Technical Limits, Ethics, Adaptations, and Evolution of Invalid AI Detectors

Universities and Educational institutions are setting up a lawsuit against them, and they will surely lose.

AI content detection tools – exemplified by Turnitin’s AI-writing detector – face significant technical limitations. These systems generally work by analyzing text for statistical patterns or anomalies believed to signal AI-generation. For instance, detectors often measure perplexity (how predictable the text is) and burstiness (variations in sentence length). The logic is that AI-generated prose is “extremely consistently average” – it tends to use the most statistically probable words in a uniform style – whereas human writing usually contains more irregularities. Turnitin’s VP of AI has described their model as flagging text that is “too consistently average.” In theory, this helps catch AI-written passages that lack the spontaneous quirks of human expression. However, a fundamental limitation is immediately clear: sometimes human writers are consistent or formulaic, and conversely, advanced AI models can inject controlled randomness. As AI improves at mimicking human style, those telltale statistical fingerprints become far less reliable. In short, no detector can definitively distinguish AI text from human text in all cases – they provide probabilistic guesses based on imperfect heuristics.

The accuracy of these tools is therefore shaky. Vendors often boast high success rates in controlled tests (e.g. claims of 95–99% accuracy are common in marketing), but independent evaluations paint a different picture. Early in 2023, Turnitin asserted its new AI detector was 98% accurate with less than 1% false positives. In practice, that figure did not hold. A false positive, by definition, means the tool incorrectly flags human-written text as AI-generated. Educators and journalists soon found false positives occurring at much higher rates than “<1%.” In one trial, Turnitin’s detector “got over half [of the tested essays] at least partly wrong,” even flagging an 8% portion of a true student essay as AI-written. Anecdotally, some observers have noted that the “state of the art” detectors have false-positive rates “as high as 50%,” essentially no better than a coin flip. Statistically, a 50% false-positive rate implies that half of genuine human submissions could be mislabeled as AI – a shockingly poor specificity. Practically, this means any given accusation from the tool might be as likely wrong as right. An educator using such a tool would risk falsely accusing one in every two innocent students, which is an untenable error rate in high-stakes contexts.

False negatives are also a concern: just as detectors often catch “innocent” writing, they can miss AI-generated text that is cleverly edited or randomized. Simply paraphrasing AI output or using an “AI humanizer” tool can circumvent many detectors. There is an inherent trade-off between catching more cheaters and not punishing the innocent. If a detector is tuned to be very strict (minimizing false negatives), it tends to generate more false positives; if tuned to avoid false positives, it will let most AI usage slip through. One online discussion vividly illustrated this trade-off: you could set an AI detector’s threshold so high that it only flags 1% of human-written papers (a 1% false positive rate), “as long as you don’t mind the 99% false negative rate” that results. In other words, one can almost eliminate false accusations by making the tool extremely conservative – but then it fails to catch actual AI-written text nearly all the time. This balance is very hard to get right in practice. Because human and AI writing overlap substantially on many dimensions, any threshold will produce some errors. And with current technology, those errors are not trivial in frequency. Even Turnitin later acknowledged their real-world false positive rate was higher than initially promised, although they didn’t disclose an exact figure. They noted particular difficulty with “mixed” writing – when parts of an assignment are AI-written and parts human. In such cases, the detector often falsely flags surrounding human sentences as AI simply because they sit near AI-generated sentences. This highlights how context and overlap confound the algorithm.

Another technical limitation is the lack of transparency and feedback. These tools typically operate as black boxes: they assign a score or label but provide limited explanation for why text was flagged. For example, Turnitin’s AI report gives a percentage of the document suspected to be AI-written, yet instructors see only colored highlights without a clear rationale. There is no straightforward way to verify the result or pinpoint which linguistic features triggered it. This opacity makes it hard for both students and teachers to trust or evaluate the output. Even the companies caution that the outputs shouldn’t be taken as ironclad proof. Turnitin, for instance, added a warning on its results stating, “Percentage may not indicate cheating. Review required.” In interviews, Turnitin’s own executives stressed that the AI detection score is meant to be an “indication, not an accusation” – essentially a tip for educators to investigate further, not a final verdict. This is an important admission: it underscores that these tools cannot reach 100% certainty. They provide probabilistic data that must be interpreted with human judgment. Unfortunately, in practice some institutions have been treating detector outputs as if they were conclusive evidence, which raises serious ethical issues.

