Authored By: Thato Seoka
University of South Africa
Introduction
A ruling was made in a New York court on February 2026, whereby a university was discovered to have acted without any reasonable justification, in judgement that a first-year student’s history assignment was entirely created by an AI.[1] The student had been busy with the assignment, for twenty hours and had assistance from tutors, and had run it on two different detectors of which identified as “ human-written”.[2] This case reveals a more comprehensive legal issue: AI detection technology is deemed reliable evidence of dishonesty even though it was never invented for such standards.
Ensuing the introduction of big language models into the open domain in late 2022, institutions of higher learning, businesses, publishers, and clients have commenced in the usage of AI-monitoring tools, such as Turnitin, GPTZero, for the verification of integrity of written material. Another tool called OpenAI’s own classifier, was established from its program’s output, demonstrated itself capable only of identifying 26 percent of texts written by AI, while misidentifying 9 percent of writings created by humans as AI-generated, and removed off the market within six months due to its “low precision.”[3] Test results from self-reliant peer-appraised studies have not been any more successful.[4]
Based on the article, it is procedurally unsound, but also legally perilous for anyone to rely on AI detectors for the dismissal of someone’s original work. Such conducts make academia’s more vulnerable to the possibility of defamation, illegal administrative actions, and regulatory penalties for institutions and employers who will rely on the AI detector for their findings. Moreover, the individual accused of using AI would not be able to protect himself or herself, unless before the wrongful action is taken against the accused individual, substantiating evidence is available
In Section II, the legislative structure around the scope of defamation, administrative fairness, and consumer protection as pertaining to allegations against AI-based detection systems is unveiled. Section III probes the respective case-law in the area. Section IV assesses the issues quality-wise and Section V breakdowns the same developments internationally.
The Legal Framework
The Technology and Its Limitation
General AI text detecting program’s function using “perplexity,” a parameter because machine-generated text must be smoother than human-written text. This method is not accurate, as simple texts can be tagged as “AI” even if they were written by a human. Based on a vast research studying the effectiveness of fourteen AI detectors, none have shown a good result while resolving the authorship of texts, like simple paraphrasing in the original text, permitted the authors to fool most of the detection systems.[5] In another research studying seven most popular detection programs, it was discovered that more than half of papers written by non-native English speakers are misclassified as AI-generated texts, while in case of papers written by native speakers this error is practically absent.[6] Such cases define ordinary procedures of the technology used.[7]
Defamation
In an institution or in situation whereby an employer makes a claim of dishonesty due to AI to some third parties, such as a colleague, disciplinary board or future boss, all action could amount to defamation. In the South African context, defamation entails that someone published a defamatory statement about the plaintiff[8] and such conduct is wrongful and intentional. After the publication of a defamatory statement was confirmed, it is certainly assumed that it was wrongfully done, and intention, whereas the defendant must provide some defence, such as the statement being true for the public interest or some other defence.[9] As the claim of fraudulent authorship affects the academic and professional integrity of an individual, it can damage the reputation of a person among any reasonable audience, thus defamatory in nature.[10] Hence, if results of the detector are incorrect, it will be battle to defend against the claim of truth.
Administrative and Procedural Fairness
In circumstances where a governmental entity or state institution files a charge, any finding of academic dishonesty forms administrative action, which must comply with the principles of legality, rationality, and the procedural fairness.[11] This decision can be challenged, inter alia, but on grounds of lack of connection between the finding and evidence presented to the decision-maker.[12] As findings made on the basis of the proprietary detector score, with no evidence that connects the finding to already publicly available information or an actual chance of the accused party’s defence, will fail, regardless of the jurisdiction from which review is being sought.[13]
Consumer and Regulatory Protection
The laws governing consumer rights and competition deliver yet another, poorly examined means of shielding oneself from unscrupulous detector suppliers. In 2025, the FTC (Federal Trade Commission) secured a judgement against a firm called Workado, LLC for affirming that its AI Content Detector obtained 98% accuracy in spite of tests showing that their actual accuracy was only 53%.[14] The judgement prohibits the firm from making unsupported claims and requires it to provide evidence supporting similar claims in the future.[15] Although, it fails to compensate the people who were wrongly accused or harmed in the process, it does establish that claims regarding vendor accuracy should not be blindly trusted, as verification will be done by an independent regulatory body.
