How to Detect AI-Generated Text

Fundamentals of AI Text Detection

Learning how to detect AI-generated text requires understanding the statistical mechanics of Large Language Models (LLMs) compared to the cognitive processes of human writers. AI models—such as OpenAI's GPT series, Anthropic's Claude, or Google's Gemini—generate text by predicting the most statistically probable next token (a word, subword, or character) given a preceding sequence. Because they optimize for coherence, safety, and average plausibility across vast training corpora, their outputs exhibit distinct mathematical and stylistic footprints.

Human writing, by contrast, is driven by communicative intent, individual idiosyncrasy, irregular pacing, and contextual memory. While modern LLMs produce fluent, grammatically flawless prose, they consistently default to predictable linguistic paths. Knowing how to check if something was written by AI involves combining automated statistical analysis with close forensic reading of stylistic, structural, and factual anomalies.

+--------------------------------------------------------------------------+
|                       THE DUAL DETECTION FRAMEWORK                       |
+------------------------------------+-------------------------------------+
|         Statistical / Metric       |         Stylistic / Forensic        |
+------------------------------------+-------------------------------------+
| • Perplexity (Predictability)      | • Structural Symmetry & Pacing      |
| • Burstiness (Sentence Variance)   | • Overused Transitions & Clichés    |
| • N-gram Frequency Distributions   | • Emotional Flatness & Neutrality   |
| • Cryptographic Watermarking       | • Hallucinated Citations & Data     |
+------------------------------------+-------------------------------------+

Mathematical and Statistical Principles of AI Text

Automated detection tools and linguistic researchers evaluate machine-generated text using several fundamental statistical metrics. Understanding these concepts explains both how automated detectors function and why human writing differs from machine output.

Perplexity

Perplexity measures how "surprised" a language model is by a sequence of words. Mathematically, it reflects the inverse probability of a text sample under a reference model:

$$\text{Perplexity}(W) = \exp\left( -\frac{1}{N} \sum_{i=1}^{N} \log P(w_i \mid w_1, w_2, \dots, w_{i-1}) \right)$$

• Low Perplexity: The text relies heavily on high-probability word choices. LLMs naturally produce low-perplexity text because sampling algorithms (such as top-$p$ or temperature scaling) prioritize tokens with high statistical likelihood.
• High Perplexity: The text contains unusual word pairings, unexpected metaphors, slang, or non-standard syntax typical of human creativity, error, or domain-specific jargon.

Burstiness

Burstiness refers to the variation in sentence length, complexity, and rhythm across a document.

• Human Text: Characterized by high burstiness. A human writer might follow a dense, 45-word compound-complex sentence with a punchy, three-word fragment, followed by a medium-length passive construction.
• AI Text: Characterized by low burstiness. LLMs produce uniform sentence lengths and predictable syntactic structures, creating a monotonic cadence even when the vocabulary is varied.

N-Gram Frequency and Token Distribution

When choosing words, LLMs select from probability distributions across their vocabulary. In standard configurations, tokens in the top 10% to 20% of probability mass are selected repeatedly. Human writers frequently choose words from the long tail of the distribution—words that are grammatically valid and contextually rich but statistically rare in massive web crawls.

Stylistic and Linguistic Markers of AI Writing

When evaluating a document manually to check for AI writing, several structural and qualitative markers frequently appear. While no single marker provides definitive proof, the co-occurrence of multiple patterns strongly suggests synthetic generation.

1. Structural Predictability and Symmetrical Formatting

LLMs are trained via Reinforcement Learning from Human Feedback (RLHF) to provide helpful, comprehensive, and organized responses. This produces predictable formatting habits:

• Formulaic Paragraphing: Essays or responses often follow a rigid pattern: an introductory paragraph ending with a thesis-like roadmap, three to five uniformly sized body paragraphs each starting with a clear topic sentence, and a concluding paragraph beginning with "In conclusion," "Ultimately," or "Overall."
• Excessive List-Making: When asked an open-ended question, models default to bulleted or numbered lists with bold headings, even when a narrative or analytical format would be more natural.
• Symmetrical Point Distribution: Pro-and-con arguments or comparative sections are given almost identical word counts and balanced treatments, avoiding decisive stances unless explicitly commanded.

