Pangram Labs (formerly Checkfor.ai): Product, team and how it was built
Pangram Labs (formerly Checkfor.ai) (AI detector for text, images and video, sold to individuals, schools, publishers and platforms, United States): Paid plans and an API on sale, November 2026. Evidence grade E: no revenue figure was checked. Team: 2 co-founders at the October 2023 launch; headcount not published; the company's site named 11 people with roles in November 2026, 3 of them interns. Launched 12 October 2023; still running.Added
- Outcome
- Paid plans and an API on sale, November 2026
- Evidence
- E no revenue figure was checked
- Launch
- 12 October 2023
- Status
- Still running
- Team
- 2 co-founders at the October 2023 launch; headcount not published; the company's site named 11 people with roles in November 2026, 3 of them interns
- Outside funding
- Funded: SiliconANGLE reported US$9 million in July 2026; the June 2025 round is US$4 million in a Reuters headline and US$2.7 million in SiliconANGLE
- Group
- Learning and study
- Main channel
- Not stated
What is not verified
No revenue figure was found. A payment-processor profile, a founder interview with figures and a company filing were looked for and not found. The statement that customers pay rests on the company's price list and on its own pages. No independent source confirms that these customers pay. The June 2025 round is US$4 million in a Reuters headline and US$2.7 million in a SiliconANGLE article. Wikipedia retells a newspaper article on user counts and revenue growth. That article could not be opened, so those figures are not used here. No page states the headcount. The company's site names 11 people with roles, 3 of them interns, and its job board lists 14 open positions. This story is grade E: the business is real and its product, price and customers are described from public pages, but no revenue figure was checked. Aiolum shows it as an idea worth studying, not as proof that it makes money.
Pangram is an AI detector: software that estimates whether a person or a language model wrote a text. It sorts passages into human-written, AI-assisted and AI-generated, and it also checks images and video. The company's home page says that it serves "educators, publishers, law firms, HR departments, foundations, and more". It sells monthly plans to individuals and pay-per-use access to developers.
Max Spero (chief executive) and Bradley Emi (chief technology officer) founded it in Brooklyn, New York. The company's About Us page says that they met in their first year in a Stanford dormitory. It launched on 12 October 2023 as Checkfor.ai and took the name Pangram Labs on 8 April 2024. The founders' launch post called the company "a small team of experienced machine learning researchers and engineers".
This story has no revenue figure. SiliconANGLE reported on 29 July 2026 that the company raised US$9 million, led by Menlo Ventures. A Reuters headline on the company's press page reports US$4 million in June 2025, and SiliconANGLE states US$2.7 million for that round. The status "paid plans live" comes from the company's own pricing page.
Why did Pangram Labs (formerly Checkfor.ai) work?
Four causes, in the order they mattered. Each one is a reusable lesson.
Both founders had shipped machine learning products before they started
The About Us page lists the first founder's work at Nuro, Google, Two Sigma and Yelp. It lists the second founder's work at Absci and on Tesla Autopilot. The launch post says that the team had shipped real-world machine learning at Google, Tesla Autopilot and Nuro. No source tests whether this experience explains the product.
It trades some missed AI text for fewer false accusations
Max Spero says in a June 2026 interview that the model is calibrated to 1 false positive in 10,000. He says that the company works to catch more AI text at that rate. A reviewer of the product writes that the model treats a false positive as 10 times worse than a false negative. Epoch AI's test shows the other side: some AI text that imitates an author is not flagged.
It charges people per month and developers per use
The company's pricing page lists a free daily allowance, an Individual plan and a Professional plan. It also lists an API that charges per 100 words of text, per image and per second of video. The same text model serves a browser extension, an API and learning management systems.
How it was built
These details come from the company's pages and a founder interview. They are grade C, and the page source gives grade B for the web tools. [8, 16, 21, 15, 2, 1]
- In a June 2026 interview Max Spero says that the founders "decided to train a deep-learning classifier" instead of measuring perplexity, the unexpectedness of each word. He says that the model is retrained all the time.
- The company says that it adds human texts that the model wrongly flagged, together with AI versions of them, to the training data. It calls this hard negative mining.
- The company's Pangram 4 post says that the model is over six times larger than its predecessor. The Pangram 4.1 model card says that the model is meant for texts of at least 50 words.
- The privacy policy says that the company uses Amazon Web Services for data hosting and Stripe for payment processing.
