We’ve read so much about how AI is being used in writing, how it has impacted students and professors and how AI is used unethically in most cases. Almost every claim made about AI and writing today rests on one of two things: a headline number with no methodology behind it, or an anecdote from a single classroom. This report is an attempt to do better, using two independent sources of evidence that Quetext is unusually well positioned to bring together.
The first is anonymized data. Quetext processes millions of documents each year, and this report analyzes aggregated, anonymized results from ~500K AI checks and ~500K plagiarism checks, comparing each month with the same month in the previous year. Please bear in mind, no individual users or documents were identified or reviewed. Instead, the analysis focuses on overall patterns in how AI-assisted writing and plagiarism changed over time.
The second is a survey. We divided the audience into 3 segments: students, educators and professionals. We selected these segments to get an answer from three different vantage points: students who write the assignments, educators who assess them, and professionals who write as part of their paid work. The survey is what makes the data legible. Scan data can tell you that a flag rate moved; only people can tell you why they think it moved, what they were doing with AI, and what it cost them when someone got the call wrong.
Executive summary
The flag rate went down. The AI did not go away.
In 2026, 84.64% of documents submitted were flagged at 80% or higher on our AI-likelihood scale. That is a large number by any standard, roughly five in every six documents. It is also 12.4 percentage points lower than 2025, when the figure was 97.04%.
In contrast to expectation, a falling flag rate suggests an easy headline: AI writing is in retreat. Our data does not support that reading, and our survey actively contradicts it. Among the writers we surveyed, AI use looks anything but receding. Most students use AI often or for most writing tasks. Nine in ten of the professionals we surveyed do. And the educators assessing that work are far more likely to say suspected AI use has risen over the past year than to say it has fallen. What appears to have changed is not whether people use AI, but how visibly.
The key finding is that AI-assisted writing and plagiarism are now two different issues. In 2026, AI flags were about seven times more common than plagiarism flags and were higher every month. This means they should not be treated as one problem. Institutions that use plagiarism alone to measure AI use are likely missing what’s really happening.
The six findings that matter · 6 findings
01
The decline is real, but it is a decline from near-universal.
84.64% still means about five out of six documents containing AI-assisted writing. The use of AI hasn’t disappeared — it has declined from the very high levels seen in 2025. The gap grew throughout the year, from slightly higher than 2025 in January to much lower by September, with May the only month that stayed about the same.
02
AI detection and plagiarism detection now measure different problems.
A 72.5-point gap between the two flag rates across the year, with AI flags exceeding plagiarism flags in every month, is not noise. Plagiarism-style similarity looks for text that already exists somewhere else. AI likelihood estimates how machine-generated a passage appears. In 2026 these moved independently, and they peaked in different months.
03
Flags are falling while AI use stays routine.
Most students told us they use AI often, and nine in ten professionals said the same. Educators, meanwhile, were four times as likely to say AI use had gone up over the past year as to say it had gone down. Set against a falling flag rate, this is the report’s central tension. Several explanations are consistent with both datasets; the data cannot arbitrate between them, and we do not pretend otherwise.
04
Using AI and passing it off are two very different populations.
Students estimated that around two-thirds of their cohort uses AI on assignments, and that resonated well with the survey’s actual responses. But only one in four says they have ever submitted AI-generated text as their own original work. Broad assistance is near-universal; passing work off is a much smaller behavior. Educators, for their part, estimate that 41% of what they receive is undisclosed AI.
05
The false-positive problem is now large enough to matter on its own.
More than four in ten students have been suspected or accused of using AI. Most of them say they had not used AI for the assignment in question. On the other side of the desk, a majority of educators either admit to having wrongly suspected a student or cannot rule it out.
06
Almost nobody has been given a clear rule about using AI.
This is the most important piece of the puzzle. Most students, educators, and professionals say they still don’t have clear rules about how AI can be used in their workflows. Students who were clearly told what was allowed were about one-third as likely to submit AI-generated work as their own compared with students who received no guidance.
Part I
The Signal: What 1 Million Scans Show
Throughout, a document is counted as AI-flagged if it scored 80% or higher on our AI-likelihood scale, and plagiarism-flagged if it crossed a 20% similarity threshold. Year-over-year comparisons use matched calendar months.
· 6 sections
1.1How common is AI-flagged writing?
Of the 500K documents run through AI detection in 2026, 450K scored 80% or higher on Quetext’s AI-likelihood scale, an 84.64% flag rate. That is down from 97.04% in 2025, a drop of 12.4 percentage points, or a 12.78% relative decline.
