Issue #360

Hemmingway AI Co-Founder: Seven Questions on AI Writing

Yashodan Naidoo answers seven questions on what separates human-sounding AI text from writing that actually helps readers.

AI & TechHemmingway AI Co-Founder: Seven Questions on AI Writing

A Model Developer Who Read Issue No. 277 Sent Me an Email

Reader, back on September 27 in Issue No.277, I covered a model called Hemmingway-1 — a 27-billion-parameter writing model trained to “write like a human.” I wrote then that its scores needed to be read with some caveats attached: the developer built and ran the benchmark itself, the grading was done by another AI model rather than a human, and the model itself is English-centric.

About a month after that piece ran, on October 21, Yashodan Naidoo — co-founder of Hemmingway AI, the company behind the model — sent me an email.

He read Issue No.277 as arguing that “writing should be judged by how easily a reader can understand and act on it,” and said that was a useful challenge to the way his own company has been describing itself as sounding “human.” Then he asked me one question: what everyday writing situation best captures the problem of reader effort?

My answer was meeting minutes. Minutes usually contain everything that matters — the decisions made, who owns the next step, the deadline — but it’s all buried inside the discussion log. If a model fills in an owner or a date that nobody actually agreed to, the minutes still read smoothly, but they send people in the wrong direction. The landlord text message in the Hemmingway-1 model card has the same shape. What the reader needs is the one line that tells them what to do next.

In that reply, I asked for an interview. I also mentioned something specific to Korea: disguising the fact that AI wrote a piece has become its own small skill — deleting em dashes, avoiding words like “chuk” (a stiff, formal word meaning roughly “axis”) or “cheung” (meaning “layer/floor”), running the draft through a “humanizer”1** tool before posting it**. That’s why several of the 7 questions point in that direction.

Yashodan agreed to the interview, and we conducted it in English. Below, I’ve set out the questions and answers in order. I translated his answers into Korean for Korean readers, adding in between the facts and my own thoughts that readers should know. I’ve also left in place the points flagged in Issue No.277 that he might not agree with.

Let me introduce the company briefly first. Hemmingway AI is registered in South Africa, with team members based in both South Africa and Switzerland. It ships a writing app together with its own model and an API. The first model, Hemmingway-1, launched on September 20, 2026, and is a further-trained version of Qwen3.8-27B. A follow-up model, Hemmingway-1.5, launched on October 25 and is now the default model for both the app and the API. Since I sent my questions on October 21, they don’t mention 1.5 at all.

One correction: in Issue No.277, I wrote that Hemmingway-1’s license was Apache-2.0, but the official model card now lists it as CC BY-NC 4.0. Hemmingway-2, set to release later, will follow the same licensing policy — free for non-commercial use, with commercial use requiring a separate agreement with the company.

The Goal of “Writing Like a Human” and How It Was Scored

Q1. Why did you set the goal as “write like a human”? Now that you’re looking at reader effort, how does that reshape the direction you’re training Hemmingway-2?

Most AI writing tools produce text that feels off once you actually place it in the context it was meant for. Tone and style matter when you’re building writing ability, but both sit beneath context. For instance, a message consoling someone who’s lost a loved one and a message updating that same person on a project’s progress can both be equally “clear,” yet they demand completely different sensibilities. Writing isn’t just about transmitting information — it’s about managing the relationship between the writer and the reader. (I came to Hemmingway from a background in music and writing, so I care as much about the effect a piece of writing leaves on its reader as about the information it carries.)

This is exactly what I want Hemmingway-2’s training and evaluation to focus on: understanding the situation the writing will be used in, preserving the content of the original draft, and producing writing that makes it easier for the recipient to grasp what matters. I expect some progress with Hemmingway-1.6, coming out this week, but the real leap forward will be Hemmingway-2.

He named three goals for Hemmingway-2: grasp the situation, preserve the content of the draft, and make it easy for the recipient to see what matters. The third is what I was asking about with reader effort; the second connects to the meeting-minutes story.

Q2. CommunicationBench, Human-Likeness, and StoryBench are benchmarks you built in-house, and they were graded by a model. Why a model judge instead of human evaluators?

What questions do you consider still open in that evaluation?

This question repeats, word for word, the condition I flagged in Issue #277. All three benchmarks were built by Hemmingway itself, and the method places two models’ answers side by side and has a different AI model — a judge model2 — pick the better one. As of November 10th, the official model card marks all three benchmarks as “internal” evaluations, with no indication of who scored them or how.

