Overview of The Verge Cast — “The one AI detector people actually trust”
In this Verge Cast episode, The Verge talks with Pangram CEO Max Spiro about the growing problem of AI-generated text and why Pangram has become one of the few AI detectors people feel they can actually rely on. The conversation covers how Pangram was built, why earlier detectors failed, how the model estimates confidence, where it’s being used, and what the rise of “humanizers” and AI-assisted writing means for education, publishing, and online trust.
Quick news roundup from The Verge
Before the main interview, the show’s “90 Seconds on The Verge” segment highlights a few tech-policy developments:
- OnePlus is exiting the US and Europe, a major shift for a brand once known for affordable, high-spec phones.
- The EU is forcing Google to make Android more open in Europe, including more access for competing AI assistants.
- The FCC plans to revisit the national broadcast ownership cap, which could allow more consolidation of US media ownership.
Why Pangram stands out in AI detection
The core question of the episode is simple: why do people trust Pangram when so many other AI detectors have a bad reputation?
Max Spiro says Pangram earned trust by:
- Publishing technical papers and benchmark results
- Getting validation from respected researchers
- Focusing on accuracy and false-positive reduction
- Building credibility in research circles before broad product adoption
According to Spiro, Pangram’s standout claim is its extremely low false-positive rate:
- Early version: 1 in 1,000
- Current version: 1 in 10,000
That low false-positive rate is a big reason users are willing to take its results seriously.
How Pangram says it works
Pangram’s approach is different from older AI detectors that mostly relied on perplexity — a measure of how predictable or “unsurprising” text is to a language model.
Older detector problem: perplexity
Spiro explains that perplexity-based detection breaks down because:
- Memorized text can look AI-generated even when it isn’t
- Example: the Declaration of Independence
- English language learners may write simply, which can also be misclassified as AI
Pangram’s approach: active learning + “synthetic mirrors”
Pangram uses a machine-learning strategy called active learning:
- It scans large corpora of human writing.
- It looks for examples where the model makes mistakes or is uncertain.
- For those cases, it creates AI “mirror” versions of the same content.
- It trains on the pair: human text vs. AI-generated version.
This helps the model learn the subtle stylistic differences between human and AI writing, especially in edge cases where the distinction is hardest.
Training on hard examples
Pangram also focuses on the most ambiguous cases — the texts that are hardest to classify. Spiro says those edge cases are where the strongest signal lives, because obvious AI text is usually easy to spot.
What the detector actually looks at
A major theme in the interview is that AI detection is not as simple as identifying one suspicious sentence.
Spiro says Pangram works holistically:
- It evaluates the whole document
- It also breaks text into sections
- It scores parts individually
- Then it aggregates those scores into an overall confidence estimate
So if a document is, for example, 50% AI, that’s not a literal sentence-by-sentence verdict — it’s the model’s estimate based on multiple parts.
Important caveat: the model is partly a black box
The team does not fully know why every specific decision is made. Like many modern AI systems, Pangram is partly a black box. To make results more readable, Pangram provides supporting evidence based on patterns associated with AI writing, but Spiro says the true model behavior is more holistic than a simple checklist.
Who uses Pangram
The episode outlines several major user groups:
- Educators and universities
- To check whether students used AI to fully generate assignments
- Publishers and literary organizations
- To verify that submitted work is human-authored
- AI companies
- To check whether datasets or expert contributions were actually written by humans
Pangram is also integrated into workplace systems, such as Canvas in higher education, so instructors can see results where they already manage assignments.
How to interpret a Pangram result
Spiro’s advice is cautious but confident:
- Longer text = more reliable result
- Short text = more uncertainty
- A flagged result should be treated as a strong signal, not automatic proof
For example:
- A 50-word tweet flagged as AI is likely correct, but still not absolute
- An 80,000-word novel that Pangram says is 90% AI can be treated with much more confidence
The general message: trust the signal, but use judgment.
The role of AI in education and work
The episode broadens beyond detection into the cultural effect of AI writing:
- Students may use AI because of pressure, lack of time, or lack of confidence
- Professionals increasingly use AI for emails, summaries, and writing help
- This creates a kind of cognitive offloading, where people stop doing the thinking themselves
Spiro’s view is that AI writing is not inherently bad, but it becomes a problem when people pretend it was human-written or use it to avoid thinking altogether.
The “humanizers” arms race
A major future challenge is the rise of humanizer tools — services that paraphrase AI-generated text so it won’t trigger detectors.
Spiro describes this as an ongoing adversarial battle:
- Pangram is building models specifically to detect humanized AI
- The company collected data from humanizer tools
- It also built internal humanizers to simulate their behavior and train against them
The next Pangram model is expected to be better at:
- Detecting humanized AI
- Estimating the degree of AI assistance
- Handling newer, more capable AI systems that use long context and research-like workflows
Notable example: the Commonwealth Prize controversy
The interview briefly discusses a recent scandal involving a literary prize-winning story, “The Serpent in the Grove”, which Pangram flagged as 100% AI-generated. The author denied using AI and said he used voice-to-text.
Spiro says he is skeptical of that explanation, citing:
- The story’s writing style
- Inconsistencies in the author’s interview
- The fact that voice-to-text alone would not necessarily produce the kind of output Pangram flagged
The episode uses this case to show how AI detection is increasingly used alongside human judgment and contextual evidence, not in isolation.
Main takeaways
- Pangram has become a trusted AI detector because it focused on accuracy, research validation, and low false positives.
- Older AI detectors often failed because they relied too heavily on perplexity.
- Pangram uses active learning and paired human/AI “synthetic mirror” texts to learn subtle differences.
- Detection is probabilistic, not absolute — longer documents yield better confidence.
- AI detection is now a real need in education, publishing, and AI data quality.
- The bigger issue isn’t just AI text; it’s honesty, authorship, and whether real human thought went into the work.
Bottom line
The episode argues that AI detection has finally become more practical — not because the problem got easy, but because companies like Pangram built systems that are harder to fool and easier to trust. Still, the show makes clear that no detector should be treated as perfect evidence on its own. The real value is in pairing detection with context, conversation, and judgment.