In summary, AI text detectors suffer from high error rates (especially false positives), an unavoidable trade-off between sensitivity and precision, and technical opacity. A purported false-positive rate of 50% is tantamount to random guessing, illustrating just how unreliable these tools can be when pushed to their limits. The technology simply cannot definitively tell human and AI writing apart in all cases, because the differences are often subtle and shrinking. This technical reality sets the stage for the legal and ethical dilemmas that follow when such fallible tools are used in educational settings.

Legal and Ethical Concerns of AI Detection in Education

Relying on AI detection tools in schools and universities poses a minefield of legal and ethical issues. At the heart of the matter is trust and fairness: a student can be falsely branded dishonest based on an algorithm’s mistake, which is a profoundly unjust outcome. False accusations carry severe consequences for students’ academic and personal lives – and these tools do falsely accuse students, as we established. Even a small false-positive rate can translate into large numbers of students wrongly flagged when applied across many assignments. For example, one analysis noted that if 22 million college essays were submitted by U.S. freshmen in a year (a realistic scale), a 1% false-positive rate would mean about 223,500 entirely human-written essays could be mislabeled as AI-generated. Each of those flags represents a student potentially – and erroneously – under suspicion of cheating. Now consider that some evidence suggests the real false-positive rates might be an order of magnitude higher. The ethical stakes are enormous.

From a due process standpoint, it is deeply problematic to treat an AI detector’s output as sole evidence of academic misconduct. Unlike plagiarism, where a teacher or honor board can see the copied source, AI-generated content leaves no tangible proof – the detector’s judgment is essentially the only “witness.” This flips the burden of proof onto the student to somehow prove a negative (i.e. prove they didn’t use AI), which is often impossible. Students have reported being put through academic integrity investigations “based solely” on a Turnitin AI flag, with “no evidence required beyond the score.” In such cases, the student is presumed guilty by algorithm. This violates principles of fairness and open justice, and it likely contravenes school policies or legal standards that require evidence for punitive action. The lack of transparency compounds the issue – because the detection algorithms are proprietary black boxes, neither students nor educators can scrutinize how the AI arrived at its conclusion. Was the student flagged because of a certain phrasing, an unusual consistency in tone, or something else entirely? No one can say, not even Turnitin’s support reps. This opacity undermines the student’s ability to contest the accusation. In an ethical academic process, an accused person should be able to challenge the evidence against them. Here, the “evidence” is essentially inscrutable. Educators themselves have no clear standard to judge if the AI’s flag is credible or a fluke.

These concerns are not just theoretical. They manifest in real student experiences. Cases have emerged of diligent students – including ESL (English as a Second Language) writers and neurodivergent students – being falsely flagged and then struggling to defend themselves against the machine’s verdict. The emotional and educational toll can be severe: “To be falsely accused felt devastating,” said one student who was wrongly flagged by an AI detector. Students have described anxiety and distrust after such incidents, with some going so far as to document every step of their writing process (e.g. taking photos of draft progress) just in case they need to prove their work is original. This is a toxic learning environment – one where students feel they are under constant suspicion and must compile a defence dossier for their own writing. It runs counter to the supportive, honest atmosphere that educators aim to foster.

Furthermore, the errors made by AI detectors are not evenly distributed; they reflect and potentially amplify biases. Research indicates that AI writing detectors disproportionately flag certain kinds of writing and writers. Notably, texts by non-native English speakers are more likely to be misclassified as AI-generated. A Stanford study found that these tools can penalise non-standard syntax or less common vocabulary, which often correlates with ESL writing styles. In fact, controlled experiments have shown non-native English essays being falsely flagged at double the rate of native English essays, even though all were human-written. This means international students and others writing in a second language face a higher risk of unwarranted suspicion – essentially, they are punished for their writing style. There are similar concerns for other marginalised groups. Reports have warned that Black students may be more likely to be accused of “AI plagiarism” by teachers, likely due to biases in how their natural voice is perceived or how detection tools were trained. Likewise, neurodiverse students, who might write in atypical styles, have been flagged by detectors at higher rates. These disparities raise serious equity and discrimination issues. Wrongly accusing any student is bad enough, but if false positives skew toward already disadvantaged groups, the tool could exacerbate existing inequalities in education. In the U.S., such patterns could even provoke legal scrutiny under anti-discrimination laws. The ethical imperative for educators is clear: one cannot justify using a “cheating detector” that unfairly targets certain demographics. Doing so erodes trust and could violate students’ rights to equal treatment.