Case Law Analysis
Matter of Newby v Adelphi University
An autistic freshman, Orion Newby, had submitted a historically researched paper that was assessed by a tutor, at a university’s disability-support programme[16] The paper was tagged 100% AI-written by Turnitin,[17] yet two independent detectors discovered it to be human-made. The university’s integrity officer rejected the human writing claims despite this second evidential finding, and the same integrity officer was responsible for the appeal ruling.[18] The New York Supreme Court reversed the university’s ruling on two [19]grounds, of which is unverified proof for the judgement and a violation of Newby’s right to a real appeal process as stated in the university regulations.[20] This case personifies both grounds for challenge, for a decision lacking sound evidential basis, and a procedure that denied the student genuine due process.
In re Workado, LLC
Based on the tale of Workado, the infamy of the precise product prolongs into the sphere of overall reliability of detectors. The FTC ingrained that the detector used by Workado has been trained using only literature and was sold as universally applicable.[21] As the claimed accuracy of 98% dropped to only half of the time, which is comparable to the random prediction.[22] The final order obliges Workado to provide proof for future accuracy claims.
Critical Evaluation
The issue is a misplaced preponderance of evidence. The output from a detector is nothing more than a statistical estimation which is based on a model formulated by a private entity, of which the details have not been made known to anybody, and was tested under conditions that hardly copy the conditions widespread in real life.[23] Regrettably, a massive amount of institutions and employers like to regard this estimation as irrefutable, which turns the onus of proof upside down. Hence, instead of having to prove misconduct using real evidence, the accused is expected to dismiss some unseen percentage, which cannot be explained in any way by anyone.[24] The heftiest burden is thrown on writers who adopt a style that is different from the narrow training basis for the detector, such as: non-native English speakers, neurodivergent writers, and writers who write plainly and in a formulaic fashion, as they stand no chance of rebutting the accusation based on the information provided by the detector.[25]
This discussion the fact that AI technology does play an important role. The Vanderbilt University reduced the ability of Turnitin to identify AI-written papers completely in August 2023. It turned out that the university calculated that only 1 percent of papers referring to its 75,000- papers that were sent earlier that year might result in 750 innocent accusations every year.[26] Reactions of such is based on the fact that this software can be a good source of information for reaching a logical result but cannot provide the final argument. Should it happen that the accusation was disclosed to a third party and it is not based on enough evidence and was made by someone who can also evaluate the case, then the possible risks explored above in part II of this article are not theoretical, as the Newby case has already shown how it works in practice.
Comparative Perspectives
The Newby case in the United States appears to be the one of many cases of this kind that have recently emerged against universities,[27] proposing that courts start to formulate a doctrine regarding the types of finding made by detection devices. The Federal Trade Commission’s move against Workado stands for a second case, showing that legal action is now being taken by not just against universities but also against suppliers of this technology.
Though South Africa does not have an existing ruling regarding AI detection, the country can employ its current delict and administrative law structure instead of creating new legislation. A ruling made by the Constitutional Court in the case of Khumalo v Holomisa institutes that if a defamatory statement is made, it shall be considered wrongful and intentional. In this case, it will be the onus of the party alleging dishonesty to provide evidence supporting their claim and not for the accused to disprove it.[28] The legislation in this case is the Promotion of Administrative Justice Act 3 of 2000, which proposes that the disciplinary action taken by a public university can be appealed on the same grounds of rationality and fairness, of which were mentioned in the case of Newby.[29] In conclusion, regardless of legislation and jurisdiction, legal systems based on rational decision-making and protection of one’s reputation are able to reach comparable judgement, when faced with a case based on a faulty algorithm.
Conclusion
This article asserts that viewing results from AI-detection software as evidence of dishonesty is empirically illegitimate and legally uncertain. As software measures patterns and not the authorship, and self-reliant studies find that AI detectors’ accuracy is deficient for serious claims and normally such unduly affects creators who are in a vulnerable situation.[30] The Matter of Newby v Adelphi University shows that rulings made based on such evidence is not legally defensible unless it has been verified and properly explored.
