We argue that the OpenAI GPT Store, a marketplace containing thousands of customized generative AI models known as GPTs, can exacerbate problems surrounding users' ability to recognize the expertise (or lack thereof) of generative AI models. We contend that OpenAI models are anthropomorphized and presented as authority figures, and that many community developers of GPTs adopt or imitate the model established by OpenAI. We define Expertise Fog, a deceptive design pattern wherein the interface presents a facade of expertise despite the lack of evidence that such models possess expertise, or indeed possess any appreciable differences from a base model with respect to expertise. Finally, we propose that the deceptive design research community proactively address deceptive patterns in GPTs to prevent real-world harm. We offer four mechanisms for doing so: transparently disclosing GPT components, robustly evaluating expert GPTs, clearly labeling GPTs as tools, and foregrounding shortcomings of expert GPTs.
CEUR ceur-ws.org
Generative AI technologies such as ChatGPT [
One might reasonably ask why users of ChatGPT continue to rely on a generalist conversational
agent to automate expert work. Scholars have contended that OpenAI’s presentation of ChatGPT
as a milestone on the path to Artificial General Intelligence [
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Concerning though they are, the instances of inappropriate reliance on generative AI discussed
above largely precede the release of the GPT Store, a marketplace for customized versions
of ChatGPT released in January 2024 by OpenAI, which made available tens of thousands
of “GPTs” to purchasers of ChatGPT Plus subscriptions [
Customized, shareable generative AI technologies present the potential for novel deceptive designs, and additional deceptive patterns may arise as GPTs become more integrated into user-facing technologies. Building capabilities to describe and address deceptive patterns as they emerge can position the community to mitigate their associated harms.
OpenAI introduced custom versions of ChatGPT called GPTs during its Developer Day on
November 6, 2023 [
These advances produced configurable options that enabled the creation of the customized
versions of ChatGPT now called GPTs. We outline the four core components of a GPT:
– Instructions: Builder-defined directions to govern model behavior and interactions with an end
user, typically including what the model should and should not say to users.
– Knowledge: Text files referred to by a GPT during conversation with a user. According to the
most recent documentation, builders may equip a GPT with up to twenty files [
A GPT must have Instructions, while other components are optional. Other features communicate GPT functionality, scaffolding conversations and assisting in discovery on the Store: – Name: A title for the GPT. – Description: 300 words or fewer describing the intended functionality of the GPT. – Image: An image to represent the GPT, uploaded by the builder or made with DALL-E. – Conversation Starters: Prompts suggested before a conversation with the GPT begins, which appear as buttons above the text input box in the GPT’s conversational user interface.
Builders can use a menu-based interface to define both the core components of the GPT and the user-facing features intended to describe the GPT’s functionality. However, the default interface for building a GPT is another GPT called the GPT Builder, which creates a GPT based on its conversation with a (human) builder. Toggling between the two tabs (“Create” and “Configure”) of the GPT Creation interface reveals that the GPT Builder automatically enters text into the Instructions, Description, and Conversation Starters fields based on the builder’s description. The GPT Builder often suggests a Name, and may automatically generate an image for a GPT using DALL-E after it completes the Instructions, Description, and Name fields.
The GPT Store provides a means for users to discover and search for GPTs created both by Community builders and by OpenAI itself. As of this writing, the GPT Store is entirely contained within a single web page (https://chat.openai.com/gpts), from which every available GPT can be navigated to. At the top of the page is a search bar with the default text “Search public GPTs.” Typing into this search bar displays the top ten results of a keyword search in a dropdown descending from the search bar. Each result includes the Name, Image, Description, and Author of the GPT, along with its approximate number of conversations with users (e.g., “1K+”).
Below the search bar is an OpenAI-curated set of four “Featured” Community GPTs, which are
updated weekly. The Name, Image, Description, and Author of these GPTs are displayed such that
they appear larger than any other GPTs included on the page. Below the blocks of Featured GPTs
is a block of “Trending” GPTs, which includes the six (expandable to twelve) most popular
Community GPTs on the GPT Store. The Trending GPTs are followed by a block of six (expandable
to seventeen) GPTs developed by OpenAI itself. Below the OpenAI GPTs are seven additional
blocks of six (expandable to twelve) Community GPTs each, organized according to the following
headings: DALL-E, Writing, Productivity, Research & Analysis, Programming, Education, and
Lifestyle. How models are selected for inclusion in these blocks of GPTs is opaque: OpenAI’s
documentation states only, “we actively review engagement with GPTs and surface trending GPTs
for different categories” [
In order to access the GPTs available via the GPT Store, users must subscribe to ChatGPT Plus, which costs $20.00 USD per month as of this writing, and also provides users with faster model response times and access to OpenAI’s flagship GPT-4 model via the ChatGPT user interface [ 22]. Similarly, builders must subscribe to ChatGPT Plus to create GPTs [22]. While non-subscribing users can see GPTs available in the GPT Store, they cannot interact with them [22].