The interesting thing is that the rate did not fall evenly. The first two months of 2026 were nearly identical to the same months a year earlier; January actually ran slightly ahead. From March onward the two years pulled apart.
2025–20262024–2025
Figure 2. Monthly AI flag rate, 2026 vs. 2025. Shaded area shows the year-over-year gap.
Professional writer / copywriter, Quetext Writing Integrity Survey
That single sentence describes a behavior that no scan can see directly: a writer spending real time not on the writing, but on the appearance of the writing. If this behavior holds at large, a falling flag rate measures something quite different from falling use of AI.
Key takeaways
AI flags fell 12.4 percentage points year over year, from 97.04% to 84.64%, but 2026 remained high in absolute terms.
The two years were nearly identical in January and February, then diverged sharply from March onward.
Self-reported AI use is high and routine, and educators believe suspected use has risen, which rules out “less AI is being used” as a self-evident explanation for the decline.
As users become more experienced with AI tools and spend more time editing, rewriting, and personalizing AI-generated drafts, the resulting writing is less likely to trigger a high AI score, even though AI continues to play a major role in the writing process.
1.2Two checks, two different problems
Plagiarism-style similarity checks flagged a far smaller share of documents than AI checks did, in every single month of 2025–2026. Averaged across the year, 12.10% of plagiarism-checked documents crossed the 20% similarity threshold, against 84.64% of AI-checked documents crossing the 80% AI-likelihood threshold. The average similarity score was 9.48, comfortably below the flag line.
Consistency matters more than the size of the gap. AI flags were higher than plagiarism flags in every month of the year, often by a wide margin. This wasn’t a one-time spike or a seasonal trend — it shows that AI detection and plagiarism detection are measuring two different behaviors.
Why does this matter? Because a lot of institutional policy still treats “plagiarism” as the umbrella category and AI as a subspecies of it. The 2026 data suggests this is far from reality. Copying is a text-provenance problem: the words exist somewhere else and can be located. Machine assistance is an authorship problem: the words may be entirely novel and still not be the writer’s own thinking. A misconduct process designed around the first will systematically mishandle the second.
Figure 3. AI flag rate and plagiarism flag rate, 2025–2026 annual averages, with the monthly range for plagiarism flags.
Survey insight
The people being surveyed don’t think it’s one category either
When asked whether using AI to write an assignment counts as cheating, a majority of students chose neither “yes” nor “no” but “depends on how it’s used.”Professionals said the same thing.
Around one third of students said outright that it is not cheating, and around one in seven said it is. Almost nobody treats it as the settled binary that plagiarism has been for a century. The data and the survey are describing the same shift from two directions: the old category has stopped doing the work.
Key takeaways
AI flags ran about seven times higher than plagiarism flags across 2026, and exceeded them in every month of the year.
Average similarity across all plagiarism checks was 9.48, well under the 20-point flag threshold.
Both audiences most likely to be judged — students and working professionals — reject a binary definition of AI cheating, choosing “depends on how it’s used” over yes or no.
1.3Rethinking when AI-flagged writing peaks
Going in, we expected AI-assisted writing to track the academic calendar, spiking around December finals and the May end-of-term crunch. The 2025–2026 data supports half of that and flatly contradicts the other half.
May was indeed the peak month for AI flags, at 99.23%, almost exactly matching May 2025’s 99.22%, making it the one month of the year with essentially no year-over-year change. December, however, was the year’s lowest month, not its highest, at 75.35%.
Plagiarism flags kept their own separate calendar entirely. January was the peak month at 18.61%; November was the low point at 8.96%. The two metrics did not peak together, did not trough together, and did not move together — further evidence that they are tracking different behaviors rather than two symptoms of one.
AI flag rate
Highest monthMay 2026 · 99.23%
Lowest monthDec 2025 · 75.35%
Plagiarism flag rate
Highest monthJan 2026 · 18.61%
Lowest monthNov 2025 · 8.96%
This matters for planning. Many institutions focus extra integrity efforts in December, but our data shows December had the lowest AI flag rate of the year. If AI use is the main concern, more attention should go to May. If plagiarism is the concern, January is the better time to focus.
1.4The gap widened through the first three quarters
Figure 4. Year-over-year change in AI flag rate by month, in percentage points.
Key takeaways
The year-over-year gap widened from +2.26 points in January to −24.13 points in September — a 26-point swing — before narrowing slightly in the final quarter.
September 2025 recorded the sharpest single-month year-over-year drop in the dataset.
The gradual shape of the decline is more consistent with a progressive change than with a single step change, though the data cannot identify a cause.
