Pangram CEO Max Spero says the internet is "dangerously close" to the dead internet theory becoming reality within a few years, as AI-generated text and images push into job applications, product reviews, and insurance claims. The AI detection startup recently raised $9 million and signed a partnership with Substack, which now uses Pangram's technology to show readers which of their favorite newsletter authors write with AI assistance. Spero made the comments on an Equity podcast episode published September 2, 2026, and the argument sits at the center of a fast-growing category: startups selling themselves as the trust layer for a web where nobody knows what a human wrote anymore.
The dead internet theory — the idea that most online content and engagement is machine-generated rather than human — was fringe when it surfaced around 2021. Spero's pitch is that generative models have collapsed the timeline. Pangram's own product surface has expanded to match: alongside its text classifier, the company recently shipped an AI image detection tool aimed at the same trust problem in visual media.
The Substack deal is the more concrete signal. Substack is a platform whose entire economic proposition rests on readers paying for a specific human voice. Surfacing which writers lean on AI, and how heavily, changes the subscriber calculus in a way a generic content-moderation label cannot. Pangram is being paid to answer a question the platform cannot credibly answer itself.
Key facts
- 01Pangram raised $9M for its AI detection system and partnered with Substack to flag which newsletter authors write with AI.
- 02CEO Max Spero says the dead internet theory could become reality within a few years as AI text floods reviews, job applications, and insurance claims.
- 03Pangram recently launched a new AI image detection tool alongside its text classifier.
- 04Spero argues detection should measure how much AI went into content, not just label it AI or human.
Spero's second argument is subtler than the first, and it's the one that matters for how detection tools get built. A binary AI-or-human label, he says, is the wrong output. The useful question is how much AI went into a piece of writing — the difference between a human draft polished with a model, a model draft edited by a human, and a fully synthetic post is enormous, and lumping them under a single flag destroys the signal.
That framing is also a hedge against the failure mode that has haunted every AI detector shipped so far: false positives. Pangram has previously said its classifier is tuned for a low false-positive rate, and Spero highlighted the stakes of getting it wrong on the podcast, especially when the flagged content is a sensitive image or a job application that decides whether someone gets hired. A detector that wrongly brands a human writer as a bot is worse than useless — it becomes the injustice, not the solution.
The labor-market read from Spero is bleaker and more specific than the usual AI-and-jobs discourse. Cheap, generic writing work — SEO filler, low-end copywriting, mass-produced blog posts — is, in his view, already gone. The counterweight is that genuinely good human writing becomes more valuable, not less, because it becomes scarcer and more identifiable against a background of synthetic text. Detection tools are part of what makes that scarcity legible to a reader or a buyer.
Whether Pangram's technical claims hold up under adversarial pressure is the open question. Every detector released to date has faced the same arc: strong benchmarks at launch, followed by rapid degradation as writers learn which stylistic tells the model keys on and adjust. Frontier labs including OpenAI have quietly shelved their own detection products after concluding accuracy was not defensible. Pangram is betting that a focused startup, iterating faster than the model releases it is trying to catch, can hold a line the labs abandoned.
The $9M round, which we covered last week, is small by AI-infrastructure standards but sized appropriately for a company whose product is a classifier, not a foundation model. The more interesting metric is distribution: a Substack integration puts Pangram in front of every paid newsletter reader on the platform, which is the kind of embedded surface area that competing detectors have not secured.
Pangram is a wedge into a market that will either become essential or become impossible. If detection accuracy holds, trust layers become as standard as spam filters, and companies like Pangram sit under every platform that sells human-generated content — publishing, hiring, insurance, e-commerce reviews. If detection breaks under adversarial pressure from the next model generation, the dead internet Spero is warning about arrives on schedule, and the trust layer becomes a story about a category that could not scale fast enough to matter.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
