Yes, AI detection tools can give false positives on human writing, but the likelihood is not identical across products or writing situations. A person can write a passage independently and still receive an AI-likely result because detectors estimate authorship from patterns in text rather than checking a document’s actual history. That distinction matters: a score is an inference, not proof.
The question has become more consequential as schools, universities, publishers and businesses use automated classifiers to review written work. A false positive can be more than an irritating technical error. It can lead to a student being questioned, a writer losing credibility, an employee having to defend routine work, or an editor spending time investigating a submission that was never generated by a language model.
The evidence is also more nuanced than either extreme suggests. Some recent evaluations report very low false-positive rates for selected tools on controlled human-written samples. Other research finds substantial inconsistency across detectors, genres and mixed-authorship conditions. Turnitin itself says its model may misidentify human-written text and should not be used as the sole basis for adverse action. GPTZero similarly advises educators not to punish students on the basis of a classifier result.
This guide separates what a detector score can tell you from what it cannot establish. It examines why false positives happen, what recent studies actually measured, how to interpret products such as Turnitin, GPTZero and Originality.ai, and what a fair review process looks like. The central finding is straightforward: detection can be a prompt for further inquiry, but authorship should be assessed through evidence of the writing process, not a percentage alone.
What a False Positive Actually Means
In classification, a false positive occurs when a system labels a genuinely human-written sample as AI-generated. It is different from a false negative, where AI-generated text is labelled human. Both errors matter, but they create different risks. A false positive can unfairly cast suspicion on an authentic writer; a false negative can allow AI-generated material to pass without being identified.
The denominator is important. A false-positive rate is normally the number of human samples incorrectly flagged divided by all human samples tested. It is not the same as the percentage shown beside an individual document. If a report says “80% AI”, that does not automatically mean 80% of the words were written by AI. For example, GPTZero’s current guidance explains that its percentage expresses confidence in a document-level classification, rather than the share of sentences or words produced by AI. Other products may define their displayed scores differently.
A detector can therefore make two distinct mistakes in the way people interpret it: the classifier may misclassify the text, and a reader may misunderstand what the score represents. Even when a product is correctly reporting its own model output, that output is not a forensic chain-of-custody record. It cannot, by itself, reveal whether the author typed every word, used a grammar checker, translated a paragraph, received editorial assistance or pasted text from a generative model.
It is useful to distinguish three questions. First, did the software identify patterns that its model associates with generated text? Second, how reliably does that association work on this type of writing and this particular sample? Third, does the result establish a breach of a rule? The first is a software question, the second is an evaluation question, and the third is a policy and evidence question. Treating all three as one question is the root of many unfair conclusions.
Why Human Writing Can Look Machine-Generated
AI detectors do not watch a writer compose a document. They analyse the submitted text and look for statistical or learned features associated with different classes of writing. Depending on the product, the model may use patterns involving word choice, sentence structure, predictability, repetition, or features learned from labelled examples. The precise model architecture and thresholds are often proprietary, so it is not responsible to claim that every detector uses the same formula.
Predictable prose is not proof of automation
A person writing a formal report, an assignment introduction or a standard business email may use familiar phrases and conventional sentence structures. Those choices can make the prose statistically predictable. Yet predictable writing is common in human communication: technical manuals, legal clauses, academic abstracts, grant applications and routine corporate reports all rely on repeated conventions. A detector that associates certain patterns with generated text can therefore encounter human writing that shares those features.
The reverse is also possible. A language model can produce varied, unusual or highly personal-sounding prose, especially when given examples of a particular author’s style. A detector is not measuring whether a passage is original, insightful or truthful. It is estimating whether its observed features resemble examples in the detector’s training and evaluation environment.
Short samples magnify uncertainty
A short paragraph contains fewer observations than a full essay or report. One unusual phrase, a heading, a list, a formulaic opening or a few repetitive sentences can have an outsized effect on a small sample. GPTZero’s published FAQ says its document-level classification is generally more accurate than paragraph-level classification, which in turn is more accurate than sentence-level classification. That is a vendor statement, not an independent guarantee, but it reinforces a sensible rule: avoid treating a tiny excerpt as a reliable basis for an authorship judgement.
