The Sinceerly Stack

Essay This entry carries an argument or interpretive position, not just a neutral definition.

The product

In April 2026, a Harvard MBA grad named Ben Horwitz launched a Chrome extension called Sinceerly — the doubled e being either a typo deliberately preserved or a feature, depending on how charitable one is feeling¹. The tagline reads, in the upright sans-serif that has become the house font of AI-skeptical tooling: AI to undo your AI writing.

The product is a humaniser. It takes text generated by an AI tool — typically a Gmail draft you wrote with ChatGPT or Claude or Copilot — and reverses the AI-shaped tells. From the marketing page, three product names for three intensity levels:

Subtle. Removes em-dashes, smooths the most obvious AI-isms, leaves most of the prose intact.
Human. Removes the “not just…” family of constructions. Strips em-dashes more aggressively. Adds occasional informalities.
CEO. All of the above, plus: drops words to lowercase. Adds typos. Inserts acronyms. Shortens. The marketing page advertises the result with mock email exchange in which the user’s careful corporate paragraph becomes “sent from my iPhone” plus three sentences with minor errors. The branding leans into the joke.

The founder’s own pitch, copied from the About section verbatim:

“Hi, I’m Ben. I got sick of everyone in my inbox sounding like AI. So I built Sinceerly. I tested Sinceerly by cold emailing 5 Fortune 500 CEOs. 4 CEOs replied. Of those replies, each was under 10 words. 2 replies had typos. One reply called me Larry (my name is Ben).”

The pitch is funny on at least three levels. It is also, charitably read, a sincere observation about how senior executives actually write email. CEO emails are short, slightly typo-laden, and occasionally call you by the wrong name, because the CEO is busy. The Sinceerly product packages this observation as a service.

The product page also offers a free tier (three rewrites), a Pro tier at $4.99/month for unlimited rewrites, and — this is the load-bearing detail for the entry that follows — a Bring Your Own API Key option that lets the user supply their own Anthropic key directly. The recursive stack is therefore explicit: Sinceerly is a Claude-backed humaniser of Claude-generated text. The model is being used against its own training-corpus tells.

The stack, named

The Dictionary names the structure that Sinceerly sits inside, because the name is more durable than the product:

The Sinceerly Stack: a recursive cat-and-mouse layer of AI tooling, in which each successive tool is designed to counteract the visible artifacts of the previous one. The current state of the stack, as of May 2026, has at least four layers:

AI to write. ChatGPT, Claude, Gemini, Copilot, and the broader ecosystem of language models used to generate first-draft prose.
AI to detect AI writing. Pangram Labs, GPTZero, Originality.ai, Turnitin’s AI-detection module, and competitors. The detection economy is worth several hundred million dollars per year and growing fast.
AI to remove the AI tells. Sinceerly is the worked example, but it is one of many. Humaniser tools predate Sinceerly by at least eighteen months; what is new about Sinceerly is the explicitness of the framing and the founder’s willingness to lean into the joke. Pangram’s Max Spero, speaking to Koebler in May 2026, reports the humaniser pattern is now visible in the wild even outside dedicated tools: “AI-generated marketing emails over the last year with intentional typos.” The tells-removal practice has become standard in commercial AI-assisted writing.
AI to detect the de-tellers. This is already operational, not speculative. Spero confirms Pangram has been “collecting as much data as they can to try to detect ‘humanized’ AI” and acknowledges the arms-race shape: “it’s pretty adversarial” and “there is likely to be an ongoing cat-and-mouse game between humanizer AI and AI detecting AI.” The layer is in active production.

A fifth layer is structurally implied and almost certainly already in development somewhere: AI to remove the de-teller-detection signatures, that is, tools designed to humanise the output of humaniser tools so it cannot be detected as humanised. The stack does not stop. Each layer creates a market for the next, and the recursion has no natural terminus.

Why this is a worked example of a deeper problem

The Sinceerly Stack is funny and the founder is funny and the product is funny, and the Dictionary appreciates a well-executed joke. But the entry is not, in the end, about the joke. It is about what the stack reveals about the underlying economy of AI writing.

The stack is adversarial against itself, and the user pays at every layer. Each layer charges a fee or extracts attention. The user who writes a sincere email at layer zero — using their own brain, in their own voice, without any tooling — is not in the stack at all. The user who enters the stack at layer one (AI to write) is now, structurally, a candidate for every subsequent layer. Each layer is a tax on the previous one.

The stack does not produce trust. A reader who knows the Sinceerly Stack exists cannot, in fact, trust a slightly-typo-laden email more than a polished one. The typos may be authentic. The typos may be Sinceerly-CEO-mode. The stack has hollowed out the signal that informality used to carry — this person did not have time to polish, therefore the message is sincere — by automating informality. The stack increases the cognitive cost of every reader’s distinguishing labour. The Sinceerly Stack feeds the Zombie Internet, even when the individual tools at each layer are well-intentioned.

The stack benefits the platforms that run the stack, not the writers who participate in it. Anthropic, OpenAI, and Google sell tokens at every layer. Pangram and its competitors sell detection at layer two. Sinceerly and its competitors sell humanisation at layer three. The token spend per actual email sent rises with each additional layer the user adopts, and the recipient gets the same email a sincere writer could have written for free. The stack is, in macro, a transfer of value from writers to platforms.