1.5Account type changes the shape of the risk
Different user groups show distinct patterns in how AI-assisted writing and plagiarism appear. Rather than following a single trend, each segment reflects different ways these technologies are being used.
Agency users recorded the highest rate of AI-assisted writing (91.28%), while also showing the lowest plagiarism rate (1.92%) and the lowest average similarity score (3.36%). This suggests content that is heavily AI-assisted but largely original, with little overlap with existing published material.
Professional users showed a different pattern. They had the highest plagiarism flag rate (17.83%) and the highest average similarity score (16.25%), while still maintaining a high AI-assisted writing rate. This indicates that AI use and content similarity can coexist and should be evaluated independently.
Teacher accounts are worth a look. They recorded the lowest AI-assisted writing rate (79.39%) but one of the highest plagiarism rates (16.97%). These results likely reflect the types of documents educators review and evaluate, rather than their own writing.
More broadly, the findings highlight that different user groups interact with AI and plagiarism detection in different ways, making context essential when interpreting integrity data.
AI flag ratePlagiarism flag rate
Figure 6. AI and plagiarism flag rates by account segment.
Figure 7. Full segment data
Segment
AI flag rate
Plag. flag rate
Avg. similarity
Student
86.17%
13.33%
9.28
Professional
82.76%
9.25%
8.40
Copywriter
84.62%
15.47%
11.30
Teacher
79.39%
16.97%
11.37
Business
90.82%
17.83%
16.25
Organization
87.62%
12.96%
10.78
Agency
91.28%
1.92%
3.36
Occupation is a self-reported, user-level proxy — not a document-type field.
Survey insight
Heavy use of AI and textual originality now coexist
Comparing user segments with our survey reveals an important pattern. The professional segment recorded the highest plagiarism rate and similarity score, while professionals in our survey were also the heaviest AI users — nine in ten reported using AI often or for most writing tasks, compared with roughly two-thirds of students.
These findings are not contradictory. Instead, they show how writing has changed. AI-assisted content can still be highly original, with little or no similarity to existing published material. As AI adoption grows, originality and AI use are no longer opposites, making it increasingly important to evaluate them as separate signals.
Key takeaways
The agency segment recorded the highest AI-assisted writing rate (91.28%) and the lowest plagiarism rate (1.92%), showing that high AI use can coexist with highly original content.
The professional segment recorded both the highest plagiarism rate (17.83%) and the highest average similarity score (16.25%), demonstrating that AI-assisted writing and content similarity are separate dimensions that can occur together.
Different user segments show distinct patterns of AI use and plagiarism, reinforcing that these should be measured independently rather than treated as the same issue.
1.6The shape of writing on the platform
Most submissions are medium-length documents rather than very short responses or long research papers. Around seven in ten fall between 150 and 1,200 words, with nearly half in the 150–600-word range alone. Very short submissions make up only a small share of overall volume, while documents longer than 1,200 words account for less than one-third of submissions.
This distribution highlights where writing integrity efforts matter most. The majority of checks involve everyday writing — assignments, discussion posts, blog articles, and business content — rather than lengthy dissertations or research papers. As AI-assisted writing becomes increasingly common in these routine writing tasks, institutions and organizations should ensure their policies and detection strategies address the documents people produce most often, not just the longest ones.
Figure 8. Distribution of submissions by length category across the Quetext platform. Word totals are reported as recorded and are aggregate figures; they are not directly comparable between categories.
Submissions and word totals by length category
Length category
Share
Total words
0–150 words (short answers)
1.3%
212,900,694
150–600 words (mini-essays)
49.4%
11,911,310,026
600–1,200 words (standard essays)
21.3%
3,231,308,873
1,200–2,000 words (extended essays)
11.9%
1,252,735,565
2,000–4,000 words (research papers)
6.4%
1,324,837,534
4,000–10,000 words (long-form academic)
9.8%
7,504,472,463
All categories
100%
25,437,565,155
Survey insight
This is the writing AI is used for
The survey maps almost exactly onto that distribution. The single most common assignment type students reported using AI for was routine homework and assignments, ahead of essays and research papers, which tied for second. These are medium-length, high-frequency, comparatively low-stakes pieces — precisely the band that dominates platform volume.
Scholarship and application essays — the high-stakes, career-defining documents that generate most of the public anxiety — were the least common use case students named. The integrity conversation is loudest about the writing where AI is used least.
Key takeaways
Roughly seven in ten submissions fall in the 150-to-1,200-word band; about half sit in the 150-to-600-word category alone.