Genre, language and editing change the conditions
A detector evaluated on general English essays may not perform equally well on scientific writing, second-language English, translated text, highly structured prose or hybrid documents. A 2026 study in the International Journal for Educational Integrity found that accuracy varied with text length and genre, with lower performance on scientific writing than on humanities texts in its tested corpus. This does not mean every scientific paper will be falsely flagged. It means that a score must be interpreted in relation to the kind of text the tool was tested on.
Editing also complicates classification. Human editors routinely improve grammar, consistency and flow. Translation software and accessibility tools can change surface patterns without changing who developed the ideas. Conversely, AI-generated material can be revised by a person. A final document may therefore contain several kinds of contribution that a binary human-versus-AI label cannot describe adequately.
What Recent Research Says About Error Rates
There is no single universal false-positive rate for “AI detectors”. Results depend on which product was tested, which version was used, how the human texts were collected, what counted as a positive result, how long the samples were and whether the evaluation was independent. A responsible article should report the conditions beside the number, rather than turning one benchmark into a claim about every classroom or workplace.
A July 2026 Epoch AI evaluation tested Pangram, GPTZero and Originality.ai on human-written passages as well as AI-generated passages. In its set of 495 human passages, Pangram and GPTZero returned zero false positives, while Originality.ai incorrectly flagged 19 passages, a rate of 3.84%. The result is informative but bounded: it describes those detectors, versions, samples and decision rules in that experiment. It should not be read as a permanent product-wide error rate.
A separate 2026 study by Hadra, Cambridge and Mesbah evaluated Turnitin and Originality using 192 texts, including pre-generative-AI student writing, professional human writing, AI-generated writing and hybrid compositions. The study reported overall accuracy of 0.61 for Turnitin and 0.69 for Originality in its setup. Both tools struggled with hybrid texts, and performance declined with changes in text length and genre. The authors’ result is not a universal ranking: it is evidence that accuracy can vary substantially when the task moves away from clean, fully human or fully generated samples.
A 2025 study by Cheng and colleagues compared detection tools and human reviewers across different forms of AI-generated writing. It found that tool scores varied, with inter-tool reliability ranging from 0.57 to 0.95, and concluded that detectors should not be relied upon alone. The study is useful because it shows that even when tools can distinguish some experimental conditions, their absolute scores may not agree.
The 2026 literature is not uniformly negative. Van Vlasselaer, Van Droogenbroeck and Spruyt reported that false positives were rare in their tested dataset, while also cautioning against using detectors as sole evidence in high-stakes decisions. These findings do not cancel one another out. They show why evaluation design, corpus composition and the detector version must be disclosed before a numerical result can be compared.
Selected 2025–2026 evidence (results apply only to each study’s tested conditions):
| Source | Sample / scope | Reported finding | Interpretation |
| Epoch AI, July 2026 | 495 human passages; 3 detectors | 0/495 false positives for Pangram and GPTZero; 19/495 (3.84%) for Originality.ai | A controlled benchmark, not a universal rate. |
| Hadra et al., February 2026 | 192 texts; human, AI and hybrid; 2 tools | Overall accuracy: 0.61 Turnitin; 0.69 Originality; hybrid text was difficult | Genre and composition affect performance. |
| Cheng et al., November 2025 | Multiple AI-use conditions and tools | Inter-tool score reliability ranged from 0.57 to 0.95 | Scores from different products are not interchangeable. |
| Van Vlasselaer et al., June 2026 | Synthetic test set plus 1,163 theses without known ground truth | False positives rare in the tested tools and dataset | Promising results still do not justify sole reliance. |
What the Major Detector Providers Say
Vendor documentation is not independent validation, but it is essential for understanding what a product claims, how its score should be read and what limitations the provider acknowledges. The most useful policy statements are often more cautious than the marketing language surrounding detection.
Turnitin
Turnitin’s AI Writing Report guidance states that its model may misidentify human-written, AI-generated and AI-paraphrased text. It says the report should not be used as the sole basis for adverse action against a student and calls for further scrutiny and human judgement alongside institutional policy. Turnitin also explains that the AI percentage is separate from the similarity score. Similarity checks and AI-writing detection answer different questions: matching text against sources is not the same as estimating whether text resembles model output.