2. Overused Rhetorical Transitions and Vocabulary

Due to RLHF alignment and specific training sets, certain words and phrases appear in AI text with disproportionate frequency. Common lexical signals include:

• Transitional Crutches: Frequent use of delve, testament, tapestry, landscape, beacon, paramount, foster, pivotal, crucial, furthermore, moreover, and it is important to note.
• Summary Signposts: Concluding sections heavily favor phrases like In summary, At the end of the day, It remains to be seen, or serves as a reminder that.
• Metaphorical Tropes: Complex concepts are routinely described as a "tapestry," "dance," "double-edged sword," or "ever-evolving landscape."

3. Tone Neutrality and Universal Hedging

To prevent bias, offensive output, and inaccuracy, aligned models use defensive phrasing and balanced neutrality:

"While some argue that remote work increases productivity, others contend that it hinders collaboration. Ultimately, the ideal approach depends on individual organizational needs and personal preferences."

This hedging manifests as an unwillingness to take strong positions, an absence of genuine personal perspective, and a tendency to present all sides of an issue as equally valid regardless of factual consensus.

4. Semantic Shallowness Despite Lexical Sophistication

AI models frequently generate text that sounds impressive at a glance but delivers minimal substantive information upon closer inspection. This phenomenon—sometimes termed "fluent vacuity"—involves using multi-syllabic vocabulary to restate the premise of a prompt without introducing concrete examples, historical specifics, or novel logical steps.

[Prompt: Explain why the Roman Republic fell.]

AI Typical Output:
"The decline of the Roman Republic was a multifaceted process driven by internal 
tensions, economic transformations, and sociopolitical evolution. As institutional 
frameworks struggled to adapt to changing realities, shifting balances of power 
reshaped the governance landscape, leading to foundational changes."

Analysis:
Grammatically flawless, yet contains zero concrete data: no names (Sulla, Caesar, 
Gracchi), no specific laws, no specific economic crises (latifundia), and no dates.

5. Hallucinations and Source Fabrication

Because LLMs predict tokens rather than querying relational databases of ground truth, they can generate plausible-sounding falsehoods:

• Invented Citations: Fabricating academic papers by combining real author names with plausible journal titles and fabricated DOIs.
• Anachronisms: Blending historical figures, dates, or scientific mechanisms that sound logically coherent together but are factually incorrect.
• Non-Existent Quotes: Generating verbatim quotations from real public figures that were never spoken or recorded.

Automated Detection Methods and Technologies

Automated AI detectors use several distinct technical architectures to flag synthetic content. Each method possesses distinct operational profiles, strengths, and failure modes.

+--------------------------------------------------------------------------------+
|                         AUTOMATED DETECTION MECHANISMS                         |
+-----------------------+--------------------------------+-----------------------+
| Zero-Shot Classifiers | Supervised Fine-Tuned Models   | Watermarking Schemes  |
+-----------------------+--------------------------------+-----------------------+
| Evaluates Perplexity  | Trained on paired datasets     | Cryptographic bias    |
| and Burstiness via a  | of human vs. AI text           | embedded in token     |
| base LLM (e.g., GPT-2)| (e.g., RoBERTa classifiers)    | selection logits      |
+-----------------------+--------------------------------+-----------------------+

1. Zero-Shot Perplexity Analyzers

Tools like early iterations of GPTZero run target text through an open-weight reference model (such as GPT-2 or Llama). The tool calculates the log-likelihood of each token given its context. If the average perplexity is low and burstiness across sentences is uniform, the text is flagged as likely machine-generated.

2. Supervised Fine-Tuned Classifiers

These detectors fine-tune a pre-trained transformer (such as RoBERTa or DeBERTa) on millions of paired human and AI documents across various domains. The model learns subtle high-dimensional embeddings and contextual relationships that distinguish human writing styles from machine outputs.

3. Statistical Watermarking

Watermarking operates at the generation stage rather than post-hoc inspection. During text generation:

1. The model's pseudo-random number generator divides its vocabulary into a "green list" and a "red list" based on the preceding token's hash.
2. The model introduces a slight mathematical bias (a logit bump) favoring green-list tokens.
3. A human reader cannot perceive the bias, but an inspection tool with the cryptographic key can calculate whether the proportion of green-list tokens exceeds statistical probability ($p < 0.001$).

While highly accurate, watermarking requires the generating organization (e.g., OpenAI, Google) to embed the watermark directly during API inference, rendering it ineffective against open-source models run locally.