Everything longer than 1,200 words — extended essays, research papers and long-form academic work together — accounts for under three in ten submissions.
Students report using AI most on routine homework and assignments, and least on scholarship and application essays.
Part II
The Behavior: Who Uses AI, and What For
AI-assisted writing is no longer defined by whether people use it, but by how they use it. To understand the behaviors behind the platform trends, we looked beyond detection data and explored how students, educators, and professionals use AI in their everyday writing. Their responses reveal when people rely on AI, where they draw the line between assistance and misconduct, and how expectations differ across education and the workplace.
· 2 sections
2.1AI use is no longer occasional
The most striking thing about AI adoption among the writers we surveyed is how unremarkable it has become. Among students, roughly two in three use AI often or for most writing tasks, and all but one respondent had used it at some point in the past year. Among professionals the figure climbs to nine in ten — and nearly half say they now use AI for most of their writing, full stop.
Within the student group, adoption rises with academic level rather than falling. Graduate and postgraduate students were the heaviest users of any cohort, with around three-quarters reporting frequent or near-constant use. The intuition that AI writing is primarily a problem of younger, less experienced students is not supported here; if anything the pattern runs the other way.
Tool use is plural rather than exclusive. ChatGPT was the most widely named assistant in both groups — roughly two-thirds of students and four-fifths of professionals — but Claude and Gemini were each named by around half of students, and Claude by six in ten professionals. A meaningful minority named Perplexity or something else again. Most writers are not choosing a tool; they are assembling a stack.
Key takeaways
Around two-thirds of students and nine in ten professionals use AI often or for most writing tasks.
Graduate and postgraduate students were the heaviest users of any student cohort, at roughly three-quarters.
Writers use multiple assistants rather than one: ChatGPT leads, but Claude and Gemini were each named by about half of respondents.
2.2AI is being used to think, not to ghostwrite
If AI were principally a ghostwriting technology, “generate first drafts” would top the list of use cases. It does not — in either audience. The most-selected use among students was brainstorming ideas, followed by improving grammar, summarizing information and rewriting or editing existing work. Professionals gave nearly the same ranking, with brainstorming first and saving time, grammar and comprehension close behind.
Generating a first draft sat in the bottom half of the list for both groups. So did overcoming writer’s block among students. The dominant pattern is assistive rather than substitutive: AI is being used at the edges of the writing process — before it, to think; and after it, to tidy — more than in the middle of it, to produce.
The free-text responses reinforce this to an almost monotonous degree. Asked in one sentence how AI had changed writing for them, respondents overwhelmingly described comprehension, structure and confidence rather than output. Several described using AI to check the logic of an argument they had already made. One postgraduate student described the effect as being able to write “more profoundly, critically, scientifically” — the opposite of the outsourcing narrative.
“It organizes my jumbled thoughts into a coherent text.”
Undergraduate, STEM
“AI speeds up my early-stage work and structures my scattered thoughts, but my skills in judgment, deep editing, and adding lived experience remain irreplaceable.”
Professional writer / copywriter
It would be naïve to take self-report at face value here; people describe their own AI use in the most defensible available terms, and the same survey contains a quarter of students admitting to submitting machine-generated text as their own. But the consistency of the pattern across two very different audiences, and its alignment with the platform finding that heavy-AI segments produce highly original text, makes it hard to dismiss. Substitution exists. Assistance is more common.
Why this matters
A policy built on the premise that AI use means AI authorship will misclassify the majority of what is actually happening. The writer who used AI to interrogate the structure of their own argument and the writer who pasted in a generated essay both produce a document that may be flagged — but they have not done the same thing, and no reasonable institution wants to treat them identically. Distinguishing between them is a question of assignment design and disclosure, not detection.
Key takeaways
Brainstorming was the top use case for both students and professionals; generating first drafts ranked in the bottom half for both.
Comprehension and structure dominate the free-text responses, not output generation.
Assistive use appears far more common than substitutive use — but a quarter of students separately admit to submitting AI text as their own, so both exist.
Part III
The Trust Problem
If Part II describes what people do, this part describes what they believe about each other — and it is here that the survey becomes uncomfortable reading. The problem is not that anyone is badly informed. It is that three audiences are looking at the same behavior, measuring it differently, and arriving at conclusions that cannot all be acted on at once.
· 4 sections
3.1Using AI and passing it off are not the same population
Students see AI use as common among their peers, with two-thirds estimating that at least 61% of classmates use AI for assignments. Their perceptions closely match their own behavior: about two-thirds say they use AI often and for homework or assignments. However, using AI is different from claiming AI-generated work as original. While AI assistance is widespread, only one in four students reported submitting AI-generated text as their own, showing that most students still distinguish between using AI as a tool and presenting its output as their own work.