Going forward, I plan to include human evaluators alongside model judgments in the benchmarks we release. The main reason we keep using models as “judges” is that it lets a small team like ours realistically run repeated comparisons across multiple responses to the same prompt. It lets us quickly assess the impact of swapping out a model, and compare huge numbers of responses against the same set of guidelines. That’s genuinely valuable when you’re building a model. But that doesn’t make the judgment “objective,” and it’s no substitute for actual reader feedback.

In our initial evaluations, we masked which model produced each response and compared responses in both orderings (to reduce positional bias3). That’s a useful safeguard, but it doesn’t eliminate the judge model’s own preferences, or our own biases as the people who designed the evaluation (which tasks to include, for instance). A judge model might favor responses that use more common phrasing, might fail to catch fabricated facts, or might pick the response that sounds pleasant over the one that’s actually helpful. The “Human-Likeness” evaluation is especially weak on this front. Having a (different) model judge whether a piece of writing was written by a model or a human is a poor proxy for how an actual human would judge that writing. This is where I think human evaluators will need to take on a much more direct role going forward.

Yashodan didn’t deny either of the two conditions I raised in Issue #277. If anything, he spelled out, in concrete terms, what a judge model can miss: it might favor common phrasing, might fail to catch fabricated facts, might choose the answer that sounds good over the one that’s actually helpful. He went so far as to call the human-likeness evaluation “especially weak.”

But a benchmark that includes human evaluators is still a matter for “going forward.” Every score published today was scored by a model.

Where Does Erasing the Tells Diverge from Writing Well?

Q3. In Korea, people try to make writing look human by erasing surface-level tells—like em dashes or certain words. Is what Hemmingway does the same thing, or is there a difference between hiding the tells and actually writing well? Where do you draw that line?

Shortcuts like stripping out em dashes or specific words belong to the realm of humanizer tools, and honestly, that’s just a stopgap. Hemmingway is training the model to write well, which goes far beyond what a humanizer tool does.

What we care about more is this: Did this piece of writing understand the context it’s going to be used in? Does it sound right for this relationship? Is it saying what I (the sender) meant to say? Did it fill in details I (the sender) never authorized? A piece of text can pass every single test designed to catch “AI-written” words or punctuation and still fail on every one of these questions.

We do talk about writing that “sounds human.” But we don’t think being indistinguishable from a human on a detector score is enough, by itself, to claim that a piece of writing is accurate, useful, or honestly presented.

We want people to have tools that help them express themselves without sounding like every other piece of AI-generated text flooding the world. Good writing should multiply what someone is trying to say—something they’re prepared to stand behind. It can’t substitute for actually having something to say.

There’s a fact worth placing right next to this answer. Four days after I sent my question, Hemmingway-1.5 launched, and its official model card describes it as “the undetectable AI.” According to the company’s own testing, it passed GPTZero, Pangram, and four public detection models at a 90% rate.

Here, “passing” means the detector read the text as human-written with more than 90% confidence. By the same detectors, human-written text (1,670 samples) passed at an 86% rate, while Hemmingway-1 scored 7%. The model card lists one use case as “pasting in AI-written text to make it sound human.”

By the company’s own numbers, 14% of human-written text failed to pass. That’s the same reason I wrote, back in Issue 195, that detector scores shouldn’t be treated as evidence—and this figure bears it out.

There’s a gap between Yashodan saying detector scores “aren’t enough” and a product launch that leads with its detector pass rate. That doesn’t mean the answer was wrong—he said “not enough,” not “not necessary.” Still, this answer alone made it hard to draw a clean line between the Korean trick of erasing em dashes and a model fine-tuned to pass detectors.

Q4. Hemmingway-1 is English-centric. In Korean, a lot of the AI “tell” comes from English sentence structures bleeding into Korean. What would it take to build the same kind of model for Korean or other non-English languages? Is it on the roadmap?

Our first non-English language is likely to be something widely spoken, like Arabic or Spanish. I’m quite wary of taking a concept of “good writing” that works in English, translating example sentences into another language (say, Korean), and then assuming the result counts as “good writing” in Korean too. I don’t have the linguistic expertise to know what sounds natural in Korean. So we’d need to bring in native Korean writers to help define the task, select and validate the right training data, and evaluate the output. Language isn’t just grammar and vocabulary—it’s also relationships and context. We’d also want to make sure the model is equally good, in non-English languages, at generating new sentences while preserving the facts of the original draft—names, decisions, and so on. Success will come down to task design, training data curation, and the quality of evaluation.