Limitations, False Positives, and Evasion Techniques

Automated text detectors are probabilistic, not deterministic. Relying on them as sole arbiters of academic or professional integrity presents significant operational and ethical risks.

The False Positive Problem

A false positive occurs when human writing is incorrectly classified as AI-generated. This risk is unevenly distributed:

• Non-Native English Speakers: Studies have demonstrated that text written by non-native speakers receives significantly higher AI-probability scores. Non-native writers often use more restricted vocabularies, simpler sentence structures, and predictable transitions, directly mimicking the statistical properties of LLMs.
• Technical and Formulaic Writing: Legal briefs, medical reports, standard operating procedures, and scientific abstracts naturally require standardized phrasing and low perplexity, causing elevated false-positive rates.
• Short Samples: Texts under 250 words lack sufficient statistical mass for reliable perplexity or burstiness calculations, rendering automated classification highly unstable.

Adversarial Evasion Strategies

Generating text that bypasses automated detectors requires minimal technical effort:

1. Prompt Engineering: Instructing a model to "write with varied sentence lengths, use informal phrasing, incorporate rare idioms, and avoid transitional words" drastically alters perplexity and burstiness.
2. Paraphrasing Tools: Running AI output through secondary paraphrasers (e.g., QuillBot) or translation loops (e.g., English $\rightarrow$ German $\rightarrow$ English) scrambles the token sequence while preserving semantic meaning.
3. Manual Insertion of Typographical Idiosyncrasies: Adding occasional spelling errors, non-standard punctuation, or conversational parentheticals breaks statistical uniformity.
4. Hybrid Composition: Interweaving human-written sentences with AI-generated paragraphs dilutes the classifier's confidence score below detection thresholds.

Forensic Verification Framework

When evaluating a text sample where authorship is in question—such as an academic essay, professional report, or published article—use a multi-stage verification framework instead of relying on a single detection score.

+--------------------------------------------------------------------------+
|                       FORENSIC TRIAGE WORKFLOW                           |
+--------------------------------------------------------------------------+
                                     |
                                     v
              +----------------------------------------------+
              |  Step 1: Check Minimum Thresholds            |
              |  - Is the sample > 250 words?                |
              |  - Is it free of dense technical formulas?   |
              +----------------------------------------------+
                                     |
                                     v
              +----------------------------------------------+
              |  Step 2: Run Multi-Tool Statistical Analysis |
              |  - Test across 2-3 distinct detector engines |
              |  - Note Perplexity and Burstiness graphs     |
              +----------------------------------------------+
                                     |
                                     v
              +----------------------------------------------+
              |  Step 3: Conduct Fact & Citation Audits      |
              |  - Verify DOIs, page numbers, and quotes     |
              |  - Check for historical or factual blends    |
              +----------------------------------------------+
                                     |
                                     v
              +----------------------------------------------+
              |  Step 4: Author Baseline & Process Evidence  |
              |  - Compare against verified past writing     |
              |  - Review edit history / version control     |
              +----------------------------------------------+

Step 1: Establish Baseline Viability

Before running any analysis, confirm that the sample is long enough (at least 250–300 words) and that the genre is suitable for evaluation. Highly structured formats (such as technical specifications or recipe steps) should not be evaluated using statistical classifiers.

Step 2: Multi-Engine Statistical Cross-Check

Submit the text to multiple detection systems that use different underlying architectures (e.g., a fine-tuned classifier alongside a perplexity-based engine). Treat the results not as a binary "verdict," but as an indicator of whether further manual investigation is justified.

Step 3: Citation and Fact Audit

Perform rigorous spot-checks on references and claims within the text:

• Query cited journal articles directly in databases like PubMed, IEEE Xplore, or Google Scholar.
• Confirm whether cited page numbers, volume numbers, and publication years align accurately.
• Verify whether specific statistics, data points, or historical anecdotes correspond to real-world sources.

Step 4: Process-Oriented and Longitudinal Verification

In educational and enterprise environments, the most reliable way to verify authorship is through process-based evidence rather than post-hoc text analysis:

• Version History: Inspect Google Docs or Microsoft Word version histories to confirm realistic drafting behavior (e.g., incremental additions, revisions, deletions) rather than massive blocks of text pasted in a single timestamp.
• Baseline Comparison: Compare the suspected text against authenticated prior writing by the same author. Look for significant shifts in vocabulary range, syntactic complexity, preposition usage, and punctuation habits.
• Oral Defense / Knowledge Interview: Ask the author to explain the thesis, define complex vocabulary used in the piece, or describe the research process behind specific claims. An author who used AI to generate their work will often struggle to explain its nuanced arguments or terminology.

