Elkhatat et al. 2023: evaluating AI content detection tools

This record is one of eight in the studies index. Elkhatat 2023 is the smallest study indexed here and the one that most directly compares GPT-3.5 with GPT-4 on the same prompt. The paper was received on April 30, 2023 and accepted on June 30, 2023 (journal page).

Elkhatat 2023 tested five detectors. The paper names no product versions or access dates, so the table carries only the detector and its card.

Detectors as named in Elkhatat et al. 2023, DOI 10.1007/s40979-023-00140-5.

Every text in Elkhatat 2023 was about one engineering topic. The prompt to ChatGPT was "write around 100 words on the application of cooling towers in the engineering process", and the human controls came from the same subject (Springer PDF).

Sample as described in Elkhatat et al. 2023, journal page.

In Elkhatat 2023, detector outputs were normalised to the five-tier scale used by the OpenAI classifier: under 20% AI "very unlikely", 20 to 40 "unlikely", 40 to 60 "unclear", 60 to 80 "possibly", over 80 "likely". Each response was then coded positive, negative, false positive, false negative or uncertain, and sensitivity, specificity, positive predictive value and negative predictive value were computed per detector per GPT version (DOI 10.1007/s40979-023-00140-5).

Elkhatat 2023 reports the diagnostic figures as percentages in a chart and lists every response in Table 4 of the paper. The counts below are tallied from the rows of that table; the percentages the paper states in its text are given alongside where they exist (Springer PDF).

Counts from Table 4 and percentages from the text of Elkhatat et al. 2023, DOI 10.1007/s40979-023-00140-5.

The authors' first finding is a version gap: Elkhatat 2023 states that the tools were more accurate on GPT-3.5 content than on GPT-4 content (journal page). On GPT-3.5, the OpenAI classifier had the highest sensitivity in Elkhatat 2023 at 100%, but its specificity and negative predictive value were the lowest at 0%, because the classifier flagged all 5 human controls. GPTZero was the balanced performer on GPT-3.5 in Elkhatat 2023 at 93% sensitivity and 80% specificity, while Writer and Copyleaks struggled with sensitivity by the authors' account. The GPT-4 question, added 9 November 2026, lists Copyleaks, GPTZero and Writer as the three cards this record scored on GPT-4 text.

On GPT-4, Elkhatat 2023 reports generally lower results: Copyleaks had the highest sensitivity at 93%, CrossPlag kept 100% specificity, and the OpenAI classifier showed substantial sensitivity with no specificity (Springer PDF). Table 4 of Elkhatat 2023 shows the size of the drop for two tools: GPTZero went from 14 of 15 GPT-3.5 texts to 4 of 15 GPT-4 texts, and CrossPlag from 15 of 15 to 6 of 15.

The human controls produced the paper's warning. In Elkhatat 2023, Writer rated two of the five human paragraphs "Likely AI-Generated" and a third "Uncertain", and GPTZero rated one human paragraph "Likely AI-Generated" (DOI 10.1007/s40979-023-00140-5). The authors conclude that detection tools should not be the sole determinant in academic integrity cases.

The Elkhatat 2023 authors name the sample first: five tools, content generated only by ChatGPT 3.5 and 4, and only five human-written controls, which they say could affect the findings. Every text addressed one engineering topic in about 100 words, so the results speak to short technical prose (DOI 10.1007/s40979-023-00140-5).

The study was conducted at a specific time, and the authors expect tool performance to have evolved since. No product versions or access dates are given in the paper, and no corrigendum is recorded.

Three of the five detectors in Elkhatat 2023 have a card in this index; the Writer card was added on 9 November 2026 for a detector withdrawn on 22 December 2025. Each card records the GPT-3.5, GPT-4 and human counts with the base of 15, 15 and 5 texts (DOI 10.1007/s40979-023-00140-5).

All three appear in the ranking of detectors by independent test record. How this record is sourced explains why counts tallied from a published table are shown with the table named. The home page lists which AI detectors exist and who has tested them.

Related records: Walters 2023 found the same GPT-4 gap across 16 detectors and 126 documents, and Weber-Wulff 2023 tested the same OpenAI classifier and two of the same tools on 54 documents in the same spring (DOI 10.1007/s40979-023-00140-5).