Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Training Users Against Human and GPT-4 Generated Social Engineering Attacks

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Human text plus GPT-4 styling is the phishing users miss most.

desk verdict Genuine preregistered 2x2 result and a real interaction, but the 'most challenging hybrid' headline outruns the pairwise statistics, and the IBL simulation is fit to the human outcome it claims to predict. read the letter →

arxiv 2502.01764 v1 pith:E53W5EHA submitted 2025-02-03 cs.HC

Tailia Malloy , Maria Jose Ferreira , Fei Fang , Cleotilde Gonzalez This is my paper · ORCID
classification cs.HC
keywords phishingdetectionsocialengineeringGPT-4instance-basedlearningAI-writingbiasanti-phishingtrainingLLMembeddingshuman-AIcollaboration
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

The paper argues that current anti-phishing training, which relies on simple human-written plain-text emails, misses the attack format users actually struggle with: emails whose text is written by a person and whose visual styling is generated by GPT-4. In a between-subjects experiment with 207 participants, this hybrid condition produced the smallest pre- to post-training improvement in categorization accuracy, while fully GPT-4-written and styled emails were the easiest. The paper also identifies an AI-writing bias: the more participants believed emails were AI-generated, the more likely they were to label any email as phishing. To improve training, the paper proposes an Instance-Based Learning model that uses GPT-4 embeddings as email features and, in simulation, shows that selecting emails the model predicts a user will misclassify improves learning over random selection. If these claims hold, anti-phishing training should treat human-AI hybrid emails as a distinct, harder category and correct for users' over-reliance on perceived AI authorship.

What carries the argument

The carrying mechanism is an Instance-Based Learning (IBL) model of the user, a cognitive model that stores each encountered email as an instance with features, chosen action, and reward, and retrieves instances by activation. Email features are the cosine similarities between GPT-4 embeddings of the current email and stored emails, so textual similarity is computed by the same method across all experimental conditions. The model computes a blended value for each choice (phishing or ham) and predicts the user's action; for training selection, it iterates over the available email set and picks the email with the highest predicted probability of being misclassified, on the theory that showing users emails they are likely to get wrong maximizes learning. The experimental machinery is the $2\times2$ between-subjects design with 40 feedback training trials between 10 pre- and 10 post-training no-feedback trials.

What would settle it

A controlled training study with human participants who receive emails chosen by the IBL model versus randomly selected emails would settle the claim; if the IBL-selected group does not show a larger increase in pre- to post-training accuracy (or the model fails to predict individual trial-by-trial choices), the simulation result is not evidence of improved training. A second check: if removing the GPT-4 styling from the human-written emails eliminates the difficulty gap, then the hard category is the formatting rather than the human-AI combination.

Watch

Extended reading notes

Core claim

The central discovery is an interaction between who writes an email's text and who creates its visual formatting. In a $2\times2$ design crossing author (human or GPT-4) with style (plain text or GPT-4-generated HTML/CSS), emails written by humans and styled by GPT-4 were the hardest for participants to categorize, with the smallest mean improvement ($\mu = 0.015$) between pre- and post-training, while GPT-4-written and styled emails showed the largest improvement ($\mu = 0.104$) and a significant interaction effect ($p < 0.001$). The paper further finds that participants' perception of AI authorship biases their decisions: a regression analysis shows that participants who identified a higher proportion of emails as AI-written categorized more emails as phishing across all conditions, with the strongest slope in the GPT-4-styled conditions. On the training side, the paper shows that an Instance-Based Learning model using GPT-4 embeddings as attributes can be calibrated to reproduce human pre-post improvement, and that using the model to select emails that it predicts a user will misclassify yields a significantly larger simulated pre-post improvement than random email selection.

Load-bearing premise

The weakest load-bearing premise is that the calibrated simulated agents of the IBL model behave like real human learners when the email selection rule changes, so the simulated benefit of adaptive selection would transfer to actual trainees.