The stack assumes the wrong winning condition. Each layer is structured around the question can the AI provenance be detected?, which is the wrong question. The right question, as the parent entry AI Writing argues, is would I want to be reading this if I knew? The stack does not answer that question. It cannot. It is a stack of detection-and-evasion tools, and detection-and-evasion is not the same thing as trustworthy communication. The stack is therefore solving the wrong problem at every layer.

Max Spero, Pangram Labs’s CEO and one of the people most directly profiting from the stack’s existence, said the quiet part out loud in the Koebler piece: “It’s kind of looking grim for the future of the internet.” This is not the founder of a flat-earth conspiracy commenting on geography. It is the CEO of an AI-detection firm commenting on the stack his own product sits inside. The structural shape is bad enough that even the people running it are saying so.

What the Dictionary recommends

Do not enter the stack.

The recommendation is structural, not moralising. The Dictionary does not claim Ben Horwitz is wrong to have built Sinceerly. It does not claim the Pangram Labs detection economy is fraudulent or that the humaniser-detection arms race is going to ruin civilisation. The recommendation is operational: the stack is bad for the participating writer, in ways that compound across the layers, and the writer’s best move is to opt out at layer zero.

Opting out at layer zero means: when an email or message or post or essay matters, write it yourself, in your own voice, with the cadence of your own reading voice in your head. Use AI tools where they make sense for the task — research, summarisation, copy-editing, brainstorming — but produce the final prose with your own attention on the page. This is the cooperative-writing practice named in AI Writing and detailed across this cluster. It is the alternative to the stack.

The stack will continue to exist. Many users will participate in it. The Dictionary’s bet, in AI Writing, is that the cooperative-writing top — writers whose writing is read, who care about their own register — will be more visible against a background dominated by the stack, not less. Sinceerly’s existence is, in a sense, evidence for the bet. A market for AI-to-undo-AI emerged because the AI default is recognisable enough to be worth undoing. The undoing layer is itself an admission that the AI default has a fingerprint, and the fingerprint matters to the reader, and the writer wants to hide it.

The simpler move is not to leave the fingerprint in the first place. Write the email yourself. Layer zero is the recommendation.

A small note in defence of Ben

The Dictionary has been hard on the Sinceerly Stack in this entry, but is not, in the end, hard on Ben Horwitz. Two observations in his defence:

The product is funny on purpose. Horwitz is a Harvard MBA who has cold-emailed five Fortune 500 CEOs in order to learn that they write like exhausted humans. He has packaged that learning as a product. The Sinceerly marketing page is more honest about what the product does than most AI-tooling marketing pages of 2026, because the founder is in on the joke. The product is a satire that also works as a product. This is a non-trivial achievement and the Dictionary respects it.

The product reveals the problem more clearly than serious products do. Pangram Labs and competitors are not in on the joke. They sell detection as if it were unambiguously good and structurally fair, when, as The Olang’ Trap documents, the detection has systematic bias. Sinceerly, by leaning into the absurdity, exposes the structural problem with the entire detection economy more cleanly than any serious analysis has. Sinceerly is not the problem. The detection economy that Sinceerly satirises is the problem. The Dictionary recommends opting out of both; it does not blame the satirist for the thing he is satirising.

A coda on the name

The name Sinceerly — with the doubled e — is doing more work than the founder may have intended.

Prof. Langenkamp’s essay The Sincere Society (Substack, May 2026) names the structural condition the Sinceerly Stack sits inside: feedback systems that have learned to reward the performance of a virtue rather than the virtue itself, in the way the 18th-century imperial Chinese examination system, the post-Revolutionary French Terror, and modern student-evaluation forms each, in their separate centuries, learned to reward the performance of cheng (誠 — alignment of inner state with outer expression) over the thing itself. The essay’s argument is that cheng in such systems is not a sentiment to be felt but an engineering requirement to be designed in — a structural property of the feedback loop, not a personal virtue of any one participant.

Sinceerly, the product, is precisely the failure mode the Sincere Society essay diagnoses, made into a Chrome extension. It names itself with a sincerity-marker (sincere) and a doubling cute-flag (the second e), because it has reduced sincerity to the stylistic-fingerprint problem — a thing addressable by typo-injection, em-dash-removal, and lowercase-coercion. It cannot do otherwise. A tool that fixes the fingerprint cannot fix the substance. If the underlying email is sincere, the AI fingerprint is a cosmetic concern. If the underlying email is insincere, removing the fingerprint does not make it sincere; it makes it insincere-without-the-fingerprint, which is worse, because it is now undetectable.

The satirist almost certainly knows this. The product almost certainly leans into it. Sinceerly is funny because the name is doing the diagnosis the founder did not have to write. It is sincerity-by-other-means, packaged as a $4.99-per-month subscription, with a free tier of three rewrites for the curious. The Dictionary’s recommendation, restated for this section: do not enter the stack. The simpler move is to write the email in a register where the question of fingerprint does not arise, because the fingerprint is yours, and yours alone, and that is the only durable kind of sincerity in a Zombie Internet. Sinceerly cannot give you that. Neither can anything else for sale.