What Makes AI-Drafted Content Read Human: First-Hand Data

What actually makes AI content read machine-like, and can it be changed before drafting starts?

Fixed section shapes and generic material read the most machine-like in our data. The material, at least, is decided before drafting starts, so that part can change before a word is written.

A January 2026 Medium post on beating AI detectors documents one team's attempts. Their Claude pipeline, prompted with E-E-A-T principles, citation standards and formatting rules, produced output the author calls "excellent for SEO/GEO" but "terrible for humanness score" on GPTZero. A loop that regenerated whole drafts left the score at zero. Letting the model call GPTZero as a tool landed scores between 0 and 25. Rewriting only the flagged sentences, one at a time, worked best of the three. The author still says no method works consistently 90% of the time or more, and that "Tools like GPTZero, ZeroGPT, and Quillbot improve their models almost weekly, breaking what worked yesterday."

That test measured one commercial detector, not what makes text read human in general. Our own measurements point at structure and material. In 146 AI-drafted articles, 56 contained at least one passage that read human, but an article is judged by its most machine-like passage, so 144 of 146 were still called AI. The opening was the most machine-like part: 87% of first passages read AI, against 64-75% for later passages.

Some writers now change the first step instead of the last. Developer Elio Struyf describes in Getting interviewed by AI to write your technical content or blog posts how an AI interviewer, in sessions of about 10 minutes, drew out the material for three of his blog posts. His reason: "The writing process filters out a lot: the tangential insights, the 'oh by the way' details, the thought process behind technical decisions." He is careful to add that the approach does not automate writing ("it doesn't, not really").

This article separates what we measured from what we infer. Measured results are presented as findings; practical advice drawn from them is labelled as guidance.

How will the gap between human-reading and machine-reading content shift over the next 12-24 months?

We expect the gap to follow the material: sections built from real profiles, curated links or a company's own records should keep reading human more often than generic drafting. None of this is certain.

The starting point is measured. Author bios built from a real profile read human 51% of the time in our September 2026 scoring, calls to action 33% and curated resource lists 21%. Body sections, the bulk of every article, read human 7% of the time and AI 81% of the time. Those rates come from all 146 articles across 13 websites. The 37% against 71% result for record-heavy sections is a separate comparison, from one sports equipment manufacturer's site.

Why the model-light blocks read human is a plausible explanation rather than a finding: they carry text the model did not write from scratch. Whether that holds as models and detectors change is the forward-looking question. Our view, labelled as prediction, is below.

The forecast changes if new models start producing specific, varied prose by default, which would shrink the premium on first-hand records. Our September 2026 data does not show that yet.

Does the style signature of AI drafting explain what makes content read machine-like?

Only partly. Style habits such as dash rate are real and measurable, but removing them in our pipeline did not stop generated sections reading AI, and some human writing gets misread too.

Human writing is the first check. On 2,868 Medium posts archived before the end of 2020, the previous version of our detector called 4 AI (0.14%). Three were institutional prose (a government programme update, a health policy analysis and a corporate design case study), and one was a list of 100 headline-style post titles. None were AI-written. The current version calls 2 of the same 2,868.

Our research reads that pattern this way: the human writing most often mistaken for AI is neutral, well-organised institutional prose, because that is the register AI models imitate by default. That is a plausible explanation, not something four posts prove. It fits a second test, where the previous version called 8 of 637 US federal agency documents AI, above our 0.5% cap, before retraining brought it down to 2.

Style drift in AI drafts is also measurable. In September 2026, a fleet of AI-drafted articles contained zero em-dash characters, because a sanitizer had swapped each one for a spaced hyphen, yet carried 12.9 dash-punctuation marks per 1,000 words, against 0.8 per 1,000 words in one company's own pre-AI blog. After we replaced the writing rule with a budget of 2 dash marks per 1,000 words, new articles measured 1.69 dash marks per 1,000 words. That brings the surface closer to the human baseline. It does not change what the article is built on.

Cleanup of that kind changed the surface, not the verdict. Removing reused phrases, dashes and fixed skeletons made our articles stop looking alike, but 81-100% of every generated prose section type still read AI to a detector trained on the pipeline, with shared phrases at 0.5% or less. Outside research points the same way: Reinhart et al. (PNAS, February 2025) describe a noun-heavy, information-dense model style, with a larger gap from human writing for instruction-tuned models than for base models.

Readers lean on style cues too, and those cues misfire. In a Stanford HAI study (Hancock et al., March 2023), people judging dating, professional and hospitality profiles told human from AI text with 50-52% accuracy, and wrongly read grammatical correctness, first-person pronouns, family references and informal language as signs of a human writer.

We therefore report two readings separately. Origin is who most likely wrote the text: AI, a person, or not called. Style is how generic or distinctive the writing reads. The two do not move together. On a sports equipment manufacturer's site, 288 posts by one human author and 30 AI-drafted buying guides scored in a similar style band (57 against 68 on our style scale), while our detector flagged 0.7% of the human posts and 73% of the guides.

So style explains part of what reads machine-like, and editing style is legitimate work in a production pipeline. It does not settle authorship on its own, and it cannot add material that no generic draft could have included.

What does the detector data show about which AI-assisted articles actually read human?

Four open detectors called none of 223 held-out articles from our pipeline AI, yet a classifier trained on that pipeline's output separated them perfectly. Reading human depends on who is asking.

The open detectors were desklib, Fakespot, Binoculars and Fast-DetectGPT. For each, we set the threshold so that at most 1% (and, separately, 0.5%) of verified human documents would be called AI, then confirmed that rate on 8,540 other human documents. At both thresholds, 0 of 223 held-out pipeline articles were called AI, while the same thresholds caught 13-80% of ordinary public AI text (Fakespot 80%, desklib 66%). The detectors even scored our articles as more human than real human writing, with AUROC between 0.22 and 0.40.