Privacy and intellectual property concerns provide another layer of legal/ethical trouble. Most AI detection services (Turnitin included) require uploading student work to an external server, where it is analysed and often retained. This raises questions under privacy laws like FERPA (which protects education records in the U.S.) about sending student essays to third-party companies. Students and parents might rightly worry: Who owns the data once a paper is submitted to an AI detector? Will it be stored indefinitely, and for what purpose? In Turnitin’s case, the company explicitly demands broad rights over submitted student papers – “perpetual, irrevocable, non-exclusive, royalty-free, transferable and sublicensable” rights, as one investigative report uncovered. Turnitin has amassed a giant database of student writing over decades and uses it not only to check plagiarism, but now to train and market its AI detection algorithms. Ethically, students are often not fully aware that by submitting an assignment to Turnitin, they are potentially contributing their original writing to a for-profit company’s intellectual property. This has led to long-standing debates about student copyright: in some countries and jurisdictions, there have been challenges arguing that such practices infringe on students’ ownership of their work or their consent. Legally, universities have had to examine whether using these tools might conflict with policies that promise students control over their academic work. At minimum, transparency is needed – students should know if their work will be stored or repurposed beyond just checking their own paper.

Finally, consider the broader impact on the academic ethos. When teachers lean on AI detection as a crutch, it may undermine the pedagogical relationship. There’s an ethical slippery slope in outsourcing one’s judgment to an algorithm. If instructors start to treat the detector as the ultimate arbiter of integrity, it can erode the trust between student and teacher. Educators have voiced concern that an over-reliance on AI detection cultivates a “gotcha” mentality rather than a culture of learning and honesty. It shifts focus from teaching good writing and critical thinking to policing and surveillance. A recent ethics review argued that excessive AI-policing “fosters a culture of suspicion rather than support” in the classroom. Students may feel less inclined to take intellectual risks or develop their own voice, fearing that anything slightly different could trigger the AI alarm. This runs counter to the educational mission of encouraging originality and growth. Thus, from an ethical standpoint, educators must ask: Are we protecting academic integrity, or are we harming it by using tools that can mislabel creativity as cheating? The answer isn’t straightforward, but the potential for chilling effects on student writing is real.

In summary, the use of current AI content detectors in education raises red flags on multiple fronts. Legally, institutions could face challenges if students are disciplined without solid evidence or if data privacy is breached. Ethically, false positives and biased outcomes risk damaging student welfare, amplifying inequality, and corroding the trust that is fundamental to the educational process. As one expert bluntly put it, investing in these flawed detectors might be better replaced by investing in supporting teachers and students directly, given the limited value of a tool that can so easily be wrong or gamed. The next section looks at how educational institutions are responding to these concerns, as many are beginning to recognise that the promise of easy AI detection has not materialised in a fair or reliable way.

How Educational Institutions Are Adapting to Unreliable AI Detection

Faced with the inaccuracies and controversies of AI detectors, many educational institutions are rethinking their approach to AI in academic work. The initial wave of enthusiasm for “catching cheaters with AI” is giving way to a more cautious, nuanced strategy. One clear trend is that some schools have pulled back from using AI detection tools altogether. In August 2023, for example, Vanderbilt University made headlines by disabling Turnitin’s AI-writing detector campus-wide. This decision came after several months of trial and considerable concern among faculty. Vanderbilt’s leadership cited the tool’s unreliability, lack of transparency about its inner workings, and the risk of false accusations as key reasons for the ban. Importantly, this university framed the move as protecting students’ interests: they determined that until detection technology significantly improves (if it ever does), using it could do more harm than good to the academic integrity process. Vanderbilt was not alone. A “significant majority” of universities in the UK also initially opted not to enable Turnitin’s AI detection when it was first rolled out, with many UK educators explicitly asking Turnitin to hide or disable the AI-generated score on submissions. These institutions essentially decided that they couldn’t trust the detector enough to act on its results, so it was safer to avoid exposing faculty and students to a potentially misleading metric.