Turnitin’s documentation also describes a threshold treatment for low scores. Its current guidance says scores from 1% to 19% are not surfaced as a numerical percentage in the report, in order to reduce potential false-positive incidents; they are represented with an asterisk. The company’s release notes show that the model has continued to change, including an October 2026 English-model update intended to improve detection while minimising false positives. The same notes say that previously generated reports are not retroactively updated, so a new report may differ from an older one.
GPTZero
GPTZero’s FAQ describes its classifier as a tool for identifying possible AI use and says its results should not be used to punish students. It acknowledges edge cases in both directions and recommends reviewing a student’s writing process, drafts and understanding. Its October 2026 support documentation further notes that accuracy improves with more text and that the system can sometimes flag highly procedural or other machine-generated text. It also cautions that its training data is predominantly English prose written by adults, which is relevant when interpreting results for other genres and populations.
GPTZero’s published material also distinguishes document-level classifications such as Human, Mixed and AI from the confidence attached to them. The displayed percentage is not the proportion of words written by a machine. Readers should check the current report definitions rather than assume that a 70% result means 70% of the document is AI-authored.
Originality.ai
Originality.ai’s help centre acknowledges that human-written content can receive a high AI score and lists short text, structured writing, editing and translation among conditions that deserve careful interpretation. Its public support pages also describe common false-positive concerns. These are useful statements about the product’s own guidance, but vendor-authored explanations should not be substituted for independent testing.
Originality.ai’s results in independent evaluations have varied with the dataset and benchmark. Epoch AI’s July 2026 test found 19 false positives among 495 human passages. Hadra and colleagues also found that performance changed across text types and that hybrid writing was challenging. These are not necessarily contradictory: the studies used different corpora, test designs and evaluation criteria.
How to interpret vendor guidance:
| Provider | Documented caution | Practical implication |
| Turnitin | AI report may misidentify text; not sole basis for adverse action. | Treat the result as a lead for review, not a verdict. |
| GPTZero | Edge cases exist; results should not be used to punish students. | Review full documents and process evidence. |
| Originality.ai | Human text can score highly, especially in short, structured, edited or translated samples. | Check the sample and editing context before interpreting a score. |
Who May Face a Higher Risk of Misclassification?
It is tempting to create a fixed list of people whose writing is likely to be falsely flagged. The evidence does not support a universal ranking of individuals or groups. It is more accurate to identify writing conditions that may deserve additional care and then ask whether a particular detector has been independently tested on them.
Second-language writers are one important case. Some earlier studies found that certain detectors disproportionately classified essays by non-native English writers as AI-generated. However, later studies have not always reproduced the same pattern, and results vary by detector, corpus and model generation. The correct conclusion is not that every non-native English writer is at high risk. It is that fairness must be tested across language backgrounds rather than assumed from an overall accuracy score.
Highly formulaic and procedural writing is another condition worth checking. A methods section, lab report, legal clause, standard operating procedure or structured abstract may naturally use repeated terminology and conventional syntax. A tool trained on a different distribution of text may find those features unusual or machine-like. GPTZero explicitly warns that highly procedural text can sometimes be flagged.
Short submissions and isolated paragraphs also deserve caution because they provide fewer signals. Translated and heavily edited text may be difficult to classify because the final wording does not map neatly to a single authoring process. A person may draft in one language, translate into another, then revise with a human editor. A single score cannot reliably separate those contributions.
Finally, hybrid writing is increasingly common. A writer might use a spelling assistant, ask a colleague to edit a paragraph, use a grammar tool, or incorporate a permitted AI-generated outline. Whether any of those actions violates a policy depends on the rules. The detector’s job is not to decide that policy question. Institutions should distinguish prohibited generation from permitted assistance and should tell writers what forms of support must be disclosed.
Why a High Score Is Not Proof of Misconduct
A score is a model output, not a direct observation of authorship. It does not identify the keyboard used, establish the time at which a sentence was created, or reveal whether a writer used an AI system. It cannot independently prove intent, concealment or a breach of a school or employer rule.
The difference between probability and proof is especially important when the consequence is serious. A false accusation can affect a student’s academic record, scholarship, professional reputation or relationship with an institution. A responsible process should therefore require corroborating evidence and give the writer a meaningful opportunity to respond.
