Three years ago, Artur Zhdan was applying to U.S. colleges and using AI to help with his essays. Then detection tools started spreading across campuses. Essays got flagged. His applications were rejected.

No safety net, no obvious next move. Just a laptop and a problem he understood better than most.

GPTinf launched in late 2022 not from a business plan, but from that experience. People were already writing with AI, and the text often showed it. Flat sentences, repetitive structure, no real voice. Zhdan wanted to fix that. Today the platform has 25 million users across education, marketing, and professional writing, covering detection, humanization, plagiarism checking, and structured writing.

AI-assisted writing is now normal. So is skepticism toward it. GPTinf lives in that gap, helping writers produce content that is clear, coherent, and actually theirs. That raises real questions about how detection works, what rewriting does to meaning, and where editing ends and evasion begins.

We interviewed Zhdan about all of it. How the platform is built, what the AI can and cannot do, and what two years at this particular intersection has taught him.

The Problem

1. Before we get into the product itself, can you tell the story of how GPTinf started, and how your own experience with AI-generated text and detection tools pushed you to build it?

When I was applying to U.S. colleges in 2022, I used AI to help draft my essays — not to cheat, just to get past a blank page as a non-native English speaker. Then detection tools rolled out, my essays got flagged, and applications were rejected. That’s when I realized detection wasn’t just a policy debate — it had real consequences, and the tools making those decisions were far from perfect. I didn’t set out to build a company. I wanted to understand why AI text gets flagged, what makes it different from human writing, and whether you could close that gap without destroying meaning. That curiosity became GPTinf.

2. For readers who may not know GPTinf yet, how would you describe the product today and the main problem it is trying to solve in the AI era of writing?

GPTinf is a writing refinement platform for the AI era. AI-generated text has a recognizable fingerprint — uniform sentence length, inflated vocabulary, formulaic structure — and detection tools catch exactly that. Our platform helps writers check their text for AI signals, refine it so it reads naturally, and verify originality. We serve 25 million users across education, marketing, and professional writing. The common thread: they all use AI at some stage, and they all need the result to sound like them, not like a language model.

3. GPTinf covers several stages of the writing process, from drafting and editing to checking and final polish. Where does AI play the biggest role across that workflow?

AI plays the biggest role in detection and humanization. Our detector scores how likely a piece of text is AI-generated. The humanizer then rewrites flagged text to match how humans actually write — shorter sentences, natural word repetition, simpler vocabulary, varied rhythm. Think of it as two AI systems in tension — one that understands what makes text look AI-generated, and one that’s learned to fix it. If our own detector still flags the output, the humanization wasn’t good enough.

Humanization and Preserving Meaning

4. When GPTinf helps refine AI-generated text so it sounds more organic, coherent, and human, what is the system actually changing most: wording, syntax, rhythm, tone, structure, or something deeper?

All of those, but the most important changes happen at the statistical level — things a reader wouldn’t consciously notice but a detector absolutely would. AI text tends to be too uniform: sentence lengths cluster in a narrow band, vocabulary is artificially diverse because models avoid repetition, and the overall rhythm is flat. Our system shifts those patterns toward how humans actually write. It’s not about swapping a few words. It’s about making the entire texture of the writing feel organic.

5. How do you preserve authorial voice, credibility, and original meaning when the starting point is already AI-assisted writing?

Users have a “Freeze Keywords” feature to lock specific terms — names, technical vocabulary, citations — so those pass through untouched. After every humanization, we automatically run detection on the output so the user sees the before and after. It’s not a black box.

6. How do you evaluate whether that process is working well? What matters most internally: readability, semantic fidelity, tone, detection quality, or something else?

Two things gate everything: semantic fidelity and detection quality. If the rewrite loses meaning, it’s a failure regardless of how human it sounds. If it preserves meaning but still gets flagged, that’s also a failure. So the system is built around a dual standard — quality must clear a strict threshold before detection results even matter. Beyond that, we monitor length fidelity (output shouldn’t be noticeably shorter or longer than input) and benchmark against known patterns in human writing to make sure we’re trending in the right direction.

