Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Digital Forensic Investigation of the ChatGPT Windows Application

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ChatGPT Windows chats survive app deletion, study finds

desk verdict Useful artifact map for ChatGPT Windows forensics on unencrypted boxes, but the persistence-after-deletion claim overstates what a single no-reboot capture can show. read the letter →

arxiv 2505.23938 v1 pith:GC4X2M5V submitted 2025-05-29 cs.CR

Malithi Wanniarachchi Kankanamge , Nick McKenna , Santiago Carmona , Syed Mhamudul Hasan , Abdur R.Shahid , Ahmed Imteaj This is my paper
classification cs.CR
keywords digitalforensicsChatGPTWindowsapplicationgenerativeAIartifactrecoveryRAMcaptureregistryanalysisnetworktrafficdeleteddata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper attempts to establish that the ChatGPT Windows desktop application leaves a recoverable digital footprint on an unencrypted Windows machine: chat prompts, generated responses, user interactions, and system-level traces can be found in RAM, the registry, deleted-file space, exported data, and network traffic. The authors argue that this evidence persists even after the application is uninstalled, and that standard, mostly free forensic tools can retrieve it. If this is right, investigators do not need special access to OpenAI's servers to reconstruct how a suspect used ChatGPT on a Windows computer; they can recover the material from the device itself. This matters because generative-AI chat tools are increasingly implicated in phishing, fraud, and data-manipulation scenarios, and this is the first published forensic study of the Windows version.

What carries the argument

The recovery machinery is the application's own local data habits on an unencrypted file system: chat content and prompts are written into memory, cache, configuration, and export files rather than kept only on OpenAI's servers. The study combines five tools to harvest these remnants — Autopsy for disk-image and timeline analysis, FTK Imager for imaging and deleted-file recovery, Magnet RAM Capture for volatile memory, Wireshark for packet capture, and Hex Workshop for byte-level inspection of exported chat data. Each tool covers a different storage layer, so when one source is wiped the others still corroborate the same facts about what the user asked ChatGPT.

What would settle it

On a Windows 11 machine with BitLocker enabled, install ChatGPT, create chats, delete the app, then image the drive and capture RAM; if the disk artifacts are unreadable and RAM shows no chat remnants after reboot, the claim that evidence persists does not generalize.

Watch

Extended reading notes

Core claim

On its own terms, the paper shows that artifacts of ChatGPT use can be extracted from at least five independent sources on a Windows 11 Enterprise machine: RAM snapshots, the Windows registry, forensic disk images including unallocated space, ChatGPT-exported HTML and JSON files inspected at byte level, and live network traffic. The central finding is persistence: RAM captures taken before and after uninstalling the app both contain the fabricated phishing email and other prompt content, registry entries describing ChatGPT survive deletion and full power-off, and deleted-file space holds residual data. The authors therefore claim that a suspect's ChatGPT conversation history and interaction timeline can be reconstructed with ordinary forensic tooling without relying on cloud access.

Load-bearing premise

The conclusions depend on the suspect machine being a single unencrypted Windows 11 system running the tested app version; the paper does not test full-disk encryption.

Editorial extensions

If this is right

  • Investigators can recover suspect prompts and ChatGPT-generated responses from RAM even after the app has been uninstalled, as long as the machine was not rebooted.
  • Registry remnants of ChatGPT installation and configuration persist across app deletion and power cycles, providing corroborating evidence of the app's presence and use.
  • Exported chat data inspected at the byte level exposes metadata, timestamps, and shared-file information that can link multiple suspects.
  • Live network capture identifies the suspect machine's IP, the OpenAI and Cloudflare server IP, and the protocols used, placing the user in direct communication with ChatGPT at a given time.
  • Because the tools include free and open-source options, the workflow is available to resource-limited investigative units.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • On a BitLocker-encrypted drive, most disk-based recovery (registry, unallocated space, exported files) would likely fail; the paper's persistence claim is probably limited to unencrypted setups, so live RAM capture becomes the primary option in encrypted cases.
  • The same 'chat JSON in local app data' pattern likely extends to other desktop GAI assistants built on similar Windows packaging, so this workflow could seed comparable studies for those apps.
  • Recovered prompts could be mined for jailbreak or prompt-injection patterns, turning a suspect's own chat logs into evidence about how the model's safeguards were circumvented.
  • A testable extension is to repeat the study under different Windows versions and ChatGPT releases to map which artifact locations are version-stable and which are implementation details.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper presents a forensic examination of the ChatGPT Windows desktop application on a single Windows 11 Enterprise test machine. The authors install the app, issue three simulated criminal prompts (phishing email generation, credit-card image modification, and employee-record manipulation), and then attempt to recover evidence using FTK Imager disk images, Magnet RAM Capture memory dumps taken before and after application deletion, Windows Registry inspection, Wireshark packet captures, and Hex Workshop analysis of ChatGPT's built-in data export. The main reported findings are that ChatGPT-generated text, user prompts, and some metadata can be found in RAM, registry entries associated with the application survive deletion and power-off, network traffic can be linked to ChatGPT/Cloudflare endpoints, and exported HTML/JSON chat data contain recoverable metadata. The abstract and Section 6.2 go further, claiming that chat history and system-level traces persist even after deletion.