The score also has to be interpreted in light of prevalence. Even a detector with a seemingly low false-positive rate can generate a meaningful number of false alarms when the share of genuinely AI-generated work in the reviewed population is low. This is a base-rate problem. If most submitted work is human-written, a small error rate among that large group may produce enough false positives to matter. A model’s accuracy on a balanced benchmark does not automatically tell an institution how often a flagged real-world submission is actually AI-generated.
Another problem is that different tools may disagree. A document might be labelled human by one detector, mixed by another and AI-likely by a third. Averaging those percentages does not solve the issue because the tools may use different training data, thresholds and definitions. Agreement can be informative, but it still does not turn a classifier into direct proof.
There is also a versioning problem. Vendors update their models. A score generated last semester may not match the score generated today, and some vendors state that old reports are not retroactively updated. Any formal review should record the tool name, model or report version where available, date of analysis, sample submitted and the exact report. Without that record, a later reviewer may be unable to reproduce the result.
A Fair Process for Reviewing a Flagged Document
A sound review process should begin with the document’s history and the relevant policy, not with an assumption that the detector is correct. The following workflow is designed for educators, editors and organisations that need to investigate a concern without treating an automated score as a verdict.
1. Preserve the original report
Save the complete report, not only a screenshot of the headline percentage. Record the product, date, model version if available, document length, language, file type and the exact text submitted. Preserve the original document. Do not repeatedly edit and rescan the only copy, because that can make it difficult to establish what was originally reviewed.
2. Check what the score measures
Read the provider’s current documentation. Determine whether the score refers to the whole document, qualifying passages, a confidence estimate, or the estimated proportion of text in a category. Check whether the product reports mixed authorship, paraphrased text or a minimum sample requirement. Never compare percentages from different products as if they shared one scale.
3. Review the text in context
Look at the full document and its genre. Is the flagged material a short introduction, a bibliography, a methods section, a template, a set of bullet points or a translated passage? Does the writing contain domain-specific terminology or standard phrasing? A flagged excerpt should be read alongside the surrounding argument and the writer’s previous work, where appropriate and permitted.
4. Ask for process evidence
Drafts, outlines, research notes, source annotations, tracked changes and version history can help reconstruct how the work developed. These artefacts are not infallible—some editors do not save every intermediate version, and a writer may draft offline—but they are closer to the actual process than a final-text classifier. The goal is to understand the work, not to demand that every writer use one particular software platform.
5. Invite a neutral explanation
If questions remain, ask the writer to explain their argument, sources, decisions and revisions. A short conversation or oral explanation can help establish understanding. The questions should be specific and non-accusatory. A person should not be expected to prove innocence merely because a software product generated a high score.
6. Apply the written policy consistently
The final decision should be made by a responsible human reviewer using the applicable rules and all available evidence. The institution should distinguish between prohibited AI generation, permitted proofreading, translation support and accessibility assistance. If the policy is unclear, the uncertainty should not be resolved by treating the detector’s classification as a substitute for policy.
This process is also more defensible for publishers and employers. A detector can direct attention to a passage that merits fact-checking or editorial review, but claims about authorship should be supported by records, disclosures and direct evidence.
How to Reduce the Chance of a Misleading Result
Writers should not have to distort their natural style to satisfy an automated classifier. The aim should be to preserve a transparent record of authorship and follow the relevant rules, not to “beat” a detector. There is no reliable, ethical way to guarantee that every detector will classify a human-written document as human.
Keep a normal drafting trail. Save outlines, notes, successive drafts and references as part of the ordinary writing process. If you use a collaborative editor, retain version history where appropriate. If you work in a local word processor, save dated versions rather than overwriting the only file. These records are useful for ordinary editing and can help resolve later questions.
Be transparent about permitted assistance. If a course, journal or employer requires disclosure of AI-assisted translation, proofreading, brainstorming or drafting, follow that requirement. Policies differ, so do not assume that one institution’s rules apply elsewhere. A grammar checker and a generative writing assistant may be treated differently.
Submit a sufficiently complete sample when the tool’s instructions require one. Do not scan a single sentence and treat the result as representative of an entire report. Avoid uploading confidential or personal material to public detectors unless the service’s privacy terms and your organisation’s rules permit it.
If a document is flagged, do not make arbitrary edits solely to change the score. Rewriting to manipulate a detector can damage clarity and may create a misleading record. Instead, preserve the result, check the tool’s interpretation guidance and provide authentic evidence of the writing process.