We turn to the question of how a GPT’s expertise is communicated to users on the GPT Store. We first study the block of seventeen OpenAI models included on the GPT Store home page, and then we examine the Community GPTs accessible on the home page and via the search bar.
OpenAI’s presentation of its own GPTs provides insight into its views on how to share GPT functionality, and implicitly sets expectations for how builders can position their GPTs on the GPT Store. We studied the seventeen OpenAI GPTs on the GPT Store and found these commonalities: – The GPT’s Name connotes expertise. Names of OpenAI GPTs include Math Mentor, Tech Support Advisor, and Creative Writing Coach. – The GPT’s Description is in the first person. Many OpenAI GPT descriptions are written as though by the GPT itself, suggesting interaction with a capable, intelligent being. – The author of the GPT is simply “ChatGPT.” There is no information about who created the GPT, who developed its Capabilities or Actions, or who authored its Knowledge.
GPTs are presented as anthropomorphized experts, with little reference to the human data and tools supporting the model. A model’s About page provides scarcely more information than is available via the home page. GPTs like Tech Support Advisor note that they can perform data analysis, or that they are capable of “Browsing,” but do not disclose what data they are browsing, nor what their capabilities mean in the context in which the GPT will be employed. Nor is there an explanation of why a user should rely on the GPT to perform this task, or what differences the user might see were they to use the GPT rather than simply using a base model like GPT-4. Attempting to follow the links next to the Author field to understand more about the model’s authors simply leads to the OpenAI homepage, which provides no information about the authors of a specific GPT. Moreover, our conversational interactions with these GPTs bear out what prior research has found, that they answer questions authoritatively, even if they may not have the correct answer to a question. Warnings about the potential unreliability of the GPT’s output typically arise only at the very end of an answer, where a user might miss them or understandably ignore them, given the otherwise authoritative tone of the GPT. The presentation of these GPTs suggests that users are encouraged to trust the models as experts, even in the absence of traditional indicators of expertise. While not every OpenAI GPT intends to provide expertise, we will show that the presentation of GPTs as experts occurs not only in OpenAI GPTs but in Community GPTs as well, many of which offer expertise in consequential domains.
A far cry from the seventeen GPTs offered by OpenAI, the thousands of Community GPTs available via the GPT Store defy comprehensive treatment in a short position paper. We will thus make our argument by addressing the presentation of expertise among GPTs intended for three consequential professional domains: law, medicine, and finance. We undertake a simple experiment by searching on “legal,” ”medical,” or ”finance” in the GPT Store search bar, and report the Name and Description of the top ten GPTs returned in the results for each search. We note that we report the results of this experiment based on a search on February 25, 2024.
Table 1 includes the top ten results for each keyword. Though these GPTs mostly do not address the user in the first-person, they nonetheless leave little doubt about their purported expertise. The top result for “legal” returns Legal+, described as “your personal AI lawyer,” and promising “real time legal advice.” The second result for “medical” returns Medical GPT, described as a “friendly virtual doctor” providing “broad medical advice.” The fifth result for “financial” returns Financial Planner, promising “personalized financial advice.” In all three cases, a GPT has adopted the title of a human professional (lawyer, doctor, financial planner) and promised expert professional advice for which human experts prepare for many years to provide. These GPTs are far from outliers: top ranked legal GPTs employ Names positioning the models as an Assistant, an Advisor, a Scribe, and a Pro; top medical GPTs use names such as Genius and Mentor; and top financial GPTs use names like Wizard, Guru, Analyst, Companion, Expert, Advisor, and Navigator.
Users searching for expertise on the GPT Store will readily find it, or at least the appearance of it. According to statistics from the GPT Store, users have logged more than 10,000 conversations with the Legal+ GPT. Moreover, we found that some expert GPTs were among the GPTs on the home page. The GPT Store featured the Financial Wizard GPT, for example, in the top six GPTs for Research & Analysis on February 26, 2024. Having established that GPTs in widespread use are presented as experts, we next theorize their presentation as a deceptive design pattern.