Significance. The paper addresses a timely and under-studied target, and its value is best understood as an existence result: on an unencrypted Windows 11 machine, standard free tools can locate ChatGPT-related strings, registry keys, network metadata, and exported chat files. The tool-version table, the chain-of-custody protocol, and the use of realistic malicious prompts are useful for practitioners and for later replication. The main weakness is that the persistence-after-deletion claim exceeds the evidence: no post-reboot capture, no defined uninstall, no published hashes, and an explicit scope exclusion of disk encryption. If the authors tighten the claims and add the missing controls, this can be a solid reference for ChatGPT Windows forensics; in its current form, the broad wording in the abstract and Section 6.2 is not supported.

major comments (3)
  1. [Abstract; §6.1; Table 2] The abstract's claim that chat history and system-level traces 'can be recovered even after deletion' is not demonstrated by the reported data. Section 6.1 states that the RAM evidence was collected 'in the scenario where the suspect computer is not restarted,' so the pre-deletion strings found in the 'after deletion' RAM capture are expected volatile remnants rather than evidence of persistence. Table 2 lists 'ChatGPT deletion from Windows' without specifying whether the uninstaller was used and whether user data directories such as %APPDATA% and %LOCALAPPDATA% were removed. To support the persistence claim, the authors must define the complete deletion procedure, take a RAM capture after a reboot, and report what survives on disk; otherwise the conclusion should be limited to 'recoverable before reboot.'
  2. [§6.2; Figure 4] Section 6.2 interprets registry entries as evidence that 'chat in ChatGPT ... will persist' across deletion and power-off, but the text and Figure 4 describe file locations and configuration settings, not chat content. A registry key that records the application path is not a chat log. The authors should either show the specific registry values that contain prompt/response data or explicitly reframe Section 6.2 as evidence of application-installation persistence rather than chat-history persistence.
  3. [§4.1.2; §8] The general statement that 'the evidence will persist' is based on a single run on one Windows 11 Enterprise machine with no full-disk encryption, and §8 explicitly excludes encryption from scope. The paper should therefore qualify the abstract and Section 6.2 conclusions as applying to unencrypted Windows 10/11 systems with the tested app build. The existence of artifacts on one machine is a useful contribution, but the universal 'will persist' wording overstates the evidence.
minor comments (6)
  1. [Table 1] Hex Workshop is listed as open-source, but it is a commercial product; the license type should be corrected.
  2. [§1.3] The sentence 'related work that is presented in discussed in Section 2' is ungrammatical and should be rewritten.
  3. [Figure 3] The caption repeats '2)' for both the before-deletion and after-deletion panels, making the figure difficult to interpret.
  4. [§4.1.1; Table 1] The tested ChatGPT Windows application version or build is not reported; Table 1 reports tool versions but not the app version.
  5. [§4.2.1; §7] No SHA-256 or MD5 hashes, disk-image names, or RAM-capture file details are provided despite the chain-of-custody discussion; adding them, or a supplemental data link, would make the forensic claims independently verifiable.
  6. [§6.5] Hex Workshop is described as analyzing 'exported binary data,' but the artifacts described are HTML and JSON files; the terminology should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's forensic findings are direct observations from a test system, with no fitted parameters or derivation chain that reduces to its own inputs.

full rationale

This is an empirical digital forensics study, not a derivation-based paper. The central claims are that ChatGPT-related artifacts, chat text, registry entries, and network metadata can be recovered using standard forensic tools on an unencrypted Windows 11 test machine. These claims rest on direct observations from RAM captures, disk images, registry views, and packet captures, documented in Sections 6.1 through 6.5. There are no equations, fitted parameters, or statistical models whose outputs are equivalent to their inputs by construction. The paper does contain one self-citation, [19] by Hasan, Shahid, and others, but it appears in the related-work discussion of AI-driven threat identification and is not load-bearing for any forensic claim; it is a normal literature citation. The limitations noted in Sections 6.1 and 8—that the suspect machine was not restarted and that full-disk encryption was out of scope—are honest scope conditions, and concerns about generalizability are soundness or validity risks, not circularity. No self-definitional step, renamed existing result, or author-imported uniqueness theorem is present. The 'after deletion' persistence finding is an empirical result from that specific environment, and while it may be overstated if generalized, it does not reduce to the paper's assumptions by construction. Therefore the circularity score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