Figure 11. Students’ own AI use, their estimate of classmates’ use, and their reported use of AI text as original work. Values approximate — reconstructed from the proportions reported in the text.
This is where the audiences diverge. We asked educators to estimate what share of submitted student writing involves undisclosed AI — the closest available analogue to the students’ own admission. Two-thirds of educators put that figure at 41% or higher. Students put it, of themselves, at around a quarter. The two questions are not perfectly comparable — one asks about a share of documents, the other about a share of people, and self-report on misconduct will always understate — but a gap of that size is not closed by definitional slippage alone.
One further signal is worth noting, with appropriate caution. Students who gave the very highest estimates of classmate AI use were also the most likely to report having done it themselves — though the relationship is uneven across the middle of the range and rests on small groups. To the extent it holds, it is what a projection effect looks like: people calibrating their estimate of the room partly from their own conduct.
Why this matters
The risk here is a category error with real consequences. An educator who reads “most students use AI” as “most students are cheating” will approach every submission as a probable violation — and on this evidence will often be wrong, because the broad behavior is far more common than the specific one. A student who knows almost everyone uses AI, and has never been told where the line falls, faces a straightforward incentive problem: the cautious route starts to feel like a penalty. Neither reading is being corrected by evidence, because almost no institution is measuring this and reporting it back to the people involved.
Key takeaways
Around two-thirds of students estimate most classmates use AI on assignments — and around two-thirds of them do. The estimate is accurate.
One in four students says they have submitted AI-generated text as their own original work — a far smaller behavior than AI use in general.
Two-thirds of educators estimate undisclosed AI in at least 41% of student writing — well above what students admit to, even allowing for under-reporting.
3.2Nobody agrees where the line is
Plagiarism has a definition. It has had one for a very long time, it is written into handbooks, and while people argue about its application they rarely argue about its meaning. AI has no such settled definition, and our survey suggests the disagreement is not at the margins — it is at the center.
Asked directly whether using AI to write an assignment counts as cheating, a majority of students chose “depends on how it’s used.” Roughly a third said plainly that it is not cheating. Only around one in seven said it is. Professionals, asked the equivalent question about undisclosed client work, produced the same ordering — “depends” first, “no” second, “yes” last — though with a larger share willing to call it cheating outright.
This is a definitional vacuum, and it is the single most consequential finding in the survey for anyone writing policy. An institution can enforce a rule that people dispute. It cannot enforce a rule that people cannot state.
Disclosure is where the vacuum becomes visible
Among the professionals we surveyed, the ambiguity has a concrete commercial expression. Only around one in six said they always disclose AI use to clients. Nearly half disclose sometimes. A quarter answered, revealingly, that clients simply don’t ask — a non-answer that is also an accurate description of the market.
Put differently: most of the professional writing produced with AI assistance reaches a client without a clear statement of how it was made, and in most cases without anyone on either side having established what the expectation was. That is not primarily a story about writers being evasive. It is a story about a service category that has not yet developed a disclosure norm.
Survey insight
A norm cannot form if nobody states one
The professional responses show how a disclosure vacuum sustains itself. Among writers whose clients had set out clear AI expectations, and among those left entirely to guess, the practice looked much the same — in both groups a substantial share said their clients simply never raise the subject. Clarity that exists on paper but is never invoked in the working relationship does not appear to change what gets disclosed.
That is a meaningfully different situation from the one in education, where students who were told clearly what was allowed reported markedly different conduct. In professional writing there is not yet a shared expectation for a policy to attach itself to. Someone has to ask the question before an answer can become a norm.
Key takeaways
“Depends on how it’s used” was the most common answer on whether AI writing counts as cheating — for students and professionals alike.
Only about one in six professionals always discloses AI use to clients; a quarter say clients simply don’t ask.
The definitional gap, not the detection gap, is what makes AI policy hard to enforce.
3.3The false-positive tax
More than four in ten students in our survey have been suspected or accused of using AI to complete an assignment. That number alone is worth pausing on: an accusation of academic dishonesty is not a routine administrative event, and it is now a common experience.
The composition of that group is what makes it serious. Slightly more students said they had been accused when they had not used AI than said they had been accused when they had. Among everyone who has faced an accusation, more than half maintain it was wrong.
Figure 12. Student experience of AI accusations, as a share of all student respondents. Values approximate — reconstructed from the proportions reported in the text.