There was no timeline for Korean in the answer. The candidates named for the first non-English language were Arabic and Spanish. But what the answer did spell out, specifically, was what it would take to build a Korean model: not translating English examples for reuse, but bringing in native Korean writers to handle task definition, data curation, and evaluation, all at once.

Responsibility, Small Models, and What Belongs to Humans

Q5. If AI-written text stops looking like AI-written text, readers lose one of the signals they used to gauge whether someone actually checked the content. Who should bear that responsibility—the people who built the model, the person sending the message, or the reader?

A tool (like an AI writing model) should produce the best output it’s capable of producing for its purpose. Deliberately making the output worse so people can tell it came from AI doesn’t make sense to me. Across writing, music, code, and video, I expect that the finished product will stop being a reliable signal of how it was made—and for many people, it’s already crossed that threshold. Responsibility for how a tool is used, and what gets put out into the world, has to rest with the person wielding it.

Our job as model builders is to shrink the distance between a thought and its finished expression. Writing a good story has traditionally required access to education and literature, plus time to practice. That access isn’t distributed equally. Someone working to make a living can still have something worth saying—they just never had the chance to develop the craft of writing.

We want those people to be able to produce high-quality work too. For people who’ve already honed that craft, we want the tool to help them go further, faster. More people should have the means to give their thinking a persuasive shape.

The question laid out three candidates for responsibility, and the answer pointed squarely at the person wielding the tool.

Q6. A 27B model beat out much larger models to take first place on everyday messaging. What does this tell us about narrowly specialized models versus general-purpose frontier models4?

My hypothesis is that it’s not that surprising for a model specialized in a specific domain to outperform a general-purpose frontier model. Focus matters, and sheer scale doesn’t automatically make up for its absence.

You see this constantly in business. Why can’t giant companies with far more money and talent beat every smaller competitor? Because having more resources and concentrating those resources on one specific problem are two different things. A small team can understand its domain more deeply and organize everything around doing that one thing well.

I think something similar applies to models. That’s what makes Hemmingway’s results interesting to me—they suggest that focused training can accomplish what a much larger general-purpose model consistently fails to.

You can imagine a future where a well-tuned system routes requests across a set of highly specialized models, and the user barely notices the model switching underneath. You’d get the strengths of multiple models inside one seamless experience. Whether it actually plays out that way remains to be seen, but I don’t assume the biggest model will always be the best model in every domain.

The ranking behind this question comes from Hemmingway’s own benchmark, CommunicationBench. By the model card’s numbers, Hemmingway-1 scored 1026 points, and the runner-up, Fable 5.1, scored 1024—a gap of just 2 points. Qwen3.8 27B, the base model underlying Hemmingway-1, scored 954. On StoryBench, which measures creative writing, Fable 5 Max (1277) and GLM-5.3 (1254) both outscored Hemmingway-1 (1197).

Q7. For Korean office workers who draft reports and messages with AI every day, what would you say should be handed off to the model, and what should still be written by hand?

Give the model as much context as you can. What do I know? What do I want the other person to understand? The better you do this, the better the output gets. I use models often when I need a structure suggested, redundancy trimmed, something explained more clearly, or help figuring out how to say something that’s hard to bring up.

Responsibility for recommendations, facts, and commitments sits with the user. If a meeting ended without agreement on who owns an action item or by when, the minutes should say so explicitly. Check numbers and sources—especially sources!—against the original material. Fluent writing isn’t the same as true writing.

I don’t really divide things up as “I’ll write this part myself.” There are plenty of times when I know exactly what I want to say but need help saying it. Before you hit send, make sure you can stand behind everything in the message. If someone questions a fact or holds you to a promise, you need to be ready to answer for it.

The point about meeting minutes comes back in this answer too. If the owner and deadline weren’t agreed on, write that they weren’t agreed on. My original worry, back in my first email, was that the model would plausibly paper over that blank space—but his answer put filling in the blank honestly as the writer’s job, not the model’s.

hemmingway founder interviewOswarld’s Lens

In issue No. 277, I put a qualifier on Hemmingway-1’s performance. Reading this round of answers, it’s the company’s co-founder who spells out that qualifier in the most detail. He wrote, in his own words, that the grading model can miss fabricated facts, and that it’s especially weak at judging “humanness.”