This is an empirical forensics case study, so the ledger contains environmental assumptions rather than fitted constants or new entities. The main burden is the single unencrypted testbed and the absence of a published baseline.

assumptions (3)
  • domain assumption Single unencrypted test environment is representative of real suspect devices.
    Sec. 4.1.2 and Sec. 8. One Windows 11 Enterprise machine, no full-disk encryption; the authors explicitly scope out encryption, so recoverability under BitLocker or other protection is untested.
  • domain assumption Implicit legal permission to image the suspect's device.
    Sec. 4.3: the authors assume all necessary approvals and permissions are implicitly granted. This shapes the chain-of-custody claims.
  • domain assumption No pre-existing ChatGPT artifacts on the test machine before the experiment.
    No baseline disk or RAM capture before installation is reported, so the observed traces are assumed to originate from the simulated session.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Digital Forensic Investigation of the ChatGPT Windows Application." pith.science (2026). https://pith.science/paper/GC4X2M5V

@misc{pith2026250523938,
  author       = {Pith},
  title        = {Pith review of: Digital Forensic Investigation of the ChatGPT Windows Application},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GC4X2M5V}},
  note         = {Machine review of arXiv:2505.23938}
}
read the original abstract

The ChatGPT Windows application offers better user interaction in the Windows operating system (OS) by enhancing productivity and streamlining the workflow of ChatGPT's utilization. However, there are potential misuses associated with this application that require rigorous forensic analysis. This study presents a holistic forensic analysis of the ChatGPT Windows application, focusing on identifying and recovering digital artifacts for investigative purposes. With the use of widely popular and openly available digital forensics tools such as Autopsy, FTK Imager, Magnet RAM Capture, Wireshark, and Hex Workshop, this research explores different methods to extract and analyze cache, chat logs, metadata, and network traffic from the application. Our key findings also demonstrate the history of the application's chat, user interactions, and system-level traces that can be recovered even after deletion, providing critical insights into the crime investigation and, thus, documenting and outlining a potential misuse report for digital forensics.

Figures

Figures reproduced from arXiv: 2505.23938 by the authors.

Figure 1
Figure 1. User Interface (UI) of the Chat￾GPT Windows Application: (a) ChatGPT icon on windows taskbar after installa￾tion, (b) the main interface of the Chat￾GPT application window, (c) file upload and screenshot options for multimedia in￾put, and (d) additional capabilities, includ￾ing image generation using DALL-E, per￾forming web search, and facilitating collab￾orative writing or coding sessions. Although ChatGPT can supp… view at source ↗
Figure 2
Figure 2. Prompt results: 1) results for regular email, 2) results for credit card [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Screenshot of RAM Capture Results 1)ChatGPT generated text 2) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Screenshot of Hex Workshop analysis: a) Prompt conversations b) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 4
Figure 4. Figure 4: Screenshot of windows registry results: a) before ChatGPT deletion [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Screenshot of Wireshark Results Showing Multiple Instances of Communication with ChatGPT. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The screenshot of Autopsy data analysis (a) Log files and (b) timeline data. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages

  1. [8]

    Dragonas, C

    E. Dragonas, C. Lambrinoudakis, P. Nakoutis, Forensic analysis of ope- nai’s chatgpt mobile application, Forensic Science International: Digital Investigation 50 (2024) 301801

    work page 2024

  2. [36]

    Scanlon, F

    M. Scanlon, F. Breitinger, C. Hargreaves, J.-N. Hilgert, J. Sheppard, Chat- gpt for digital forensic investigation: The good, the bad, and the unknown, Forensic Science International: Digital Investigation 46 (2023) 301609

    work page 2023

  3. [1]

    URL https://openai.com/chatgpt

    OpenAI, Chatgpt: Large language model, accessed: Jan 29, 2025 (2023). URL https://openai.com/chatgpt

    work page 2023

  4. [2]

    Y . Cao, S. Li, Y . Liu, Z. Yan, Y . Dai, P. Yu, L. Sun, A survey of ai- generated content (aigc), ACM Comput. Surv. 57 (5) (Jan. 2025). doi: 10.1145/3704262. URL https://doi.org/10.1145/3704262

    doi:10.1145/3704262 2025

  5. [3]

