident to occur. With driverless cars, there is no chance of human error or the car getting distracted because driving is what it was built for.
Ch. 10% AICh. 22% AIThere are not many current laws or regulations on autonomous cars, so even if Google or BMW were to finish research and create a flawless and driverless car, they wouldn't be able to manufacture them on a commercial level. This is because there wouldn't be any laws covering liability in the case of an accident, if an accident were to somehow occur. It is undetermined who would be at fault for the accident according to the state. I think the company should be more at fault, considering they are advertising these cars as "driverless".
Ch. 10% AICh. 22% AIWith my opinion on driverless cars being said, I still think they are a good idea and will benefit society in the long run and keep more people safe. There are bumps in the road regarding laws and the time it will take to get to where big name companies want to be. But, if we are all patient and let scientists conduct quality research, I see a bright future for fully autonomous cars.
Ch. 10% AICh. 23% AIScroll the chart sideways to read every paragraph.
Which AI detector is best for teachers?
It depends on how you mark and what your school already pays for.
If false alarms matter most
Choose Pangram if a wrong flag would land on a real person.
If you scan in volume
Choose Originality.ai for high detection on raw AI text.
If you need reasons
Choose GPTZero for highlights that show what drove a score.
If you share reports
Choose Copyleaks for reports staff can keep.
How did we rank detectors for teachers?
For this page we scored every detector on student work only: essays students wrote without AI, essays by students writing in a second language, and AI essays, both raw and edited by a person.
False alarms on student writing carry the most weight, because a wrong flag lands on a real student. Detection on raw AI text comes next, then detection on AI text a person has edited, then access and price.
Every number on this page comes from run RUN-2026-11-A. The CSV for that run is on the test runs page.
Read how we test| Factor | Weight | What it measures |
|---|---|---|
| False alarms on human writing | 40% | Student essays scored as AI |
| Detection on raw AI text | 30% | Unedited AI essays scored as AI |
| Detection on human-edited AI text | 15% | AI essays a person revised, scored as AI |
| Access and price | 15% | Free plan, limits, school sign up |
Every detector, the same fields
Verdict, score, who it suits, what it gets wrong, the price on the day we checked, and how it read the student set.
PGPangram
Most accurate- Verdict
- The careful second reading. Use it when a flag will be argued over.
- Score
- Test in progress
- Sub-scores
- Readings Test in progress
- Best for
- A second reading you can defend
- Main limitation
- Free plan caps daily checks
- Price and plan limit
- $20 a month, 300,000 words. Free plan: 20 checks a day. Checked on
| Sample rows | Samples | Scored as AI |
|---|---|---|
| All sample rows | 180 | Test in progress |
OAOriginality.ai
- Verdict
- Built for editors and batches, priced for regular use.
- Score
- Score pending
- Sub-scores
- False alarms 0 of 100Raw AI 20 of 20Edited AI Test in progress
- Best for
- Editors checking batches
- Main limitation
- No permanent free plan
- Price and plan limit
- $14.95 a month, 200,000 words. Free plan: Trial credits. Checked on
| Sample rows | Samples | Scored as AI |
|---|---|---|
| Student essays, no AI | 60 | 0 of 60 |
| Second language essays, no AI | 20 | 0 of 20 |
| Professional prose, no AI | 20 | 0 of 20 |
| AI essays, raw | 20 of 60 so far | 20 of 20 |
| AI essays, edited by a person | 20 | Test in progress |
GZGPTZero
- Verdict
- The pick for marking essays. Open the sentence view before you act on any flag.
- Score
- Score pending
- Sub-scores
- False alarms 0 of 100Raw AI 20 of 20Edited AI Test in progress
- Best for
- Teachers marking essays
- Main limitation
- Short texts score unreliably
- Price and plan limit
- CA$25.99 a month (Premium, monthly), 300,000 words. Free plan: Yes. Checked on
| Sample rows | Samples | Scored as AI |
|---|---|---|
| Student essays, no AI | 60 | 0 of 60 |
| Second language essays, no AI | 20 | 0 of 20 |
| Professional prose, no AI | 20 | 0 of 20 |
| AI essays, raw | 20 of 60 so far | 20 of 20 |
| AI essays, edited by a person | 20 | Test in progress |
CLCopyleaks
- Verdict
- One tool a department can share, with reports staff can keep.
- Score
- Test in progress
- Sub-scores
- Readings Test in progress
- Best for
- Departments choosing one tool
- Main limitation
- Reports are hard to read
- Price and plan limit
- $16.99 a month, 100 credits. Free plan: Limited free credits. Checked on
| Sample rows | Samples | Scored as AI |
|---|---|---|
| All sample rows | 180 | Test in progress |
TITurnitin
- Verdict
- Use it if your school has it, and read the full report, not the headline number.
- Score
- Test in progress
- Sub-scores
- Readings Test in progress
- Best for
- Schools with a licence
- Main limitation
- Only through your institution
- Price and plan limit
- Sold to institutions only. Free plan: No. Checked on
| Sample rows | Samples | Scored as AI |
|---|---|---|
| All sample rows | 180 | Test in progress |
Why do accuracy claims disagree?
Every accuracy number has a base; read it before the number.
Vendors test on their own data, independent studies on theirs. On the shared RAID benchmark (Dugan and colleagues, 2024), commercial detectors were compared at a fixed 5% false positive rate, and their ranking changed with the attack and the domain. A 2025 University of Chicago working paper found only one of four detectors met a strict false positive cap. Our run is one more base, published in full.
Read the sentences, not the number
Open the sentence or paragraph view and see what drove the score. A textbook style paragraph, like P3 above, often explains the whole flag.
Compare with work you know
Set the flagged essay beside the student's earlier writing. A sudden change in voice tells you more than a percentage.
Ask the student about the draft
Ask how they wrote it, what they changed and what they read. Drafts, notes and version history answer more than any detector.keep the version history
Record what you found
Note what you checked and what the student said, not only the score. If the case goes further, the score should never stand alone.
No detector score should be the only evidence in a misconduct case.
Questions from teachers
What is the most accurate AI detector?
On published evidence, Pangram: in a 2025 University of Chicago working paper it was the only tool tested to meet a 0.5% false positive cap. Our own readings for it are in progress; GPTZero and Originality.ai each flagged 0 of 100 human texts so far.
Do AI detectors reach 99% on every test?
Not on every set. Vendors report figures on their own data; independent tests such as Weber-Wulff and colleagues (2023) found all 14 tools tested below 80% accuracy.
How is accuracy measured here?
As counts: human essays flagged out of 100, raw AI texts caught out of 60 and edited AI texts caught out of 20, all from one dated run. How we test has the rule for a flag.
Related reading
Best AI Detectors (2026): Picks Backed by Our RunsThe overall ranking.
Best AI Detector for Teachers (2026): Classroom PicksThe pick for a stack of essays and one flag to check.
Pangram vs GPTZero: Which Detector Fits Your Work?Benchmarks, claims and false alarms.
AI Detector Test Runs: Dated Results and Raw FilesEvery run, its samples and its files.