Even at places where AI detectors are available, academic integrity offices and teaching centers are issuing strong guidelines on limited use. The consistent advice is: do not treat an AI detection score as conclusive evidence. For instance, the University of Pittsburgh’s teaching center explicitly recommends against using AI detection tools as proof of misconduct, noting they are not accurate enough to prove a violation. Instead, instructors are urged to rely on traditional methods of evaluation and, if needed, to gather more direct evidence (such as changes in a student’s writing level or confirmation from the student). More generally, experts insist that any AI flag should be just the start of a human-led inquiry, not a verdict. A recent policy brief advised institutions to never use AI detector results as the sole piece of evidence in an academic misconduct case. If a paper is flagged, the recommendation is to perform a manual review: compare the suspicious paper to the student’s past writing, discuss the assignment with the student, and look for other signs (like factual errors or fabricated sources that AI might produce). In effect, schools are adapting by re-embedding human judgment into the loop, rather than letting the algorithm be judge, jury, and executioner. Some have even formalized this: they require a committee or instructor to corroborate AI allegations with a viva voce (oral exam) or supplemental tests of the student’s knowledge before any academic penalty is imposed.

Another adaptive strategy is greater transparency and communication about AI. Rather than covertly trying to catch students using ChatGPT, many educators now openly talk with students about generative AI tools. Syllabi and honor codes are being updated to clarify what is allowed and what isn’t. For example, a professor might state that using AI to brainstorm ideas or correct grammar is acceptable, but using it to generate full answers is not – and students must disclose any AI assistance they did use. By setting these expectations up front, the hope is to mitigate misuse without needing a detector at all. Some institutions encourage faculty to have students sign an “AI usage statement” for assignments, or to include reflective questions where students document their writing process and any tools used. This shifts the focus from a “gotcha” approach to a more integrative approach: it treats AI similar to other resources (like the internet or calculators) that need guidelines and ethical norms. In line with this, educational institutions are developing AI literacy programs for both students and staff. The idea is to teach why certain AI use might be inappropriate (plagiarism, lack of learning, etc.) and also how AI can be used responsibly to enhance learning when permitted. By improving understanding, schools aim to reduce the adversarial cat-and-mouse dynamic that drove the initial demand for detectors.

Importantly, some colleges are exploring alternative assessment designs to reduce opportunities for undetectable AI cheating. This is a proactive pedagogical adaptation. For example, instructors might increase the weight of in-class writing, oral presentations, or one-on-one discussions of work, so that a student’s knowledge and voice must be demonstrated in person (where using ChatGPT live is impractical). Others are assigning more personalized or current-event topics for essays, under the reasoning that AI trained on past data will struggle more with highly specific, up-to-the-minute prompts. Another approach is scaffolding assignments – breaking a research paper into smaller components (proposal, outline, draft, etc.) – which makes it easier to spot inconsistent jumps in ability and also engages students in the process so deeply that reliance on AI is less tempting. Some institutions are even trialing supervised writing exams or using tools that lock down computers to prevent opening AI tools during tests. While such measures hark back to more controlled assessment environments, they reflect a recognition that if AI text generation is here to stay, the way we evaluate learning may need to evolve in response, rather than hoping a perfect detection algorithm will save the day.

In terms of policy, universities are starting to codify these adaptations. We see emerging policies that explicitly warn faculty not to punish students based on detector results alone, and that outline steps for verification. Some academic integrity hearings now treat AI similar to plagiarism in that the intent and context are considered – for instance, was the student properly taught about AI tool use? Did they just use it for editing or for the entire assignment? The outcome might differ in those cases. Significantly, institutions are also considering the liability aspect: if a student is falsely accused and later exonerated, that could expose the university to legal challenges. So from a risk management perspective, many are erring on the side of caution by treating detector outputs as at most a “lead.” In fact, one could argue institutions are hedging: they pay for these AI detection tools (often bundling them with plagiarism services), but quietly advise instructors to be very wary of acting on them. As one technology commentator wryly noted, colleges are spending millions on detectors that have such limited value and are easily undermined, so resources might be better spent elsewhere. This sentiment is pushing schools to think more broadly about academic integrity in the AI age.