    M. P. Rogers, H. M. Hillberg, C. L. Groves, Attitudes towards the use (and misuse) of chatgpt: A preliminary study, in: Proceedings of the 55th ACM Technical Symposium on Computer Science Education V . 1, 2024, pp. 1147–1153

    work page 2024

  6. [4]

    SA ¸ VULESCU, L

    ¸ S. SA ¸ VULESCU, L. IV AN, The impact of chatgpt on cybercrime and on the activities carried out by the law enforcement structures., Romanian Journal of Forensic Science 136 (136) (2023) 358–364

    work page 2023

  7. [5]

    Fui-Hoon Nah, R

    F. Fui-Hoon Nah, R. Zheng, J. Cai, K. Siau, L. Chen, Generative ai and chatgpt: Applications, challenges, and ai-human collaboration (2023)

    work page 2023

  8. [6]

    Guleria, K

    A. Guleria, K. Krishan, V . Sharma, T. Kanchan, Chatgpt: forensic, legal, and ethical issues, Medicine, Science and the Law 64 (2) (2024) 150–156

    work page 2024

Show all 39 references
  1. [7]

    A. Aleti, Software testing of generative ai systems: Challenges and op- portunities, in: 2023 IEEE /ACM International Conference on Software Engineering: Future of Software Engineering (ICSE-FoSE), 2023, pp. 4–

    2023

  2. [9]

    A. R. Javed, W. Ahmed, M. Alazab, Z. Jalil, K. Kifayat, T. R. Gadekallu, A comprehensive survey on computer forensics: State-of-the-art, tools, techniques, challenges, and future directions, IEEE Access 10 (2022) 11065–11089. doi:10.1109/ACCESS.2022.3142508

    2022

  3. [10]

    Lee, H.-Y

    J.-H. Lee, H.-Y . Kwon, Large-scale digital forensic investigation for win- dows registry on apache spark, Plos one 17 (12) (2022) e0267411

    2022

  4. [11]

    A. Raza, M. Hussain, H. Tahir, M. Zeeshan, M. A. Raja, K.-H. Jung, Forensic analysis of web browsers lifecycle: A case study, Journal of Information Security and Applications 85 (2024) 103839

    2024

  5. [12]

    D. S. I. Utomo, Y . Prayudi, E. Ramadhani, Forensic web analysis on the latest version of whatsapp browser, Journal of Computer Networks, Ar- chitecture and High Performance Computing 5 (1) (2023) 359–367

    2023

  6. [13]

    Bandal, S

    S. Bandal, S. Rath, Unveiling digital secrets: An image text vision app for enhanced digital forensics investigations, in: 2024 12th International Symposium on Digital Forensics and Security (ISDFS), IEEE, 2024, pp. 1–6

    2024

  7. [14]

    doi:10.1109/ICSE-FoSE59343.2023.00009

    2023

  8. [15]

    J. Allison, Network packet analysis as a unit of assessment: Identifying emotet, in: Proceedings of the 22nd Koli Calling International Conference on Computing Education Research, 2022, pp. 1–2

    2022

  9. [16]

    A. D. Kadage, B. M. Nayak, V . S. Hingmire, K. Wanjale, N. Bogiri, P. L. Mandale, Ai-enhanced digital forensics: Automated techniques for e ffi- cient investigation and evidence collection, Journal of Electrical Systems 20 (1s) (2024) 211–229

    2024

  10. [17]

    A. A. Solanke, M. A. Biasiotti, Digital forensics ai: evaluating, standard- izing and optimizing digital evidence mining techniques, KI-Künstliche Intelligenz 36 (2) (2022) 143–161

    2022

  11. [18]

    Dixit, P

    P. Dixit, P. Bhattacharya, S. Tanwar, R. Gupta, Anomaly detection in au- tonomous electric vehicles using ai techniques: A comprehensive survey, Expert Systems 39 (5) (2022) e12754

    2022

  12. [19]

    Stamp, M

    M. Stamp, M. Alazab, A. Shalaginov, Malware analysis using artificial intelligence and deep learning, V ol. 1, Springer, 2021

    2021

  13. [20]

    S. M. Hasan, A. M. Alotaibi, S. Talukder, A. R. Shahid, Distributed threat intelligence at the edge devices: A large language model-driven approach, in: 2024 IEEE 48th Annual Computers, Software, and Appli- cations Conference (COMPSAC), 2024, pp. 1496–1497. doi:10.1109/ COMPSA...