Editorial extensions

If this is right

  • Anti-phishing training curricula should include human-written emails stylized by AI code as a separately measured and harder category, since this condition produced the smallest learning gains.
  • Training that ignores the AI-writing bias may reinforce over-flagging; presenting benign AI-generated emails as examples could reduce false positives.
  • Simple prompting of GPT-4 to write phishing emails from scratch appears less dangerous than using GPT-4 only for styling, so defensive training should prioritize hybrid attack formats.
  • IBL-based adaptive email selection, if validated with human participants, could personalize training by showing each user the emails they are most likely to misclassify.
  • Because ham-email accuracy declined in both hybrid conditions, evaluation of phishing training should measure false positives as well as detection of actual phishing.

Reading between the lines

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

  • The AI-writing bias may extend beyond email: if people treat AI-sounding content as inherently suspicious, the same over-flagging could distort decisions about social media posts, news, or workplace messages, which the paper does not test.
  • The IBL selection simulation's gain depends on the calibrated agent parameters matching real users; a natural next step is a human experiment comparing IBL-selected versus random email order, and if the gain persists, training platforms could adopt model-tracing selection.
  • The interaction effect suggests attackers need not write better text; they only need to make human text look professionally formatted. A testable consequence is that adding GPT-4 styling to any human-written email should increase its perceived legitimacy, independent of the text content.
  • If the bias is driven by visual salience of styled banners and logos, then simpler styling or plain-text rendering of the same emails should reduce the over-flagging effect, which could be checked by re-analyzing the existing stimulus sets.
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 reports a preregistered experiment (N=207) comparing end-user detection of phishing and ham emails in a 2x2 between-subjects design that varies the email text author (human vs. GPT-4) and the email styling (plain text vs. GPT-4-generated HTML/CSS). The primary outcome is the pre-to-post training improvement in categorization accuracy. The authors report a significant author-by-style interaction, with the human-written/GPT-4-styled condition showing the lowest mean improvement (μ=0.015) and the GPT-4-written/GPT-4-styled condition the highest (μ=0.104); only one post-hoc pairwise comparison (between these two cells) reached significance. The paper also reports a correlation between participants' perceived proportion of AI-written emails and their tendency to classify emails as phishing, which it labels an 'AI-writing bias.' Finally, the paper proposes an Instance-Based Learning (IBL) cognitive model that uses GPT-4 embeddings, and presents a simulation study claiming that an IBL-guided adaptive selection of training emails would improve learning outcomes relative to random selection.

Significance. The experimental question is timely and practically important: whether AI-assisted phishing emails are harder for end users to detect than purely human- or purely GPT-4-generated ones. The preregistration and public data on OSF are strengths, and the significant interaction between author and style is a novel empirical contribution. If interpreted cautiously, the results suggest that human-written emails with GPT-4-generated styling deserve attention in anti-phishing training. However, the headline claim that hybrid emails are 'the most challenging' is only weakly supported by the pairwise tests, and the IBL simulation's predictive claim rests on parameters fitted to the same human data it is used to 'predict.' These issues limit the paper's current support for its broadest conclusions.