The educator side of the survey corroborates the pattern rather than disputing it. One in five educators said outright that they had wrongly suspected or accused a student of using AI. A further third said they were not sure — which, given that a wrongly accused student has every incentive to maintain innocence and no way to prove it, is close to an admission that the question is unanswerable from where they sit. Taken together, a majority of educators cannot rule out having got it wrong at least once.
One further detail deserves care, because it is easily overstated. In our survey, the educators who reported having wrongly accused a student were all regular users of AI-detection tools. The number of educators involved is small and this is a correlation in a modest sample, not evidence that detection tools cause false accusations — regular users also handle more cases and may simply be more alert to their own error rate. But it is a finding that ought to make any institution ask a question it probably has not asked: what is our false-positive rate, and who is tracking it?
“Have moved to authentic evaluations (live in person) as much as possible, written work is suspect.”
College / university instructor, 11–20 years teaching
Why this matters
Detection error is not symmetrical in its consequences. A missed case of AI use costs an institution a marginal amount of academic integrity. A false accusation costs a specific student a grade, a record, a relationship with an instructor, and — on the evidence of these responses — a lasting sense that their honest work will not be believed. When roughly one in five students reports being wrongly accused, the accusation process has itself become a source of the mistrust it was designed to remedy.
Key takeaways
More than four in ten students have faced an AI accusation; more than half of those say they had not used AI.
One in five educators admits to having wrongly suspected a student; a further third cannot be sure.
No institution in this survey appeared to be measuring its own false-positive rate.
3.4The concern gradient
We put the same statement to all three audiences — “AI is eroding trust in the authenticity of written work” — and asked them to rate their agreement. The results form an almost perfectly ordered gradient, and it is not a gradient of age, seniority or technical sophistication. It is a gradient of proximity to judgment.
Educators — the people who must decide whether to believe a piece of writing — are the most concerned by a wide margin; nearly nine in ten rated their agreement in the top two points of the scale. Students, who are both judged and users, sit in the middle. Professionals, the heaviest AI users of the three and the group least often subject to a verdict, are the least concerned.
The tempting reading is that professionals are complacent. A fairer one is that they are describing a different market. Professional writing is judged on whether it works, by clients who mostly are not asking how it was made. Academic writing is judged on whether it evidences a student’s own thinking, which makes provenance the entire point. The same technology genuinely does pose a smaller problem in one setting than the other.
There is a warning inside that comfort, though. Professionals were also the group least likely to have been given any guidance at all, and the group whose disclosure practice is least defined. A low sense of risk combined with a near-total absence of governance is not a stable position for an industry to hold indefinitely.
Key takeaways
Concern about eroded trust falls steadily from educators to students to working professionals.
Nearly nine in ten educators rated their agreement in the top two points of the scale.
The gradient tracks proximity to judgment, not familiarity with AI — the heaviest users are the least alarmed.
Part IV
The Governance Gap
Two years into a technology that every audience in this survey now uses or encounters routinely, the most common experience across all three is still the absence of a clear rule. This part looks at what institutions and employers have actually put in writing — and at what measurably changes when they do.
· 3 sections
4.1Most people have never been given a rule
The governance gap is the one finding that unites all three audiences. A majority of educators said their institution has no finished, written AI-use policy — some are in development, some do not exist, and some educators simply do not know whether one exists, which for practical purposes amounts to the same thing. A majority of students said they had not been told clearly what AI use is allowed. And seven in ten professionals said no client or employer had given them a clear rule.
Figure 14. Share of each audience reporting no clear, finished rule on permitted AI use. Values approximate — reconstructed from the proportions reported in the text.
Professionals are the least governed group of the three, which is notable given that they are the heaviest users and the ones handling paid client work. But the educator figure is the more surprising one. Institutions have had two full academic years to write this down. A majority still have not finished.
Meanwhile the assessment practices have changed anyway. Roughly three-quarters of the educators we surveyed have already altered how they assign or grade writing because of AI — more in-class writing, oral defenses, live assessment, assignments deliberately tied to personal or local context. Practice is running well ahead of policy, which means the rules students actually encounter are being set individually, classroom by classroom, with no consistency and no appeal.
“It’s shifted my focus from policing grammar to coaching the thinking process, since AI now handles the mechanics, so I assess drafts, revisions, and critical reasoning instead of just the final product.”
Tutor, 11–20 years teaching
“AI forced me to change pedagogy… For assessment, giving topics/questions related to their everyday life which dissuade them from AI use.”