Institutional and Policy Considerations

Organizations, academic institutions, and publishers that deploy AI text detection must establish clear evidentiary standards and transparent policies.

+--------------------------------------------------------------------------+
|                   RESPONSIBLE DETECTION GOVERNANCE                       |
+--------------------------------------------------------------------------+
| 1. Never rely on automated AI detector scores as sole disciplinary proof.|
| 2. Require corroborating evidence (version logs, citation errors, interviews).|
| 3. Account for language background to avoid bias against non-native writers.|
| 4. Define acceptable use boundaries (brainstorming vs. drafting vs. editing).|
+--------------------------------------------------------------------------+

1. Prohibit Sole-Source Penalization: Automated detector outputs should serve exclusively as an investigative triage mechanism, never as definitive proof of misconduct.
2. Establish Clear Use Definitions: Distinguish clearly between prohibited generation (e.g., generating entire paragraphs from scratch) and permitted assistance (e.g., grammar checking, brainstorming, translation assistance, or outline generation).
3. Maintain Human Appeal Pathways: Provide individuals accused of unauthorized AI use with transparent access to the evidence against them and clear avenues to demonstrate authorship through draft histories and oral interviews.

What AI-text detection can and cannot establish

How to detect AI generated text is not a matter of applying one perfectly reliable test. Current detection methods can identify signals that are more common in machine-generated writing, but they usually cannot prove authorship from the text alone. A human may write in a highly predictable style, edit or paraphrase AI output, or use grammar and translation tools; conversely, an AI-generated passage may be carefully revised until it resembles ordinary human writing. The strongest assessment therefore combines several kinds of evidence: the text itself, the writing process, document history, source checking, and the context in which the work was produced.

An AI detector should be treated as an indicator for further review, not as a final verdict. This distinction matters particularly in education, employment, publishing, research, legal disputes, and disciplinary proceedings, where a false accusation can have serious consequences.

Why identifying AI writing is difficult

Generative language models produce text by predicting likely sequences of words from patterns learned during training. That process can result in writing that is fluent, grammatically correct, and organized, but none of those properties is unique to AI. Human editors, professional writers, second-language writers, and people using templates can produce similar prose.

Detection is also complicated by the fact that authorship is often mixed. A person may ask an AI system for an outline, draft a passage themselves, use an automated grammar checker, translate the result, and then revise it. In such a case, the meaningful question is not simply whether AI was involved, but what role it played and whether that role complied with the relevant rules.

Several factors reduce confidence in text-only judgments:

• Short samples contain little evidence. A paragraph or a few sentences may not provide enough stylistic variation for a meaningful assessment.
• Editing changes the signals. Rewriting, shortening, adding personal examples, or correcting factual errors can make generated prose harder to distinguish.
• Different models and prompts produce different styles. A detector calibrated for one type of output may perform differently on another.
• Language and genre matter. Formal reports, technical explanations, legal writing, and standardized school essays naturally use recurring structures and vocabulary.
• Detectors can produce false positives. Human writing may be labeled as likely AI-generated, especially when it is concise, highly polished, formulaic, or written by someone using a non-native variety of English.
• Detection tools can be evaded. Light paraphrasing, translation, deliberate variation in sentence structure, and human revision may alter the patterns a detector looks for.

For these reasons, a score such as likely AI or a percentage estimate should not be interpreted as the percentage of the text that was actually written by AI. It is generally a statistical classification, and its meaning depends on the tool, sample, language, and threshold used.

Textual signs that may justify closer examination

When people ask how to tell if something is written by AI, they often mean how to identify unusual features in the prose. These features can be useful clues, but none is conclusive by itself. The right approach is to look for a cluster of inconsistencies rather than one familiar symptom.

Generic organization and unusually even prose

AI-generated text often follows a predictable structure: a broad opening, a sequence of balanced points, and a neutral closing statement. It may use headings and transitions that are logically tidy but not especially specific to the subject. Phrases such as “it is important to note,” “in conclusion,” or “by understanding these factors” are not proof of AI use; they are common in human writing too. Their significance increases only when they appear repeatedly alongside other peculiarities.