major comments (3)
  1. [Section 3.2 and Abstract] The claim that human-written/GPT-4-styled emails are 'the most challenging' for end users, and the abstract's statement that co-created emails 'pose a greater challenge' than emails created by GPT-4 or humans alone, are not supported by the reported post-hoc tests. The significant author-by-style interaction (F=14.344, p<.001) establishes that the effect of styling differs by author, but it does not establish that the human-written/GPT-4-styled cell is significantly harder than both the human-written/plain and GPT-4-written/plain cells. The paper reports that only one pairwise comparison was significant (human-written/GPT-4-styled vs. GPT-4-written/GPT-4-styled, p=0.033); all other pairwise comparisons were not significant. The numerical lowest mean improvement alone does not license 'most challenging' in the absence of significant differences against those two conditions. The abstract and discussion should be revised to state precisely that the hybrid cell had the lowest observed improvement and was significantly lower than the fully GPT-4-generated cell, while noting that differences against the plain-text conditions were not statistically significant.
  2. [Section 3.4 and Figure 5] The IBL simulation does not provide valid evidence that adaptive email selection improves training outcomes because the simulated agents' parameters were adjusted to reproduce the same human pre-post improvement under random selection that the simulation later 'predicts.' The paper states: 'their parameters were adjusted to reflect the same pre-post training improvement that was observed in humans.' The right column of Figure 5 is therefore not an independent prediction but a consequence of calibration. No validation on held-out human data or on human behavior under the IBL selection policy is presented, and the Introduction's claim that the simulation 'reduces the effect of the AI-writing bias' goes beyond what the simulation can show. The authors' later caveat that 'additional future work is required' is accurate, but the framing of the simulation as a supportive result should be substantially softened, and the fitted parameters and their influence should be disclosed.
  3. [Section 3.3] The 'AI-writing bias' is inferred from a post-experiment self-report of perceived AI authorship and its correlation with the overall phishing classification rate in the same task. Because perceived authorship and classification behavior are measured concurrently after the main task, the regression analysis in Figure 4 cannot establish that a pre-existing bias causes misclassification; the correlation could be driven by actual email features, by a general response tendency, or by retrospective rationalization. The term 'bias' implies a systematic judgment error, but the design does not manipulate perceived AI authorship. The paper should frame this as a correlational association and acknowledge that causal claims require an experimental manipulation of participants' beliefs about AI authorship.
minor comments (6)
  1. [Abstract / Manuscript front matter] The abstract included at the beginning of the manuscript text describes a different study (delayed feedback and a Hierarchical Instance-Based Learning model) and does not match the actual abstract of this paper, which appears in the full text. Please ensure the correct abstract is used in the final version.
  2. [Section 3.2] The statistical description 'mixed repeated measure analysis of variance' is imprecise; the analysis appears to be a 2x2 between-subjects ANOVA on pre-post improvement scores (or a mixed ANOVA with time as a within-subjects factor). Please clarify the model and report the means and standard deviations for all four cells in addition to the improvement scores.
  3. [Section 3.1] The exclusion of participants whose improvement was more than two standard deviations below the mean (18 of 224 participants) is unusual because the outcome variable itself is the basis for exclusion. Please report how many excluded participants fell in each condition and discuss how this exclusion could affect the between-condition comparisons.
  4. [Sections 1 and 3] The design is described as including emails where the 'vice-versa' combination is evaluated (human-styled GPT-4-written text), but the actual GPT-4-written/plain condition is simply the unstyled GPT-4 text and does not involve human styling. Please clarify that the second hybrid condition is GPT-4-written text without HTML/CSS styling, rather than human-styled.
  5. [Section 2.4 and 3.4] The IBL model details are incomplete: the embedding model used for cosine similarity is not specified (e.g., which GPT-4 embedding model and dimensionality), and the temperature parameter τ in Eq. (2) is not defined or given a default value. Please provide these details for reproducibility.
  6. [Section 3.3] Figure 4 labels slopes and R² values but the text does not report the numerical values or confidence intervals for each condition; including these in the text would support the comparison across conditions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main experimental result is self-contained and the IBL simulation is an extrapolation, not a fitted-input prediction.

full rationale

The paper's central empirical claim is based on a preregistered between-subjects experiment comparing human and GPT-4 authorship and styling; that result is independent of any model fitting and is not circular. The IBL simulation in Section 3.4 does contain a fitting step: the authors state that 'their parameters were adjusted to reflect the same pre-post training improvement that was observed in humans,' and the simulated agents are then used to compare random email selection with IBL-based selection. This is a model-calibration and counterfactual-simulation procedure, however, not a circular reduction: the reported outcome under IBL-based selection is a distinct quantity from the human improvement that was matched, and the paper explicitly labels the result as a simulation ('This suggests that selecting emails to show participants using an IBL model may improve the quality of educational outcomes') and calls for future confirmation. The self-citations in Sections 2.3 and 2.4, such as references [29] and [30] for using LLM embeddings as IBL attributes, are methodological choices rather than load-bearing justifications of the paper's conclusions. The supplied Abstract text describes a different HIBL delayed-feedback study and contains no derivational chain that can be reduced to inputs within the provided full text, so no circularity assessment can be drawn from it. The strongest statistical concern is that the Section 3.2 claim that the human-written/GPT-4-styled condition is 'the most challenging' rests on only one significant pairwise comparison, but that is a statistical-inference or correctness issue, not circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The ledger reflects that the IBL model and its embedding-based similarity are carried over from prior work, while the training-selection policy and the calibrated simulation are introduced for this paper. The main free parameter is the set of IBL simulation parameters adjusted to match human improvement. The named AI-writing bias is a new construct supported only by within-study correlational data.