Researcher, more than 20 years teaching
These are not defensive responses; they are pedagogical ones, and they point at the most durable answer anyone in this survey offered. An assignment that cannot be completed well without the student’s own reasoning, context or voice does not need to be policed. It defends itself.
Key takeaways
A majority in every audience reports no clear, finished rule on permitted AI use — professionals least governed of all, at seven in ten.
Around three-quarters of educators have already changed how they assign or grade writing.
Assessment practice is moving faster than written policy, leaving rules to be set classroom by classroom.
4.2The clarity dividend
The strongest behavioral relationship anywhere in this survey is not about detection. It is about instruction.
Among students who said their instructor or institution had told them clearly what AI use was allowed, around one in seven reported having submitted AI-generated text as their own original work. Among students told only “somewhat,” the figure roughly doubled. Among students who had been told nothing at all, it was more than four in ten — around three times the rate of the clearly-instructed group.
Figure 15. Share of students reporting they have submitted AI-generated text as their own, by the clarity of guidance they received. Values approximate — reconstructed from the proportions reported in the text.
We are careful about what this shows. It is an association within a single survey, not a controlled experiment, and causation could plausibly run in more than one direction — institutions with a strong integrity culture may be both more likely to write a clear policy and more likely to have students who behave accordingly. Students in clearly-governed environments may also be more cautious about admitting misconduct.
Even allowing for all of that, the size of the difference is hard to ignore, and it points somewhere useful. Clear guidance costs an institution almost nothing. It carries no false-positive risk. It does not require procurement, calibration, or an appeals process. Compared with every other intervention discussed in this report, it has by far the best ratio of effect to cost — and it is the one thing a majority of institutions still have not done.
Key takeaways
Students given clear AI guidance were roughly a third as likely to report submitting AI text as their own.
The relationship is associative, not causal — but it is the largest behavioral difference in the survey.
Clear guidance is the cheapest intervention available and carries no false-positive risk.
For institutions
What “clear” appears to mean in practice
The students who reported clear guidance were not describing a values statement. Across the responses, the guidance that registered was specific and situational: which tasks AI may be used for, at which stage of the work, whether disclosure is required, and what the consequence of undisclosed use is. A single line in a syllabus saying AI use “must be appropriate” is not, on this evidence, distinguishable from saying nothing at all.
4.3The detection debate splits along a predictable line
Asked whether institutions should rely more on AI-detection tools, the three audiences answered in almost perfect inverse order to how much they stand to lose from a wrong call.
Figure 16. Support for greater reliance on AI-detection tools, among those expressing a view. The educator question offered only yes or no; students and professionals could also answer “unsure,” which will widen the apparent gap. Values approximate — reconstructed from the proportions reported in the text.
Around three-quarters of educators supported greater use of AI detection tools, compared with about half of students. Professionals were the most skeptical — only about a quarter supported greater reliance, while the largest share said they were unsure.
Even among educators, support was cautious rather than absolute. Most reported only moderate confidence in their ability to identify AI-written work on their own, and many had experienced wrongly accusing a student. This suggests they see AI detection as a helpful aid, not a definitive answer.
The findings point to a clear conclusion: AI detection should be treated as one signal among many. It is most effective when it starts a conversation, not when it is used as final proof.
Key takeaways
Roughly three-quarters of educators support greater reliance on detection; only about a quarter of professionals do.
Only four in ten educators rate their confidence in identifying AI writing unaided in the top two points of the scale.
Support for detection tracks exposure to the problem, while skepticism tracks exposure to the cost of being wrong.
Part V
What This Changes
· conclusion
Conclusion: the question has moved
For the past two years, the conversation around writing integrity has focused on one question: Can we detect AI-generated writing? Our findings suggest the more important question is now How should AI be used?
Our data shows that while AI-assisted writing remains widespread, plagiarism continues to decline. At the same time, educators report seeing more AI use than ever. Together, these findings suggest that AI-assisted writing and plagiarism are no longer the same problem. As AI evolves, detection alone will become an increasingly incomplete way to assess writing integrity.
The survey highlights a different challenge. Most students, educators, and professionals still lack clear guidance on acceptable AI use. Yet the strongest behavioral difference in our data comes from something much simpler than technology: clear policies. Students who understood what was allowed were far less likely to report undisclosed AI use.
The takeaway is clear. The future of writing integrity depends not only on better detection tools, but on better expectations. Organizations that clearly define how AI can and cannot be used are better positioned to build trust, encourage responsible use, and adapt to a world where AI-assisted writing has become a normal part of the writing process.