- The website's page source points to Next.js, the Sanity content service, Amazon S3 file storage and Google Tag Manager. Job posts link to Ashby.
- The website shows a badge "SOC 2 Type II · Verified by Sensiba" and a link to a trust center. The trust center was not read.
Tools named in the sources
Each one opens its page in the Tools tab, with every story that uses it.
Who tried this before Pangram Labs (formerly Checkfor.ai), and why did they fail?
Two companies in our database of failed startups sold AI products to institutions on a promise of reliability. They fit only in part, because neither one detected AI text. The values below are from that database, which calls them AI-assisted summaries, and are not independently verified.
ROSS Intelligence
The promiseAn AI legal research platform that finds case law for lawyers faster and with more accuracy than older methods.
Why it diedThe record says that it could not gain enough traction against incumbents that had decades of data and integrated products. It names Westlaw and LexisNexis as the large players today.
Olive AI
The promiseSoftware bots that take over the administrative work of hospitals, such as prior authorizations and billing checks.
Why it diedThe record says that it sold a vision of autonomous AI that it could not deliver. It says that each hospital needed extensive customization, which broke the cost per customer.
The failed startups come from our database of 1,749 dead startups, retold here in our own words. Its entries are AI-assisted summaries of public sources, so these cards carry no evidence grade.
What changed
ROSS and Olive sold AI to professional buyers on a promise of accuracy and low effort. According to their records, one met incumbents and the other met the cost of customizing for each customer. Pangram sells a narrower promise, a verdict on one text, through a page, an extension and an API that need no customization.
The company states its accuracy as a measured error rate and invites outside tests. Epoch AI and the Reuters Institute for the Study of Journalism published results that differ from the company's figures in some conditions. The company's accuracy claims are therefore its own, and the market for AI detection is disputed.
Can one person copy Pangram Labs (formerly Checkfor.ai) now?
Only in part: the method can be copied, the accuracy claim cannot be checked here.
Two founders with machine learning experience built one detector, published prices and a free daily allowance, and sold to schools, publishers and developers. A new builder would need the same depth in model training and a large set of labelled human and AI text. Without a published error rate and outside tests, buyers have little reason to trust a new detector.
Where it is still open: Detection in areas where the company itself asks for outside tests, such as newly scanned historical documents and the output of speech-to-text programs.
- 1Pick one buyer problem where a wrong verdict has a cost, such as moderation or publishing review.
- 2Publish the error rate and invite outside researchers to test it, as the company did on 6 November 2026.
- 3Offer a free daily allowance and a monthly plan for individuals, and a per-use API for developers.
- 4State the limits in plain words, such as short texts, code and texts that imitate an author.
How much revenue does Pangram Labs (formerly Checkfor.ai) make?
Each row shows the evidence grade and the source numbers from the list at the end of this page. The grades are explained in How Aiolum verifies.
| Fact | Value | Grade | How we know | Sources |
|---|---|---|---|---|
| Company start | 2023, in Brooklyn, New York | C | The home page says "Founded in Brooklyn, NY, in 2023 by AI researchers from Tesla and Google". The launch post says that the founders had worked on the problem for about a month. | [2, 4] |
| Launch and earlier name | 12 October 2023 as Checkfor.ai; renamed Pangram Labs on 8 April 2024 | C | The launch post is dated 12 October 2023. The rename post of 8 April 2024 says that nothing under the hood changed. The terms name the legal entity as Pangram Labs, Inc. | [4, 5, 14] |
| Team | 2 co-founders: Max Spero (CEO) and Bradley Emi (CTO) | C | The About Us page names both and says that they met in a Stanford dormitory. No page states the number of staff. Author pages on the company's site name 11 people with roles, 3 of them interns. The job board lists 14 open positions. | [3, 4, 22, 23] |
| Funding | US$9 million, July 2026; an earlier round in June 2025, reported as US$4 million and as US$2.7 million | C | SiliconANGLE reported on 29 July 2026 that the company announced US$9 million, led by Menlo Ventures. It gives US$2.7 million for June 2025. A Reuters headline of 24 June 2025 on Pangram's press page says US$4 million. The Reuters and TechCrunch articles could not be opened. | [20, 6] |
| Published prices | Individual US$20 a month, Professional US$65 a month, API US$0.05 per 100 words | C | The company's pricing page. Individual starts with a 7-day free trial, then US$20 a month. The home page offers 20 free credits a day. | [1, 2] |
| Named customer | Quora, since April 2024 | C | The company's own case study says that it partnered with Quora in April 2024. The home page shows a NewsGuard logo without text or a link. | [11, 2] |
| False positive rate, company claim | 1 in 10,000 on the home page; 0.0041% for Pangram 4 | C | The home page and the founder's interview say 1 in 10,000. The company's Pangram 4 post of 29 July 2026 says 0.0041%, roughly 1 in 24,000, on internal benchmarks. | [2, 8, 16] |
| Independent test by Epoch AI | 0 of 495 human passages flagged; 30 of 297 author-imitating AI passages not flagged | C | Epoch AI published it on 15 July 2026 for Pangram version 3.3.2. It counts a "Mixed" verdict as a miss, which gives 10.1%. | [9] |
| Adversarial test reported by a journalism institute | 86% of 588 AI samples labelled human | C | The Reuters Institute for the Study of Journalism wrote this in July 2026 about a researcher's rhyme, repetition and archaic-language samples. | [10] |
| Age limit and audience | 18 or older; marketed to teachers, universities, publishers, platforms | C | The terms say that a user must be at least 18 years old. The home page lists solutions for teachers and higher education, and no page aims at students. | [14, 2] |
How to read this table: grade C is what the company states on its own pages, or what a reviewer or a journalist wrote. No value is grade A or B, and no row is a revenue figure.
| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Free | 0 | 20 free credits a day | |
| Individual | 20 | 300,000 words and 100 image scans a month, 7-day free trial | |
| Professional | 65 | 5 times the words and image scans of Individual, US$200 of API credits a month | |
| API, text | US$0.05 per 100 words, pay per use |
The pricing page shows an Individual plan with a 7-day free trial and a Professional plan billed at once. The page says "Save $60" and "Save $240" for yearly payment but shows no yearly amount. API rates are per use: US$0.25 per image, US$0.05 per image on the Flash model and US$0.05 per second of video. Bulk text API has a 20% discount, according to the page. An Organizational tab and enterprise plans were not readable as text.
When did Pangram Labs (formerly Checkfor.ai) launch, and how did it grow?
From a launch post of two founders to a company that sells several plans, in 37 months. Each entry carries its grade.
- 12 October 2023CThe founders launch Checkfor.ai, a tool that detects AI-generated text from a copy and paste. They say that they have worked on it for about a month. [4]
- April 2024CCheckfor.ai becomes Pangram Labs on 8 April. The company's case study says that it partnered with Quora in April. [5, 11]
- June 2025DPangram's press page lists a Reuters headline of 24 June 2025 that reports a raise of US$4 million. SiliconANGLE later gave US$2.7 million for this round. [6, 20]
- 15 July 2026CEpoch AI publishes tests of three detectors. Pangram flags no human passage and misses 30 of 297 passages that imitate an author. [9]
- 29 July 2026CThe company releases Pangram 4. SiliconANGLE reports on the same day that it raised US$9 million, led by Menlo Ventures. [16, 20, 6]
- 6 November 2026CThe founder asks researchers for new independent evaluations and offers credit grants with no publishing terms. [17]
What could go wrong for Pangram Labs (formerly Checkfor.ai)?
The risks are the mirror image of the causes above.
What is not verified
No revenue figure was found. A payment-processor profile, a founder interview with figures and a company filing were looked for and not found. The statement that customers pay rests on the company's price list and on its own pages. No independent source confirms that these customers pay. The June 2025 round is US$4 million in a Reuters headline and US$2.7 million in a SiliconANGLE article. Wikipedia retells a newspaper article on user counts and revenue growth. That article could not be opened, so those figures are not used here. No page states the headcount. The company's site names 11 people with roles, 3 of them interns, and its job board lists 14 open positions.
Counters, badges, logo wall and testimonials
The home page shows counters "1 in 10,000" false positives and "#1 In studies from independent researchers". A logo wall headed "Trusted by Global Brands" has eight logos without text or links. The image file names are Quora, Newsguard, Featured, Haro, Qwoted, Vibegrade, WHE and WikiEdu. Eight quotations carry a name and a role. One is signed "Cloud Architect" with no employer. A "Wall of Love" page repeats posts from X. A search for the same sentences under other names could not be made, so reuse was not checked. A newsletter article about the product says that companies pay its author, and it gives readers a discount code.