Two limits belong next to that result. A classifier trained on our pipeline's own output separates the same articles perfectly (AUROC 1.000), so a learnable fingerprint exists and general-purpose open detectors simply do not carry it. And commercial detectors such as GPTZero and Pangram were not part of the test, so the result says nothing about them.

Binoculars and Fast-DetectGPT are zero-shot methods: instead of learning from labelled examples, they score how predictable a text looks to a language model. Our 30 August audit found every zero-shot method we tried blind to humanized output, which overshoots the human unpredictability band (Binoculars about 1.06 for humanized output, against about 1.00-1.05 for human text).

Our own detector is measured on people as well. Across 9,963 verified human documents it called 4 AI. On a pool of 4,546 human documents spanning 17 kinds of writing, it called none (95% upper bound 0.08%). The cap we hold it to is 0.5% of human documents in every kind of writing, not just on average.

Public advice lands on material too. The Substack post 4 Ways to Make Your AI Content More Human (and More Searchable) warns that "Raw AI content often creates a 'sea of sameness,'" and argues that original data earns links from other websites. That is the author's opinion rather than a measurement, but it points where our section data points.

What the data supports is narrow. The open-detector pass reflects articles humanized to the level of a well-edited human writer. It is not evidence about first-hand records, and it does not make the articles untraceable. Separately, on one site, sections built on the company's own records read AI 37% of the time to our detector, against 71% for the rest, while a site that used general industry figures showed 98% against 100%.

We publish these limits, including the human writing our own detector gets wrong, because a pass rate without them means little.

Why does AI-drafted content still read machine-like after humanizer passes?

Humanizing changes how text looks more than what it is built on. After our pipeline's humanizing, 144 of 146 September 2026 articles were still called AI by a detector trained on that pipeline's output.

The same articles did pass general-purpose detectors. At thresholds set on verified human text, at most 1 of the 146 was called AI, by the two zero-shot open detectors, and the two trained open classifiers flagged none. On 223 held-out articles, a classifier trained on our pipeline's output still separated them perfectly (AUROC 1.000), and commercial detectors such as GPTZero and Pangram were not tested. So "still reads machine-like" depends on the detector: the one that knows the pipeline recognises it, as of September 2026.

The humanizer market works on a simple theory: AI content reads robotic because of surface patterns, so replace those patterns. Swap flagged vocabulary, vary sentence length, add a contraction. A June 2026 Substack roundup of five humanizer tools judges them on whether they preserve meaning, improve flow, reduce robotic phrasing and save editing time, and its author cautions that "one good output does not prove consistency."

Call this the substance test: remove the brand name from a passage and ask whether it could appear word for word on a competitor's site. If it could, surface editing has little to work with. That is editorial guidance, not a measurement, and it does not predict a detector score.

The strongest evidence behind it comes from one site. For a sports equipment manufacturer whose AI-drafted articles drew on its own sales and shipping records, such as orders per state and models shipped per county, record-heavy sections read AI 37% of the time. The other sections of the same articles read AI 71% of the time. Same pipeline, same articles, different material underneath.

The limit is just as clear. For a payments company whose articles used general industry figures, the same comparison came out at 98% against 100%, a gap of two percentage points rather than thirty-four. We did not split all 2,625 sections by type of evidence; this is one strong site result with a clear counterexample.

Across body sections from all 146 articles, the sections that read human carried 7.9 numbers per 100 words, against 2.5 numbers per 100 words in sections that read AI. Our research does not read that as "use more numbers": the payments company's figure-rich articles barely differed. A plausible explanation is that figures only one company could publish did the work, not figures as such.

The problem practitioners describe may be two problems. One is packaging: the surface habits a model brings to any draft. The other is substance: text that anyone could have written, because it rests on information available to anyone. Humanizer tools are built for the first. Our data suggests, without proving it, that the second needs different inputs.

This article reports what we measured and where the limits are. The section scores cover 146 articles for 13 websites; the first-hand records comparison comes from one site, with the payments company as its limit. Where we offer practical guidance, we say so.

What should the evidence change about how you approach AI-drafted content?

Start with what the draft is built on. On one site, record-based sections read AI 37% of the time against 71% for the rest; general industry figures elsewhere showed 98% against 100%.

Humanizing and voice are separate jobs. As a Substack guide on humanizing AI writing puts it, "A humanizer strips out the machine tics. It doesn't give you a voice. Those are two different jobs." A pipeline that treats them as one job will serve one and neglect the other.

A practical checklist (guidance, not findings):

• Collect first-hand records before drafting: your own counts, dates, places and outcomes, the facts only your company could publish.
• Draft on that material, then edit the surface. Do not expect a rewrite pass to add specificity that was never collected.
• Apply the substance test: remove the brand name and ask whether a passage could sit on a competitor's site.
• Give extra attention to openings and fixed shapes such as decision frameworks, outlook charts and myth-versus-fact boxes, which read the most machine-like in our data.
• Build author bios from a real profile rather than a generated summary.
• Check a detector's false-positive rate on human writing before trusting its verdict, and do not treat a pass on one detector as a pass on all of them.

That sequence is where the AEO Content Engine starts: articles grounded in each company's own records, written to read like expert human writing and structured so AI answer engines can cite them. To see which passages of a draft read machine-written, run it through the free AEO Content AI Detector.

This article is part of our research series on how AI writes and how humans write. The overview of the whole series is How AI Writes vs How Humans Write.