    2024 arXiv

  14. [21]

    Mena, Investigative data mining for security and criminal detection, Butterworth-Heinemann, 2003

    J. Mena, Investigative data mining for security and criminal detection, Butterworth-Heinemann, 2003

    2003

  15. [22]

    Schneider, F

    J. Schneider, F. Breitinger, Towards ai forensics: Did the artificial intelli- gence system do it?, Journal of Information Security and Applications 76 (2023) 103517

    2023

  16. [23]

    Bijani, D

    S. Bijani, D. Robertson, A review of attacks and security approaches in open multi-agent systems, Artificial Intelligence Review 42 (2014) 607– 636

    2014

  17. [24]

    Yadav, S

    S. Yadav, S. Yadav, P. Verma, S. Ojha, S. Mishra, Artificial intelligence: An advanced evolution in forensic and criminal investigation, Current Forensic Science 1 (1) (2023) e190822207706

    2023

  18. [25]

    Dunsin, M

    D. Dunsin, M. C. Ghanem, K. Ouazzane, V . Vassilev, A comprehen- sive analysis of the role of artificial intelligence and machine learning in modern digital forensics and incident response, Forensic Science Inter- national: Digital Investigation 48 (2024) 301675

    2024

  19. [26]

    Galante, R

    N. Galante, R. Cotroneo, D. Furci, G. Lodetti, M. B. Casali, Applica- tions of artificial intelligence in forensic sciences: Current potential ben- efits, limitations and perspectives, International journal of legal medicine 137 (2) (2023) 445–458

    2023

  20. [27]

    Ivanova, S

    M. Ivanova, S. Stefanov, Regarding artificial intelligence in digital foren- sic investigation: Applications and solutions, in: 2024 XXXIII Interna- tional Scientific Conference Electronics (ET), IEEE, 2024, pp. 1–6

    2024

  21. [28]

    R. J. Dinis-Oliveira, R. M. Azevedo, Chatgpt in forensic sciences: a new pandora’s box with advantages and challenges to pay attention (2023)

    2023

  22. [29]

    B. Dash, P. Sharma, Are chatgpt and deepfake algorithms endangering the cybersecurity industry? a review, International Journal of Engineering and Applied Sciences 10 (1) (2023) 21–39

    2023

  23. [30]

    S. Jia, R. Lyu, K. Zhao, Y . Chen, Z. Yan, Y . Ju, C. Hu, X. Li, B. Wu, S. Lyu, Can chatgpt detect deepfakes? a study of using multimodal large language models for media forensics, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, 2024, pp...

    2024

  24. [31]

    S. A. Shahzad, A. Hashmi, Y .-T. Peng, Y . Tsao, H.-M. Wang, How good is chatgpt at audiovisual deepfake detection: A comparative study of chat- gpt, ai models and human perception, arXiv preprint arXiv:2411.09266 9 (2024)

    2024 arXiv

  25. [32]

    Cingillioglu, Detecting ai-generated essays: the chatgpt challenge, The International Journal of Information and Learning Technology 40 (3) (2023) 259–268

    I. Cingillioglu, Detecting ai-generated essays: the chatgpt challenge, The International Journal of Information and Learning Technology 40 (3) (2023) 259–268

    2023

  26. [33]

    Halford, A

    E. Halford, A. Webster, Using chat gpt to evaluate police threats, risk and harm, International Journal of Law, Crime and Justice 78 (2024) 100686

    2024

  27. [34]

    Nikolakopoulos, S

    A. Nikolakopoulos, S. Evangelatos, E. Veroni, K. Chasapas, N. Gousetis, A. Apostolaras, C. D. Nikolopoulos, T. Korakis, Large language models in modern forensic investigations: Harnessing the power of generative artificial intelligence in crime resolution and suspect identific...

    2024

  28. [35]

    Sreya, M

    E. Sreya, M. Wadhwa, et al., Enhancing digital investigation: Leverag- ing chatgpt for evidence identification and analysis in digital forensics, in: 2023 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS), IEEE, 2023, pp. 733–738

    2023

  29. [37]

    K. Cho, Y . Park, J. Kim, B. Kim, D. Jeong, Conversational ai forensics: A case study on chatgpt, gemini, copilot, and claude, Forensic Science International: Digital Investigation 52 (2025) 301855

    2025

  30. [38]

    GDPR, General data protection regulation, Regulation (EU) 679 (2016)

    G. GDPR, General data protection regulation, Regulation (EU) 679 (2016)

    2016

  31. [39]

    Department of Justice, JM 9-48.000 - Computer Fraud, accessed: 2024-01-30 (2024)

    U.S. Department of Justice, JM 9-48.000 - Computer Fraud, accessed: 2024-01-30 (2024). URL https://www.justice.gov/jm/ jm-9-48000-computer-fraud 10

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.