For teachers and institutions, the best prevention is clear assessment design. Explain what AI assistance is allowed, provide examples of acceptable disclosure, teach students how to document their process and avoid making a detector score an automatic trigger for punishment. A well-designed assessment can include drafts, reflection, discussion, source evaluation and an opportunity to explain decisions.
What Detector Benchmarks Still Do Not Tell Us
Benchmarks are essential, but they are not a complete substitute for real-world validation. A detector may perform well on a balanced set of clearly human and clearly AI-generated samples yet behave differently on a class of student assignments, translated reports or mixed-authorship documents. A benchmark should therefore be read as a test of specified conditions, not as a guarantee about every future document.
Ground truth is a central challenge. Researchers need to know whether a text is human-written or AI-generated. Pre-2022 writing can provide a useful human baseline, but it may not represent today’s writing practices, editing tools or language models. A text labelled “human” may also have been proofread, translated or collaboratively edited. Conversely, an AI-generated text may be heavily revised. Researchers should describe how they established authorship and what kinds of assistance were permitted.
The choice of threshold matters too. A strict threshold may reduce false positives but miss more AI-generated writing. A more sensitive threshold may catch more generated text while flagging more authentic writing. There is no threshold that removes the trade-off in every context. A vendor’s “accuracy” figure may also combine false positives and false negatives, which can obscure the particular risk that matters to a school or publisher.
Genre balance is another issue. If a test set contains mostly general essays, it may not provide enough evidence about code explanations, scientific abstracts, legal writing, non-native English, poetry or short-form content. Language coverage matters for the same reason. A product should be evaluated on the actual population and material in which it will be used.
Finally, benchmark reproducibility can be limited by proprietary models. Researchers may not know the complete training corpus, internal threshold or preprocessing pipeline. A vendor update can change results without changing the user interface. Independent evaluations should report dates, versions, sample construction, confidence intervals and the exact decision rule wherever possible.
The Practical Decision: When Should You Trust a Detector?
A detector is most defensible as a low-stakes screening aid when the sample is long enough, the language and genre are within the tool’s validated scope, the result is interpreted according to the vendor’s documentation, and a human reviewer checks the wider evidence. It is least defensible as an automatic decision-maker for a short excerpt, a high-stakes allegation or a writing category that has not been adequately tested.
For students, a high score should prompt a request for clarification rather than an immediate conclusion. For teachers, it can be one reason to examine drafts or discuss the assignment, but it should not replace professional judgement. For publishers, it may justify editorial questions about disclosure and process, but not a claim of misconduct without corroboration. For businesses, it may support quality assurance, but it should not be used as a shortcut for performance management.
The most useful question is not “Which detector gives the biggest percentage?” It is “What evidence would establish how this document was produced?” That shift moves the discussion from a contest between opaque scores to a review that can be explained, challenged and repeated.
A detector may be one part of that review. It should not be the whole case.
Our Editorial Verification Process
This article was prepared as an explanatory review of AI-writing detector false positives. Research was conducted on 6 November 2026 using current vendor documentation and published research available at the time. The review prioritised primary product guidance from Turnitin and GPTZero, help-centre material from Originality.ai, and peer-reviewed or identifiable research publications from 2025–2026.
For quantitative claims, the article reports the sample size, tested products and conditions when those details were available. The Epoch AI July 2026 benchmark is described as a specific test of 495 human passages, not a universal error rate. The Hadra, Cambridge and Mesbah study is described as a 192-text evaluation of Turnitin and Originality, with its reported accuracy figures kept within that study’s scope. The Cheng et al. 2025 findings are used to illustrate inter-tool variation rather than to claim that all products perform identically.
The review did not conduct a new laboratory benchmark of Turnitin, GPTZero, Originality.ai or Pangram. It therefore makes no claim of hands-on testing by the Perplexity AI Magazine editorial team, and does not invent reproducibility results or direct quotations from named industry figures. Vendor statements are attributed to the vendors; independent study findings are attributed to their authors. Where a result depends on a particular dataset or model version, that limitation is stated.
Sitemap access limitation: the website sitemap endpoints supplied in the brief were not accessible through the research environment. Consequently, this draft does not contain fabricated internal URLs. The final internal-link pass must use verified URLs from the live sitemap or a supplied export.