Published tests and claims about accuracy differ
The company says that its false positive rate is 1 in 10,000 on its home page and 0.0041% in its July 2026 post. Epoch AI found no false positive in 495 human passages and 30 of 297 author-imitating AI passages not flagged. The Reuters Institute wrote that a researcher's 588 adversarial samples were labelled human 86% of the time. Wikipedia reports an August 2026 preprint on humanized text. On 6 November 2026 the company asked for new independent tests.
Results have been used in public accusations
Wikipedia reports that Pangram results started public accusations of AI use in a novel, an essay and prize-winning stories. It says that the people named denied or disputed them. It also says that a Wall Street Journal editor called such a tool a "defamation machine" and that The Atlantic questioned relying on it. The terms require users to be 18 or older. The product detects AI text and does not offer to hide it.
Sources
The grade next to each source is the best grade that the source can support. "Read in full" means that the text of the page was downloaded and the numbers were checked in it.
- C Pangram, Pricing, read 12 November 2026Read in full as text. The company's own page. Source for the plans, the free trial and the API rates.
- C Pangram, home page, read 12 November 2026Read in full. Source for the founding year, the counters, the logo wall, the quotations and the accuracy claims.
- C Pangram, About Us, read 12 November 2026Read in full. Source for the two founders, their earlier work and their Stanford start.
- C Pangram, blog post: AI detection just got a lot better: Announcing Checkfor.ai, 12 October 2023Read in full. Written by the two founders. Source for the launch date, the earlier name and the team description.
- C Pangram, blog post: Checkfor.ai is now Pangram Labs, 8 April 2024Read in full. Written by the chief executive. Source for the rename.
- C Pangram, Press page (list of articles), read 12 November 2026Read the list of 62 entries. Source for the Reuters and TechCrunch headlines. The two articles are not available to this pipeline (Reuters answers with a bot challenge, TechCrunch is on the blocked list).
- D Wikipedia, Pangram (AI detector), read 12 November 2026Read in full. A summary of other articles. Source for the disputes and the August 2026 preprint only. Its figures from articles that could not be opened are not used.
- C AI Policy Perspectives, Q&A: Pangram CEO, 23 June 2026Read in full. An interview with the chief executive. Source for the classifier method and the calibration of the false positive rate.
- C Epoch AI, AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors, 15 July 2026Read in full. An independent test of three detectors. Source for the counts of passages.
- C Reuters Institute for the Study of Journalism, A human wrote this - believe me, July 2026Read the section on Pangram. Source for the adversarial samples and the 99.98% claim.
- C Pangram, case study: How Quora uses Pangram to handle AI-written answers, 26 September 2024Read in full. Written by the chief executive. Source for the April 2024 partnership.
- C What's the Bet: Pangram (paid analysis newsletter), 24 April 2026Read in full. A paid analysis, as its author's note says. Source for the founder's quotation and the calibration of false positives.
- C Pangram, Wall of Love (testimonials), read 12 November 2026Read in full. Source for the repeated posts and quotations.
- C Pangram, Terms of Service, read 12 November 2026Read the sections on eligibility and the legal entity. The street address on the page is not used.
- C Pangram, Privacy Policy, read 12 November 2026Read the section on third-party services. Source for Amazon Web Services and Stripe.
- C Pangram, blog post: Introducing Pangram 4, 29 July 2026Read in full. Written by the chief executive. Source for the model size and the error rates on internal benchmarks.
- C Pangram, blog post: A call for new Pangram evaluations, 6 November 2026Read in full. Written by the chief executive. Source for the grants with no publishing terms.
- C database of failed startups, record of ROSS Intelligence, read 12 November 2026Record page read on the website on the day of writing. An AI-assisted summary, not independently verified.
- C database of failed startups, record of Olive AI, read 12 November 2026Record page read on the website on the day of writing. An AI-assisted summary, not independently verified.
- C SiliconANGLE, Pangram Labs raises $9M to launch more accurate AI detection for text and images, 29 July 2026Read in full. A journalist's report of the company's announcement. Source for the July 2026 round, its investors and the June 2025 amount.
- C Pangram, Pangram 4.1 model card, read 12 November 2026Read the section on intended usage and limitations. Source for the minimum of 50 words.
- C Pangram, site map with 23 author pages, read 12 November 2026Read each author page. 11 people have a role at the company, 3 of them interns. Guest writers and pages without a role are not counted.
- C Pangram Labs, job board at Ashby, read 12 November 2026Read the page source, which lists 14 open positions. It is not a staff count.
Run this company? Send corrections or a verified figure to aiolum@livyden.com