Finally, an adaptation worth noting is the stance of some public school districts and academic bodies that have banned or discouraged not just detectors, but even AI-generated work outright until policies catch up. For instance, some K-12 school districts initially blocked access to ChatGPT entirely. While that’s a different approach (aimed at prevention rather than detection), it shows how institutions are searching for workable solutions to uphold integrity. Over time, the pendulum seems to be swinging from extreme measures (blanket bans and heavy surveillance) toward a middle ground that emphasizes education over enforcement. The consensus among many educators and scholars is that we won’t “tech” our way out of this problem with detectors alone; instead, we must adapt teaching practices and policies to coexist with AI. This means cultivating academic cultures where integrity is emphasized and where the use of AI, if allowed, is transparent and guided. It’s a shift from trying to catch students to trying to teach students about the responsible use of technology – a positive development born directly out of the recognition that our current detection tools are too flawed to carry the load.

A Brief History of AI Content Detection and How Current Tools Work

The quest to detect AI-generated writing is almost as old as AI text generators themselves, though it truly leapt to prominence with the advent of powerful models like GPT-3 and ChatGPT (2020–2022). Early language models a decade ago produced text that was usually easy to spot – often incoherent or robotic-sounding – so there wasn’t much need for special detection. However, as AI writing quality improved, researchers began pondering detection methods. One early idea was to use AI against itself: in 2019, OpenAI released a tool to detect outputs from its GPT-2 model by training a classifier on known AI vs. human text examples. This showed some success on short texts but was far from foolproof. Another approach that emerged (and remains widely used) is based on analyzing the statistical signature of AI text. In mid-2020s research, scholars introduced metrics like perplexity to quantify how “predictable” a piece of text is. The intuition is that AI-generated text tends to have lower perplexity (the next word is too neatly probable) compared to the sometimes quirky or less optimal word choices of a human writer. Tools like GLTR (Giant Language Model Test Room) developed by MIT-IBM and Harvard in 2019 visualized this concept, allowing users to see how likely each word in a text would be predicted by a model – lots of high-probability words could imply machine writing. GLTR was more of a demonstration than a consumer tool, but it set the stage for later detectors.

The real boom in AI content detectors came after OpenAI launched ChatGPT in late 2022, which triggered both excitement and panic in education. By January 2023, GPTZero debuted, a tool created by a Princeton student that garnered widespread media attention. GPTZero specifically popularized the dual metrics of perplexity and burstiness to flag AI writing. Its premise was that human writing usually has a mix of short and long sentences and some unusual phrasings, whereas AI writing can be more uniform and “flat.” GPTZero’s public release went viral, reflecting the strong demand among teachers for an AI essay detector. Soon, a proliferation of similar services appeared – each claiming to use advanced algorithms to separate human from machine text. Many of these services touted extremely high accuracy on their websites, but these figures were typically not peer-reviewed and sometimes reflected ideal conditions. In reality, when researchers benchmarked multiple detectors on varied samples, results were far less rosy. One study in 2023 tested a suite of detectors on human-written essays and found false-positive rates ranging roughly 10% to 30% for different tools. In other words, many detectors mislabeled up to a third of genuine texts as AI – a far cry from the near-perfect accuracy some had advertised.

Turnitin’s AI Writing Detection deserves mention in this historical arc. Turnitin had long been the dominant plagiarism detection platform, founded in the late 1990s and widely adopted by the 2000s for catching copy-paste plagiarism. Its methods for plagiarism are straightforward – matching text against a massive database of sources and past submissions – but detecting AI required a new approach. Turnitin developers worked largely in secret on an AI detection model and rushed it to market by April 2023, likely to capitalize on the academic anxiety around ChatGPT. They integrated it into the existing Turnitin interface, meaning millions of assignments got an “AI score” overnight. This was historically significant: it instantly normalized AI-checking in the same workflow as plagiarism-checking. However, the launch was fraught with issues. Teachers quickly encountered cases of false flags, and Turnitin provided little transparency about how the tool worked, citing only looking for “patterns common in AI writing.” Technically, it’s believed Turnitin’s model is a trained classifier that produces a sentence-level and document-level prediction. Turnitin initially claimed extremely high confidence in its tool, but within months had to revise and add caveats as real-world feedback came in. Historically, it’s interesting that Turnitin’s credibility from plagiarism detection perhaps led some educators to over-trust its AI detection – an assumption that proved problematic once discrepancies came to light.