This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. Factual claims and citations are attributed to the sources listed below. This article does not use unverified expert quotations or unverified commercial pricing claims.
Conclusion
AI detection tools can give false positives on human writing. The size of the risk is not fixed: it changes with the detector, model version, sample length, language, genre, threshold and evaluation dataset. Recent studies include both low false-positive findings and measurable misclassifications, while vendor guidance from Turnitin and GPTZero explicitly warns against using a detector result as the sole basis for a serious decision.
That evidence does not make every detector useless. A classifier may help identify text that deserves a closer look, especially when the sample is substantial and the product has been evaluated on comparable material. But a score cannot establish authorship, intent or misconduct by itself. Nor can a percentage be compared casually across products whose definitions and thresholds differ.
The more reliable approach is procedural: preserve the report, examine the full document, review the policy, consider drafts and version history, and give the writer a fair opportunity to explain the work. Institutions should evaluate detectors on their own writing populations and publish clear rules for permitted assistance. Until those safeguards are routine, AI detection should remain a supporting signal rather than a verdict.
Frequently Asked Questions
Q: Can AI detection tools give false positives on human writing?
A: Yes. AI detectors can classify genuinely human-written text as AI-generated because they infer authorship from textual patterns rather than observing how the document was created. The risk varies by tool, version, language, genre, sample length and threshold.
Q: Does a 90% AI score mean 90% of my writing was generated by AI?
A: Not necessarily. Some tools use the percentage to express confidence in a document-level classification, not the proportion of words written by AI. Always check the specific provider’s score definition and report guidance.
Q: Can Turnitin falsely flag human-written work?
A: Turnitin says its AI writing model may misidentify human-written, AI-generated and AI-paraphrased text. Its guidance says the report should not be the sole basis for adverse action against a student.
Q: Why do AI detectors disagree about the same document?
A: Different tools use different models, training data, thresholds and definitions. A document can therefore receive different classifications from different products. Their percentages should not be averaged as if they shared one scale.
Q: Are short paragraphs more difficult to detect accurately?
A: They can be. Short samples contain fewer signals, so a small number of conventional or unusual sentences can disproportionately affect the result. GPTZero says document-level classification is generally more accurate than paragraph- or sentence-level classification.
Q: Can non-native English writing be falsely flagged?
A: Some studies have reported disproportionate false positives for particular non-native English datasets, while other studies have not found the same pattern. Risk depends on the detector and test conditions, so broad claims about all non-native writers are not justified.
Q: What should I do if my human-written essay is flagged?
A: Save the full report and original file, preserve drafts and version history, review the detector’s documentation, and ask for the institution’s review process. Explain your writing process calmly and provide authentic supporting records.
Q: Should schools use AI detectors to prove cheating?
A: No detector score should be treated as proof on its own. A fair process should combine the report with policy, process evidence, the student’s explanation and human judgement.
References
- Cheng, A., Lin, Y., Reedy, G., Joseph, C., Wirkowski, S., Mallette, V., Nagesh, V., Krieser, D., & Calhoun, A. (2025). Ability of AI detection tools and humans to accurately identify different forms of AI-generated written content. Advances in Simulation, 10, 66. View publication
- Epoch AI. (2026, July 15). AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors. View publication
- Hadra, M., Cambridge, K., & Mesbah, M. (2026). Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity, 22, Article 4. View publication
- Turnitin. (n.d.). Using the AI Writing Report. Turnitin Guides. View publication
- Turnitin. (2026). AI writing detection model: 2026 release notes. Turnitin Guides. View publication
- GPTZero. (n.d.). Frequently asked questions. View publication
- GPTZero Support. (2026, October 20). What are the limitations of GPTZero’s AI classifier? View publication
- Originality.ai. (n.d.). Human-written content flagged as AI: What to do next. View publication
- Van Vlasselaer, M., Van Droogenbroeck, F., & Spruyt, B. (2026). Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity, 22, Article 16. View publication
- Angelier, V. (2026). The pitfalls of AI detection in academic writing: Bias, false positives, and the need for inclusive assessment. AI and Ethics, 6, Article 532. View publication
Published
Spotted an error? Read our corrections policy or email editorial@perplexityaimagazine.com.