Reliability and Adaptation as AI Evolves

7. What typically triggers an update to GPTinf’s detection logic, and how often do those updates happen?

The most common trigger is a new generation of AI writing model. When a model like GPT-5.4 or Claude produces text with different properties than what our detector was created on, we see it in our benchmarks and recreat. We also monitor false positive rates on real-world human text — if a particular domain starts getting flagged more than it should, that tells us the detector needs more diverse training data. Major updates track meaningful shifts in the AI writing landscape. Smaller calibrations happen more frequently. We treat the detector as a living system, not a shipped product.

8. As AI writing models keep changing, what has become harder for GPTinf over time: detection, humanization, preserving nuance, or something else?

Both get harder, but for different reasons. On detection, each new generation of AI models writes more naturally — the surface-level gap between AI and human text is narrowing, so detectors have to rely on subtler signals. On humanization, some of the deepest signals are tied to how language models generate text at a fundamental level — properties of the generation process itself, not the words on the page.

9. False positives have become one of the biggest criticisms of AI detection. How does GPTinf handle cases where human-written text may still appear to show AI-like signals?

We take this very seriously because we’ve measured the problem firsthand. Early versions of our detector had unacceptably high false positive rates on real-world human text — casual writing, news articles, informal posts. The detector had learned what “benchmark” human text looks like, not what real writing across diverse contexts looks like. We use a conservative threshold: if the detector returns a low AI probability, we report the text as clean. We’d rather miss a borderline case than wrongly accuse someone.

Learning and Improvement

10. How do you define the line between helping people improve AI-assisted writing and helping them evade detection?

The line is whether the text gets better or just gets past a scanner. If someone’s output is clearer, more natural, and still says what they meant — that’s improvement. We can’t fully control intent, the same way spell-checkers can’t, but we control what the product optimizes for. Our humanizer has hard constraints on meaning preservation — it won’t hallucinate, drop facts, or add content that wasn’t there. If someone runs a generic AI essay through it, the output will read better, but it still won’t have their personal experience or their actual thesis. The people who use GPTinf most seriously are the ones who already have something to say.

11. How do you build a product that helps people improve AI-assisted writing without turning it into a tool for evasion?

By building around quality rather than bypass. Our humanization system is created so that it gets zero credit for beating the detector unless the output also passes a strict quality bar. On the product side, the workflow encourages engagement: users choose their style, lock keywords, and see before-and-after detection scores. It’s transparent by design. One-click bypass isn’t what the product is built to do.

12. What role does the user play in the loop? At what points does GPTinf rely on human judgment instead of automating the process entirely?

The user is in the loop at every decision point. They choose the output style, decide which keywords to freeze, see detection scores before and after, and decide whether to accept the result or run it again. We deliberately avoid a “magic button” design. Our system is good at fixing patterns that make text sound machine-generated, but whether the final text represents your argument and your voice — that’s a judgment only the writer can make.

What Comes Next

13. What is the hardest technical problem for GPTinf right now?

The hardest problem is what I’d call the deep statistical fingerprint of language models. There are properties of AI-generated text that are very difficult to change through rewriting — they’re tied to how the models generate text at an architectural level, not to word choice or sentence structure. We’re working on approaches beyond traditional rewriting to address these, and it’s an active research area for us. It’s also the kind of problem that keeps getting harder as detectors get more sophisticated.

14. Looking ahead, how do you want GPTinf to evolve as an AI product? Do you see it becoming mainly a detection tool, a writing assistant, a text-quality platform, or something broader?

We see GPTinf becoming a comprehensive text quality platform — not just detection or humanization, but the full pipeline of making AI-assisted writing genuinely good. We already offer style transformation, plagiarism checking, and a structured writing environment, and we plan to deepen each. The longer-term vision is a platform where a writer goes from a rough AI draft to a polished, original, voice-consistent piece — with every step transparent and the writer in control. We’re also expanding multilingual support beyond English. With 25 million users across very different use cases, the common need is the same: clear, original writing that’s genuinely theirs.

Origianl Creator: Genaro Palma
Original Link: https://justainews.com/technologies/natural-language-processing/from-flagged-essay-to-25-million-users-artur-zhdan-on-building-gptinf/
Originally Posted: Mon, 23 Mar 2026 12:07:12 +0000