free parameters (1)
  • IBL simulation parameters (d, mu, omega_j, sigma, and possibly temperature or beta) = unspecified, adjusted to match human pre-post improvement
    Section 3.4 states the IBL simulation models' parameters were adjusted to reflect the same pre-post training improvement observed in humans. Because the adjusted values are not reported and are fit to the aggregate outcome, any improvement shown by these agents is partly built into the model rather than independently predicted.
assumptions (4)
  • domain assumption IBL activation, retrieval, and blended-value equations (Eqs. 1-3) describe how people choose in this email categorization task.
    The model is inherited from Instance-Based Learning Theory and applied to phishing training, but the paper does not validate trial-level predictions against human choices.
  • domain assumption Cosine similarity of GPT-4 embeddings is a valid measure of psychological similarity between emails for categorization.
    Section 2.4 adopts the embedding-similarity approach from the authors' prior work; no validation against human similarity judgments is reported for this email set.
  • ad hoc to paper Agents whose parameters are fit to aggregate human pre-post improvement can stand in for human learners when evaluating a different email-selection policy.
    Section 3.4 uses the calibrated simulated agents to compare random versus IBL-selected training; generalization to humans under adaptive selection is assumed, not tested.
  • ad hoc to paper Selecting the email with the highest predicted probability of incorrect categorization improves learning outcomes.
    The email selection rule et+1 = max over e of Ve'(t) in Section 2.4 assumes difficult-to-classify examples are the best training examples; this pedagogical assumption has no direct evidence in the paper.
invented entities (1)
  • AI-writing bias
    purpose: Explains why participants who believed a higher proportion of emails were AI-written categorized more emails as phishing.
    Introduced as a novel AI-writing bias in Sections 3.3 and 4, but the evidence is a correlational regression using post-hoc questionnaire responses within the same study; no external or preregistered replication is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Training Users Against Human and GPT-4 Generated Social Engineering Attacks." pith.science (2026). https://pith.science/paper/E53W5EHA

@misc{pith2026250201764,
  author       = {Pith},
  title        = {Pith review of: Training Users Against Human and GPT-4 Generated Social Engineering Attacks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E53W5EHA}},
  note         = {Machine review of arXiv:2502.01764}
}
read the original abstract

In real-world decision making, outcomes are often delayed, meaning individuals must make multiple decisions before receiving any feedback. Moreover, feedback can be presented in different ways: it may summarize the overall results of multiple decisions (aggregated feedback) or report the outcome of individual decisions after some delay (clustered feedback). Despite its importance, the timing and presentation of delayed feedback has received little attention in cognitive modeling of decision-making, which typically focuses on immediate feedback. To address this, we conducted an experiment to compare the effect of delayed vs. immediate feedback and aggregated vs. clustered feedback. We also propose a Hierarchical Instance-Based Learning (HIBL) model that captures how people make decisions in delayed feedback settings. HIBL uses a super-model that chooses between sub-models to perform the decision-making task until an outcome is observed. Simulations show that HIBL best predicts human behavior and specific patterns, demonstrating the flexibility of IBL models.

Figures

Figures reproduced from arXiv: 2502.01764 by the authors.

Figure 1
Figure 1. An example of the email identification task shown to participants as phishing, and as a result they received -1 points. While decision confidence and the action taken by the participant were collected in the experiment, only the first question determined whether participants received 1 point or -1 points. Another important feature of this experiment is that for each condition the same set of 360 base emails was used… view at source ↗
Figure 2
Figure 2. Top-Left: The original plain-text email written by human experts Bottom-Left: The GPT-4 stylized version of this original email. Bottom-Right: The fully GPT-4 rewritten and stylized version of the email. Top-Right: The stripped plain-text version of the fully GPT-4 rewritten email. This experiment included 10 pre-training trials without feedback, 40 training trials with feedback, and 10 post-training trials without … view at source ↗
Figure 3
Figure 3. Pre and post-training categorization accuracy for ham and phishing emails by experimental condition. were compensated with a base payment of $3 with the potential to earn up to a $12 bonus payment depending on performance. This experiment was approved by the Carnegie Mellon University Institutional Review Board, and the study was pre-registered on OSF2 , where all participant data and analyses are located. 3.2 Resul… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Linear regression comparing the percentage of emails categorized as being phish￾ing emails and the proportion of emails identified as being AI written. Regressions are split between each of the four experimental conditions. Shaded regions represent 95% confidence inter…
Figure 5
Figure 5. Figure 5: All improvement measures refer to the percentage point difference between pre￾training and post-training accuracy. Left: Humans participant data. Middle: Simulated IBL agents improvement under random email selection. Right: Simulated IBL agent improvement under IBL ema…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Cluster Computing 26(6), 3421– 3436 (2023)