This is not a suggestion to ban AI-detection software but instead, it should be limited to what the evidence shows a catalyst for more investigation, but not a decision. Institutions and employers should require confirmation, such as drafts, revision history, or an interview with the author, before reaching a negative judgement, make the report available to the defendant, and ensuring that the appeal takes place before someone different from the original decision-maker. It is imperative for regulators to continue reviewing the accuracy of vendors in similar with independent evidence, which is what the Federal Trade Commission is already doing. Hence, until the technology advances closer to its advertised reliability, it is imperative to treat its findings as something to be verified first and not simply relied on.
References and Bibliography
Table of Cases
Bato Star Fishing (Pty) Ltd v Minister of Environmental Affairs 2004 (4) SA 490 (CC)
Khumalo v Holomisa [2002] ZACC 12, 2002 (5) SA 401 (CC)
Matter of Newby v Adelphi University, 2026 NY Slip Op 26021 (Sup Ct, Nassau County, 29 January 2026)
Pharmaceutical Manufacturers Association of SA: In re Ex Parte President of the Republic of South Africa 2000 (2) SA 674 (CC)
Table of Legislation
Promotion of Administrative Justice Act 3 of 2000 (South Africa)
Bibliography
Federal Trade Commission, ‘FTC Approves Final Order against Workado, LLC, Which Misrepresented the Accuracy of its Artificial Intelligence Content Detection Product’ (FTC, 28 August 2025)
Liang W, Yuksekgonul M, Mao Y, Wu E and Zou J, ‘GPT Detectors Are Biased against Non-Native English Writers’ (2023) 4(7) Patterns 100779
OpenAI, ‘New AI Classifier for Indicating AI-Written Text’ (OpenAI, 31 January 2023)
Vanderbilt University, ‘Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector’ (Vanderbilt Brightspace, 16 August 2023)
Weber-Wulff D, Anohina-Naumeca A, Bjelobaba S, Foltýnek T, Guerrero-Dib J, Popoola O, Šigut P and Waddington L, ‘Testing of Detection Tools for AI-Generated Text’ (2023) 19 International Journal for Educational Integrity 26
[1] Matter of Newby v Adelphi University, 2026 NY Slip Op 26021 (Sup Ct, Nassau County, 29 January 2026).
[2] Ibid.
[3] OpenAI, ‘New AI Classifier for Indicating AI-Written Text’ (OpenAI, 31 January 2023
[4] Debora Weber-Wulff and others, ‘Testing of Detection Tools for AI-Generated Text’ (2023) 19 International Journal for Educational Integrity 26.
[5] Weber-Wulff (n 4).
[6] Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, ‘GPT Detectors Are Biased against Non-Native English Writers’ (2023) 4(7) Patterns 100779.
[7] Weber-Wulff (n 4); Liang (n 6).
[8] Khumalo v Holomisa [2002] ZACC 12, 2002 (5) SA 401 (CC) [18].
[9] Ibid.
[10] Ibid.
[11] Promotion of Administrative Justice Act 3 of 2000, s 6(1); see also Pharmaceutical Manufacturers Association of SA: In re Ex Parte President of the Republic of South Africa 2000 (2) SA 674 (CC) [21]-[23].
[12] Promotion of Administrative Justice Act 3 of 2000, s 6(2)(f)-(h).
[13] Bato Star Fishing (Pty) Ltd v Minister of Environmental Affairs 2004 (4) SA 490 (CC) [44]-[45].
[14] Federal Trade Commission, ‘FTC Approves Final Order against Workado, LLC, Which Misrepresented the Accuracy of its Artificial Intelligence Content Detection Product’ (FTC, 28 August 2025) <
[15] Ibid.
[16] Matter of Newby v Adelphi University (n 1).
[17] Ibid.
[18]Ibid.
[19] Ibid.
[20] Ibid.
[21] Federal Trade Commission (n 14).
[22] Federal Trade Commission (n 14).
[23] Weber-Wulff (n 4).
[24] Liang (n 6).
[25] Matter of Newby v Adelphi University (n 1).
[26] Vanderbilt University, ‘Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector’ (Vanderbilt Brightspace, 16 August 2023)
[27] Mark Lesko, attorney for Orion Newby, stated that he had “heard from other parents regarding similar cases” (CBS News New York, 10 February 2026).
[28] Khumalo v Holomisa (n 8).
[29] Promotion of Administrative Justice Act 3 of 2000 (n 12).
[30] Liang (n 6).