I think Answer 2 tells us more about this company than the fact that it ranks first on CommunicationBench. The ranking is right there on the model card, but what the benchmark fails to measure isn’t on the model card — it’s in this answer. Still, words and product need to be checked separately. Whether results involving human evaluators actually get published, and whether Hemmingway-1.5’s marketing copy leads with something other than its detector-passing rate, will tell us how much weight this answer deserves.

Answer 3, which calls humanizers a stopgap, runs in the same direction as what I wrote in issue No. 213. Back then I argued that hunting down lists of “how to scrub the AI smell off your writing” is beside the point. Deleting em dashes and avoiding words like “axis” and “layer” is just the latest version of that same list.

The four questions Yashodan offers instead aren’t only for AI-written text. Did it grasp the context? Does it fit the relationship? Is it what I meant to say? Did it include anything I didn’t authorize? You can apply the exact same questions when reviewing a human-written report or message.

The reason I opened with the meeting-minutes example is that minutes filled with owners and dates nobody actually agreed to can read perfectly smoothly — and still send people down the wrong path. His fourth question, and Answer 7’s line, “if it wasn’t agreed on, write that it wasn’t,” point to exactly that spot.

In Answer 5, though, he places the responsibility on the person wielding the tool. I agree that the sender has to own their own message. But making sure unauthorized content doesn’t get inserted in the first place is also the model-builder’s job. Given that he names “preserving the content of the draft” as a goal for Hemmingway-2, it looks like he, too, sees this as homework for his own team.

Answer 6 overlaps with something I say all the time: not every person, not every task, needs top-tier frontier-model-level intelligence — the right model depends on the job and the setting. I myself hand off long or repetitive work to DeepSeek or GLM these days, and reserve Claude or GPT-6 for overall planning and initial design work.

Still, it’s too early to use this particular result as evidence for that claim. The gap between first and second place is 2 points, and Yashodan himself explained, right there in Answer 2, the limits of the benchmark that produced that score. He called it a hypothesis himself.

Closing

Yashodan took the criticism in Issue No. 277 not as an attack but as the start of a conversation. Thanks to that, we were able to run his response, in which the developer directly addressed the weaknesses in his own evaluation. This interview is also being published in an English edition, and that edition carries his answer verbatim, in the original wording.


💬 What’s the last thing you check before sending a message you drafted with AI? Leave a line in the comments.

📨 If you have a colleague who drafts reports or messages with AI, pass this piece along to them.

Your take shapes the next issue

What resonated most in this issue, or where has your experience been different?

Any registered reader can comment for free.

References & Further Reading

Primary sources

  • Interview with Yashodan Naidoo (co-founder, Hemmingway AI), November 2026. ··· These are the seven answers featured in this issue. The original is in English; the Korean translation was done by INLEVEL9 Letter.
  • Hemmingway AI, “Hemmingway-1 model card”, accessed 2026.11.10. ··· This is the source for the CommunicationBench, Human-Likeness, and StoryBench scores, as well as the current license.
  • Hemmingway AI, “Hemmingway-1.5 model card”, accessed 2026.11.10. ··· This is the source for the 90% detector-pass rate, the 86% figure for human-written text, and the list of detectors and pass criteria used in testing.
  • Hemmingway AI, “What is Hemmingway AI?”, accessed 2026.11.10. ··· This lays out the company’s location, the release dates for each model, and Hemmingway-1’s 7% detector-pass rate.

Related past issues worth reading


Illustrated portrait of Kwangseob Ahn (Oswarld)

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

  1. Humanizer: A tool that rewrites AI-generated text to look more human, lowering its detection score. ↩

  2. Judge model: An AI model that reads answers written by other models and either scores them or picks the better of two. It’s faster and cheaper than human evaluators, but the judge model’s own preferences end up mixed into the scores. ↩

  3. Position bias: The tendency, when two answers are shown side by side, to favor whichever one appears first (or last) regardless of content. Running the comparison twice with the order swapped can reduce it. ↩

  4. Frontier model: A term for the large, general-purpose AI model that is currently the most advanced available at a given point in time. ↩