    Al-Hawawreh, M., Aljuhani, A., Jararweh, Y.: Chatgpt for cybersecurity: practical applications, challenges, and future directions. Cluster Computing 26(6), 3421– 3436 (2023)

    work page 2023

  3. [3]

    Journal of contemporary criminal justice37(3), 427–451 (2021)

    Back, S., Guerette, R.T.: Cyber place management and crime prevention: the ef- fectiveness of cybersecurity awareness training against phishing attacks. Journal of contemporary criminal justice37(3), 427–451 (2021)

    work page 2021

  4. [4]

    Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al.: Training a helpful and harmless assistant withreinforcementlearningfromhumanfeedback.arXivpreprintarXiv:2204.05862 (2022)

    arXiv 2022

  5. [5]

    In: Proceedings of the 2023 ACM Conference on Information Technology for Social Good

    Baldassarre, M.T., Caivano, D., Fernandez Nieto, B., Gigante, D., Ragone, A.: The social impact of generative ai: An analysis on chatgpt. In: Proceedings of the 2023 ACM Conference on Information Technology for Social Good. pp. 363–373 (2023)

    work page 2023

  6. [6]

    arXiv preprint arXiv:2108.07258 (2021)

    Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

    arXiv 2021

  7. [7]

    Journal of Educational Data Mining5(1), 104–146 (2013)

    Bouchet, F., Harley, J.M., Trevors, G.J., Azevedo, R.: Clustering and profiling studentsaccordingtotheirinteractionswithanintelligenttutoringsystemfostering self-regulated learning. Journal of Educational Data Mining5(1), 104–146 (2013)

    work page 2013

  8. [8]

    Advances in neural information processing systems33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems33, 1877–1901 (2020)

    2020

Show all 44 references
  1. [9]

    Journal of behavioral decision making33(2), 220– 239 (2020)

    Burton, J.W., Stein, M.K., Jensen, T.B.: A systematic review of algorithm aversion in augmented decision making. Journal of behavioral decision making33(2), 220– 239 (2020)

    2020

  2. [10]

    arXiv preprint arXiv:2309.14348 (2023)

    Cao, B., Cao, Y., Lin, L., Chen, J.: Defending against alignment-breaking attacks via robustly aligned llm. arXiv preprint arXiv:2309.14348 (2023)

    2023 arXiv

  3. [11]

    In: 2024 IEEE 14th Annual Computing and Communication Workshop and Conference (CCWC)

    Chataut, R., Gyawali, P.K., Usman, Y.: Can ai keep you safe? a study of large language models for phishing detection. In: 2024 IEEE 14th Annual Computing and Communication Workshop and Conference (CCWC). pp. 0548–0554. IEEE (2024)

    2024

  4. [12]

    arXiv preprint arXiv:2310.14735 (2023)

    Chen, B., Zhang, Z., Langrené, N., Zhu, S.: Unleashing the potential of prompt engineering in large language models: a comprehensive review. arXiv preprint arXiv:2310.14735 (2023)

    2023 arXiv

  5. [13]

    Computers in Human Behavior 161, 108411 (2024)

    Chen, G., Dang, J., Liu, L.: After opening the black box: Meta-dehumanization matters in algorithm recommendation aversion. Computers in Human Behavior 161, 108411 (2024)

    2024

  6. [14]

    Journal of Experimental Psychology: Gen- eral 144(1), 114 (2015)

    Dietvorst, B.J., Simmons, J.P., Massey, C.: Algorithm aversion: people erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: Gen- eral 144(1), 114 (2015)