The bottom line
AI-flagged writing is falling as a measured signal while remaining entirely routine as a lived practice. That divergence is the defining fact of this report, and it will not be resolved by better detection alone. The institutions best positioned for what comes next are the ones that stop treating writing integrity as a detection problem and start treating it as a design problem: clearer rules, better-designed assignments, honest disclosure norms, and detection used as one input among several rather than as a verdict.
Implications
For educators and institutions
Clear AI policies are associated with lower rates of undisclosed AI use.
AI detection should support human review rather than replace it.
Integrity efforts may be most effective when aligned with periods of higher AI and plagiarism activity.
For professional writers, agencies and the clients who hire them
Organizations should define AI expectations before work begins.
AI-assisted writing and textual originality can coexist, so similarity scores alone are no longer enough to evaluate AI use.
Clear client and employer policies can reduce uncertainty around acceptable AI use.
For students
Understanding an institution’s AI policy reduces uncertainty about acceptable use.
Keeping drafts and revision history provides valuable evidence of the writing process.
AI assistance and plagiarism are different issues and should not be treated as the same behavior.
Methodology
Platform data
Platform analysis includes 500K AI checks and 500K plagiarism checks conducted in 2025–2026. AI flag rate represents the percentage of documents with an AI score of 80% or higher, while plagiarism flag rate represents the percentage of documents with a similarity score of 20% or higher. Average similarity is calculated across all plagiarism checks, not just flagged documents. Year-over-year comparisons use matched calendar months. Where results are shown by user segment, they reflect differences between groups within the Quetext user community. Document length analyses are based on a broader set of submissions and should be interpreted independently from the AI and plagiarism datasets.
Survey
The Quetext Writing Integrity Survey gathered responses from three groups within the Quetext user community: students, educators, and professionals who write as part of their work. Respondents self-identified their group, and parallel question sets were used to enable comparisons across audiences. Results are reported as the percentage of respondents to each question.
Most comparisons are made directly. Two questions required special handling due to differences in survey design: one agreement question was normalized to a common scale, and one question about AI detection tools used a different response format for educators. These differences are noted where relevant.
This survey reflects the views of Quetext users and is not intended to represent all students, educators, or professionals. Throughout this report, the strongest insights come from comparing how different audiences respond, rather than treating the percentages as representative of the broader population.
What this data can and cannot show
The data can establish that flag rates moved and by how much. It cannot establish why. It cannot distinguish between writers using less AI assistance, writers editing machine output more heavily before submitting, shifts in who is using Quetext, changes in detection behavior on either side of the exchange, or deliberate flag avoidance. Any causal claim beyond “the flag rate declined” goes beyond what this dataset supports, and we have avoided making one. Similarly, the survey establishes association, not causation: where two survey findings move together — clarity of guidance and reported conduct, for example — we say so and explain the alternative readings rather than asserting a mechanism.
Frequently asked questions
Does a falling AI flag rate mean less AI-assisted writing?
Not necessarily, and our survey suggests probably not. This dataset shows that the flag rate fell; it cannot on its own distinguish between reduced AI use, heavier human editing of machine drafts, a shift in who is using Quetext, changes in the detection model, or writers deliberately working to lower their scores. Self-reported AI use among the writers we surveyed is high and routine, which makes “less AI is being used” the least well-supported of those explanations.
Why is the plagiarism flag rate so much lower than the AI flag rate?
Because the two checks measure different things. Plagiarism similarity looks for matches to text that already exists. AI likelihood estimates how machine-generated a passage appears. A document can score high on one and low on the other, and in 2026 the two moved independently — including peaking in different months.
Is a flagged document proof that someone used AI?
No. A flag is a probability estimate about a piece of text, not a determination about a person. This report describes how that estimate behaved across a year at scale; it makes no claim about any individual document or writer. The survey findings on false accusations are a direct argument for treating flags as one input among several.
How large was the survey, and is it representative?
The survey was fielded to Quetext’s user community across three audiences and is not a nationally representative sample. It is modest in size, which is precisely why it is reported here in proportions and directional comparisons rather than point estimates — those are the findings least sensitive to sample composition. It should be read as a well-sourced snapshot of the people who use writing-integrity tools, not as a population estimate. The educator and professional audiences are smaller than the student audience, and individual figures drawn from them are best cited alongside the direction they point in rather than on their own.
Citation and reuse
This report is free to cite, quote and excerpt with attribution. Charts may be reproduced provided the source line remains legible.
For media inquiries, additional cuts of the data, or high-resolution versions of any chart in this report, contact the Quetext communications team via [email protected].