Another development in AI detection has been the idea of watermarking AI-generated content. Researchers, including a team at OpenAI in early 2023, proposed embedding a hidden signal in AI text – for example, having the AI subtly prefer certain uncommon words or punctuation in a detectable pattern. If universally adopted by AI generators, this could allow a reliable “watermark detector” to later identify machine-written text with high certainty. However, as of 2025, watermarking has not become standard. OpenAI did not deploy a visible watermark in ChatGPT’s outputs. Thus, external detection still relies on analyzing the text after the fact, with all the uncertainties that entails. Users found that even simple tricks can defeat statistical detectors: adding a few random uncommon words, intentionally varying phrasing, or using one AI to paraphrase another’s text can lower an AI-detection score dramatically. Detection methods have known blind spots that savvy users can exploit, creating an “arms race” between AI generation and detection.

A major inflection point in the history of AI detectors was the stance of AI developers themselves. OpenAI discontinued its own AI text classifier in July 2023 due to a “low rate of accuracy.” The tool had only been launched in January 2023, but by summer OpenAI admitted it was not reliable enough and pulled it. OpenAI’s classifier in tests could only correctly identify a small portion of AI-written text while mislabeling human text too often. The fact that the creator of ChatGPT publicly walked away from AI detection sent a strong signal to educators: if even OpenAI doesn’t have confidence in detection technology yet, educators should be very careful about depending on third-party detectors for serious decisions. It was a reality check on the limits of current algorithms.

In summary, the evolution of AI content detection has been swift but rocky. We went from rudimentary classifiers and concept demos in 2019, to an explosion of commercial detectors in 2023 trying to meet an urgent need, to a 2024–2025 landscape where many of those tools are either struggling, pivoting, or being used with caution. Technologically, current detectors function mostly by statistical pattern recognition (perplexity, burstiness, and AI likelihood scores) rather than any deep “understanding” of content. They are inherently fallible because human and AI writing are not binary categories – they overlap, and the boundary grows blurrier as AI models improve. Each new iteration of GPT or other large language models forces detectors to catch up, often lagging behind. Many experts believe truly reliable detection may never be fully achievable if AI-generated text becomes indistinguishable from human text in style and quality. The history so far teaches humility: initial confidence in these tools has given way to recognition of their limitations. Early attempts made the gray area of AI-assisted writing seem “black-and-white,” but honest students have sometimes been caught in the crossfire of this arms race.

Reflective Conclusion: AI content detectors, as they stand today, are blunt instruments ill-suited for the delicate task of evaluating academic work. Their technical shortcomings produce too many false alarms, raising legal risks and moral questions when used to police students. The rapid adoption and subsequent pullbacks by institutions illustrate a collective learning process: promoting academic integrity in the age of AI will require more than a quick tech fix. It calls for a thoughtful blend of technology use, policy development, and pedagogical change. Until AI detectors become far more accurate, educators are wise to treat them with skepticism – or avoid them entirely – and instead focus on constructive strategies like student engagement, honor codes, and assessment design that uphold integrity without sacrificing trust. Maintaining academic integrity has always been about a partnership between teachers and students grounded in honesty and understanding. No AI, either as friend or foe, should be allowed to erode that partnership. The ongoing challenge is to integrate new tools like AI in a way that enhances learning and integrity rather than undermining them. It’s an evolving story, reminding us that technology solutions must be evaluated critically, especially when core values and rights are at stake.

References (APA 7th Edition)

Chechitelli, A. (2023, March 16). Understanding false positives within our AI writing detection capabilities. Turnitin Blog. turnitin.comturnitin.comColey, M. (2023, August 16). Guidance on AI detection and why we’re disabling Turnitin’s AI detector. Vanderbilt University, Brightspace Support. vanderbilt.eduvanderbilt.eduD’Agostino, S. (2023, June 1). Turnitin’s AI detector: Higher-than-expected false positives. Inside Higher Ed. insidehighered.cominsidehighered.comFowler, G. A. (2023, April 3). We tested a new ChatGPT-detector for teachers. It flagged an innocent student. The Washington Post. washingtonpost.comwashingtonpost.comHirsch, A. (2024, December 12). AI detectors: An ethical minefield. Center for Innovative Teaching and Learning, Northern Illinois University. citl.news.niu.educitl.news.niu.eduMathewson, T. G. (2025, June 26). Costly and unreliable: AI and plagiarism detectors wreak havoc in higher ed. CalMatters. calmatters.orgcalmatters.orgSouthern, M. G. (2023, July 25). OpenAI shuts down flawed AI detector. Search Engine Journal. searchenginejournal.comsearchenginejournal.com