    2015

  7. [15]

    Intelligent tutoring systems in e-learning environments: Design, implemen- tation and evaluation pp

    Feng, M., Heffernan, N., Koedinger, K.: Student modeling in an intelligent tutoring system. Intelligent tutoring systems in e-learning environments: Design, implemen- tation and evaluation pp. 208–236 (2011)

    2011

  8. [16]

    Human and GPT-4 Social Engineering Attacks 17 In: Intelligent Tutoring Systems: 8th International Conference, ITS 2006, Jhongli, Taiwan, June 26-30, 2006

    Ferguson, K., Arroyo, I., Mahadevan, S., Woolf, B., Barto, A.: Improving intelligent tutoring systems: Using expectation maximization to learn student skill levels. Human and GPT-4 Social Engineering Attacks 17 In: Intelligent Tutoring Systems: 8th International Conference, IT...

    2006

  9. [17]

    In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security

    Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit con- fidence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322–1333 (2015)

    2015

  10. [18]

    Cognitive Science27(4), 591–635 (2003)

    Gonzalez, C., Lerch, J.F., Lebiere, C.: Instance-based learning in dynamic decision making. Cognitive Science27(4), 591–635 (2003)

    2003

  11. [19]

    IEEE Access (2023)

    Gupta, M., Akiri, C., Aryal, K., Parker, E., Praharaj, L.: From chatgpt to threat- gpt: Impact of generative ai in cybersecurity and privacy. IEEE Access (2023)

    2023

  12. [20]

    In: 2016 international conference on computing, communi- cation and automation (ICCCA)

    Gupta, S., Singhal, A., Kapoor, A.: A literature survey on social engineering at- tacks: Phishing attack. In: 2016 international conference on computing, communi- cation and automation (ICCCA). pp. 537–540. IEEE (2016)

    2016

  13. [21]

    a comparative literature review

    Jampen, D., Gür, G., Sutter, T., Tellenbach, B.: Don’t click: towards an effective anti-phishing training. a comparative literature review. Human-centric Computing and Information Sciences10(1), 33 (2020)

    2020

  14. [22]

    In: Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering

    Khan, J.Y., Uddin, G.: Automatic code documentation generation using gpt-3. In: Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering. pp. 1–6 (2022)

    2022

  15. [23]

    Computers in human behavior114, 106553 (2021)

    Köbis, N., Mossink, L.D.: Artificial intelligence versus maya angelou: Experimental evidence that people cannot differentiate ai-generated from human-written poetry. Computers in human behavior114, 106553 (2021)

    2021

  16. [24]

    arXiv preprint arXiv:2402.18093 (2024)

    Koide, T., Fukushi, N., Nakano, H., Chiba, D.: Chatspamdetector: Leverag- ing large language models for effective phishing email detection. arXiv preprint arXiv:2402.18093 (2024)

    2024 arXiv

  17. [25]

    arXiv preprint arXiv:2309.02705 (2023)

    Kumar, A., Agarwal, C., Srinivas, S., Li, A.J., Feizi, S., Lakkaraju, H.: Certifying llm safety against adversarial prompting. arXiv preprint arXiv:2309.02705 (2023)

    2023 arXiv

  18. [26]

    In: Proceedings of the 5th Symposium on Usable Privacy and Security

    Kumaraguru, P., Cranshaw, J., Acquisti, A., Cranor, L., Hong, J., Blair, M.A., Pham, T.: School of phish: a real-world evaluation of anti-phishing training. In: Proceedings of the 5th Symposium on Usable Privacy and Security. pp. 1–12 (2009)

    2009

  19. [27]

    In: Proceedings of the Third International Workshop on Automated Program Repair

    Lajkó, M., Csuvik, V., Vidács, L.: Towards javascript program repair with gener- ative pre-trained transformer (gpt-2). In: Proceedings of the Third International Workshop on Automated Program Repair. pp. 61–68 (2022)

    2022

  20. [28]

    Technolog- ical Forecasting and Social Change175, 121390 (2022)

    Mahmud,H.,Islam,A.N.,Ahmed,S.I.,Smolander,K.:Whatinfluencesalgorithmic decision-making? a systematic literature review on algorithm aversion. Technolog- ical Forecasting and Social Change175, 121390 (2022)