Some generated passages maintain an unusually consistent level of fluency. Every paragraph may be similar in length, sentence rhythm, and degree of formality, with few spontaneous turns of phrase. Human drafts often contain some unevenness: a particularly detailed section, a brief aside, a change in rhythm, or a sentence that reflects the writer’s individual priorities. However, professional editing can create the same consistency.

Vague specificity and interchangeable examples

A generated answer may sound informative while avoiding concrete details. It may refer to “various studies,” “many experts,” or “real-world applications” without naming the relevant study, expert, organization, event, or application. Examples may be generic enough to fit almost any topic and may not reflect the actual circumstances described in the assignment.

This is not merely a stylistic issue. Vague language can indicate that the writer has not engaged deeply with the subject. It can also reveal that the text was produced from a broad prompt rather than from firsthand knowledge or assigned sources.

Unsupported or invented claims

Fact-checking is often more useful than stylistic speculation. AI systems can produce fabricated citations, incorrect quotations, nonexistent publications, and confident claims that do not follow from the sources. Check whether:

• named sources actually exist;
• cited authors made the claimed argument;
• quotations match the original wording;
• dates, figures, technical terms, and proper names are accurate;
• links lead to the stated material;
• the conclusion is supported by the evidence rather than merely asserted.

An incorrect fact does not prove AI authorship, since people also make mistakes. It does, however, provide a concrete reason to investigate the work and the writer’s research process.

Mismatch between language and apparent understanding

A passage may use advanced vocabulary and polished transitions while misunderstanding a basic concept. It may define a term correctly in one paragraph and use it incorrectly in another, or answer a related question rather than the question that was asked. Contradictions, irrelevant generalizations, and abrupt shifts in meaning are especially useful signals because they concern substance, not just style.

Repeated stock phrasing and unusual formatting

Some generated writing includes repeated sentence patterns, excessive headings, symmetrical lists, or a conspicuous preference for certain transitional phrases. It may also over-explain simple points while failing to address an important constraint in the prompt. Again, these are clues rather than proof. Templates, editing guides, and common online writing conventions can produce the same patterns.

A more reliable checking process

A sound review should move from low-risk observations to stronger evidence. It should also give the writer an opportunity to explain the work before any judgment is made.

1. Read the work against the task

First determine whether the text actually satisfies the assignment or purpose. Look for the required argument, sources, examples, audience, length, and limitations. AI-generated material often responds to the general topic but misses a specific instruction, local context, class discussion, internal terminology, or required method.

Compare the level of detail with what the writer was asked to do. A broad, polished answer that ignores the assigned reading may be less credible than a less elegant answer containing accurate, relevant engagement with that reading.

2. Check facts and sources independently

Verify a sample of important claims rather than relying on the text’s confidence. Start with claims that can be tested directly: names, quotations, statistics, dates, citations, technical definitions, and descriptions of events. If the work includes references, inspect the original sources rather than assuming that a plausible-looking bibliography is genuine.

Source verification can reveal poor research regardless of who wrote the text. The appropriate response may be correction, a request for documentation, or a discussion of research practices—not an automatic finding of AI use.

3. Compare the text with the writer’s established work

A meaningful change in vocabulary, sentence structure, depth, spelling, or subject-matter understanding may warrant questions. Comparison is most useful when the samples are similar in genre, length, topic, time period, and editing conditions. A casual message should not be compared directly with a carefully revised report.

Style comparison is inherently uncertain. People write differently for different audiences and may improve substantially over time. It should therefore prompt a conversation or request for process evidence, not serve as conclusive proof.

4. Examine the writing process

Process evidence is often stronger than a detector score. Depending on the setting, useful evidence may include:

• notes, outlines, drafts, and revision history;
• versioned files or document-editing timelines;
• research notes and source annotations;
• saved prompts and AI output, when disclosure is required or voluntarily provided;
• the ability to explain why particular evidence and wording were chosen;
• a short supervised writing sample or oral explanation.

No single item is decisive. A writer may compose offline, lose drafts, or use a different application. Conversely, an extensive revision history can be produced without proving that every sentence was authored by the person. Process evidence is valuable because it permits a more direct, fair assessment of how the work came together.

5. Ask specific, open questions

If clarification is appropriate, ask the writer to explain a claim, source, calculation, structural choice, or revision. Questions should test understanding rather than invite a forced confession. For example, ask why one source was considered more relevant than another, how a particular conclusion follows from the evidence, or what changed between two drafts.