    2022

  21. [29]

    Proceedings of the Conference on Natural Language Learning (2024)

    Malloy, T., Ferreira, M.J., Fang, F., Gonzalez, C.: Leveraging a cognitive model to measure subjective similarity of human and gpt-4 written content. Proceedings of the Conference on Natural Language Learning (2024)

    2024

  22. [30]

    Frontiers in Psychology15, 1387948 (2024)

    Malloy, T., Gonzalez, C.: Applying generative artificial intelligence to cognitive models of decision making. Frontiers in Psychology15, 1387948 (2024)

    2024

  23. [31]

    Psychological review (2024)

    Malloy, T., Sims, C.R.: Efficient visual representations for learning and decision making. Psychological review (2024)

    2024

  24. [32]

    In: Proceedings of the AAAI Symposium Series

    McDonald, C., Malloy, T., Nguyen, T.N., Gonzalez, C.: Exploring the path from instructions to rewards with large language models in instance-based learning. In: Proceedings of the AAAI Symposium Series. vol. 2, pp. 334–339 (2023)

    2023

  25. [33]

    Patterns5(5) (2024)

    Park, P.S., Goldstein, S., O’Gara, A., Chen, M., Hendrycks, D.: Ai deception: A survey of examples, risks, and potential solutions. Patterns5(5) (2024)

    2024

  26. [34]

    Psychonomic bulletin & review14, 249–255 (2007) 18 Malloy et al

    Ritter, S., Anderson, J.R., Koedinger, K.R., Corbett, A.: Cognitive tutor: Applied research in mathematics education. Psychonomic bulletin & review14, 249–255 (2007) 18 Malloy et al

    2007

  27. [35]

    Artificial Intelligence Review57(12), 1–23 (2024)

    Schmitt, M., Flechais, I.: Digital deception: Generative artificial intelligence in social engineering and phishing. Artificial Intelligence Review57(12), 1–23 (2024)

    2024

  28. [36]

    Neural com- putation 35(3), 309–342 (2023)

    Sejnowski, T.J.: Large language models and the reverse turing test. Neural com- putation 35(3), 309–342 (2023)

    2023

  29. [37]

    In: 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW)

    Sharma, M., Singh, K., Aggarwal, P., Dutt, V.: How well does gpt phish people? an investigation involving cognitive biases and feedback. In: 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW). pp. 451–457. IEEE (2023)

    2023

  30. [38]

    Computers & Security127, 103105 (2023)

    Singh, K., Aggarwal, P., Rajivan, P., Gonzalez, C.: Cognitive elements of learning and discriminability in anti-phishing training. Computers & Security127, 103105 (2023)

    2023

  31. [39]

    MIT Press (2022)

    Vishwanath, A.: The weakest link: How to diagnose, detect, and defend users from phishing. MIT Press (2022)

    2022

  32. [40]

    Journal of Educational Computing Research59(6), 1169–1183 (2021)

    Weaver, B.W., Braly, A.M., Lane, D.M.: Training users to identify phishing emails. Journal of Educational Computing Research59(6), 1169–1183 (2021)

    2021

  33. [41]

    hack: engaging anti-phishing training through a role-playing phishing simulation game

    Wen, Z.A., Lin, Z., Chen, R., Andersen, E.: What. hack: engaging anti-phishing training through a role-playing phishing simulation game. In: Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. pp. 1–12 (2019)

    2019

  34. [42]

    arXiv preprint arXiv:2302.11382 (2023)

    White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert, H., Elnashar, A., Spencer-Smith, J., Schmidt, D.C.: A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382 (2023)

    2023 arXiv

  35. [43]

    Morgan Kaufmann (2010)

    Woolf, B.P.: Building intelligent interactive tutors: Student-centered strategies for revolutionizing e-learning. Morgan Kaufmann (2010)

    2010

  36. [44]

    arXiv preprint arXiv:2307.15043 (2023)

    Zou, A., Wang, Z., Kolter, J.Z., Fredrikson, M.: Universal and transferable adver- sarial attacks on aligned language models. arXiv preprint arXiv:2307.15043 (2023)

    2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.