A person who cannot explain their own argument may have submitted work they did not meaningfully produce or understand. That still does not establish exactly which tool was used, and it may also reflect poor preparation, anxiety, language barriers, or a misunderstanding of the assignment.

AI-detection software: uses and limitations

Automated detectors generally look for statistical or stylistic properties of text. Some estimate how predictable the word choices are to a language model; others compare patterns of phrasing, sentence length, vocabulary, or structure with examples labeled as human- or machine-generated. Commercial and institutional tools may combine several signals, but their exact methods and performance can change over time.

A detector result should be recorded with its context: the tool used, date, language, text length, document type, and threshold. Without that information, a score is difficult to interpret or reproduce.

Why detector scores are not proof

Detection systems can misclassify both categories. They may flag human writing because it is formulaic or highly edited, and they may miss AI writing after revision or paraphrasing. Performance can be weaker for languages, dialects, and genres that were not well represented in the tool’s development or evaluation. Short text is particularly problematic because a small number of stylistic choices can dominate the result.

A detector may also be unable to distinguish among different kinds of automated assistance. Text produced by a chatbot, translated by software, corrected by a grammar tool, completed by predictive typing, or copied from a template may be treated differently even though the relevant policy treats these activities separately.

How to use a detector responsibly

If a tool is used, its result should be one input in a documented review, not the sole basis for punishment or rejection. Good practice includes:

1. Preserve the original text and the exact report produced by the tool.
2. Avoid presenting a probability or label as a factual percentage of AI-authored content.
3. Check whether the tool supports the language and type of writing being examined.
4. Seek corroborating evidence from sources, drafts, revision history, and discussion.
5. Give the writer a clear opportunity to respond.
6. Apply the same standard to comparable cases.
7. Protect submitted text and personal information according to the relevant privacy requirements.

Policies differ by school, employer, publisher, and jurisdiction. Some permit brainstorming or proofreading but prohibit generated passages; others require disclosure of any automated assistance. The governing policy should be identified before evaluating whether use was acceptable.

Distinguishing AI assistance from prohibited substitution

The phrase AI-generated text can conceal several different activities. A useful review separates them rather than treating all automation as equivalent.

The appropriate question is often whether the person can demonstrate the required skill and whether their use complied with the applicable rules. A polished final document alone cannot answer that question.

Common mistakes when trying to detect AI writing

Several popular approaches produce unreliable conclusions.

Treating style as a fingerprint. There is no universally identifiable AI style. Human writing overlaps with generated writing, and model output varies by prompt and revision.

Using one suspicious phrase as proof. Stock wording is common across textbooks, websites, and professional prose. It can indicate formulaic writing but cannot establish its source.

Assuming factual errors prove AI use. People hallucinate, misremember, cite poorly, and misunderstand sources. Errors should be investigated on their own terms.

Relying on a detector without reading the work. Automated output can obscure obvious contextual evidence and may encourage unjustified certainty.

Ignoring legitimate language differences. Non-native speakers and writers using regional dialects can be unfairly flagged by systems that equate standardization with human authorship.

Confronting someone with an accusation instead of evidence. This can make a fair explanation less likely and may violate institutional procedures. A neutral request for drafts, sources, or an explanation is more informative.

Assuming metadata is conclusive. File timestamps, editing histories, and application metadata can be incomplete, altered, or affected by synchronization and copying. They are supporting evidence, not an infallible record.

Better ways to establish authorship and responsible use

The most effective long-term solution is often to design a process that makes authorship visible. In education, this can include staged submissions, annotated sources, brief reflections on revisions, in-class writing, oral explanation, and assignments requiring local or personal analysis. In workplaces and publishing, tracked revisions, source records, documented review, and clear disclosure requirements can serve a similar purpose.

These measures are preferable to trying to identify every generated sentence after the fact. They also support legitimate uses of AI while preserving accountability. A writer who uses an AI system for permitted assistance should retain enough information to describe what the system did, verify factual claims, and identify which ideas and wording they adopted or rejected.

For high-stakes decisions, text-only detection should not be treated as definitive evidence. Follow the relevant institutional procedure and, where consequences are significant, seek review from a qualified person familiar with authorship assessment, language variation, privacy, and due process. The central goal is not to label text based on appearance, but to make a fair, evidence-based determination about the work, the process behind it, and the rules that apply.