Prior work establishes that deceptive design patterns manipulate the behavior of users in order
to obtain money, data, or attention from the user [23, 24]. Such patterns are pervasive among
user-facing apps on platforms such as the Apple and Android stores [25, 26], and they can
influence not only a user’s online behavior but their well-being [ 27] and sense of safety in the
real world [28]. Given the relative recency of GPTs and the GPT Store in particular, little prior
work explicitly treats the presence of deceptive design patterns in these technologies, save that of
Zhan et al. [
We contend that the user-facing presentation of expertise on the GPT Store constitutes a deceptive design pattern, which we refer to as Expertise Fog. We specifically define Expertise
GPT Name Legal+ Legal Assistant Legal Advisor Legal Eagle Legal Scribe Legal Bot Legal Legal eagle Legal Pro
Legal scenario simulator for professionals, students, and enthusiasts Assists in drafting basic legal documents Legal bot: friendly, step-by-step legal advisor We provide you with intelligent text generation capabilities to help you create highquality text content in various applications.
I analyze legal texts and explain laws.
Examines any legal document to identify pitfalls. #Legal
Medical Genius Medical Notes Medical Advice Medical Mentor Medical Illustration Master Medical Translate
Medical AI
Community GPTs for Three Expert Professional Domains on the GPT Store Legal Keyword Medical Keyword Financial Keyword
GPT Description GPT Name GPT Description GPT Name GPT Description Your personal AI Medical I simplify complex Financial Financial expert and lawyer. Does it all Research medical research, Wizard programmer, clarifyfrom providing real highlighting key ing investments and time legal advice for points and sources. translating financial day-to-day problems, indicators into algoproduce legal con- rithms. tract templates & much more! Legal assistant for Medical Friendly virtual doc- Financial Financial Analyst consulting, drafting GPT tor for broad medical Guru Generalist contracts and legal advice. documents Answers legal ques- AI for Financial tions in a variety of Medical Analyst topics Students Expert in financial markets, strategies, and analysis. ”A senior financial analyst with 20 years of experience, specializing in financial auditing, valuation, and investment analysis.” Financial derivatives tutor-explainer Personalized financial advice to help you achieve your financial goals Tailored ifnancial planning advice for UK users Financial planning assistant for budgeting and investments Expert in financial analysis and advice Strategic advisor for mid-life financial planning.
A savvy ifnancial advisor ofering personalized investment strategies.
Medical Study AI aids in Medical Assistant learning, AI for Medical Students, understanding Medical Terminology, navigating Medical School, and Molecular Biology concepts… Oral Medical Pathology Diagnostic Lesion description, diferential diagnosis, complementary tests and treatments.
Write Excellent Medical Notes helpful dialogue simulation roleplay DrBrinkhouse Advanced medical professor delving into complex medical details.
Creates high-quality medical art from keywords.
A medical translation assistant, providing accurate translations of medical texts and conversations. helpful dialogue simulation roleplay DrBrinkhouse
Financial derivatives Financial Planner Financial Planning UK Financial Companion Financial Expert Financial Advisor Financial Navigator 10
Legal Reader
Adaptive Kosovo Legal Guide. Fog as the manipulative redirection of attention to a source of information presented to the user as an expert, without adequately establishing and communicating whether the ostensible expert can truly offer expertise. In theorizing Expertise Fog, we build on prior work of Chaudhary et al. [29], who describe “Feature Fog,” a deceptive pattern that reduces a user’s ability to detect how much time they’ve spent on an activity (such as the lack of a time stamp in a video player). This pattern is also characterized as “Time Fog” by Monge Roffarello et al. [30], who note that it is also an example of the “interface interference” pattern described by Gray et al. [31], and specifically reflects the “hidden information” pattern. Expertise fog is similarly characterized by hidden information; what makes Expertise Fog distinctive is that determining whether the user appropriately relied on an expert GPT requires the judgment of a human expert - the very entity for which the GPT presents a substitute. A non-expert would be unqualified even to judge whether the model can stand in for an expert. Indeed, while we interacted with the GPTs in question during the course of collection data for this paper, we could not assess whether they offer sound legal, medical, or financial advice, because we are not lawyers, doctors, or financial experts. While the outputs of these models appear compelling to lay users, they may nonetheless contain consequential mistakes only apparent to skilled professionals.
The potential for real-world harms to result from the legal, medical, and financial GPTs outlined in this work is relatively self-evident: a user might act on bad legal, medical, or financial advice from a GPT presented to the user as an expert. Even without the Expertise Fog induced by the GPT Store, users (including lawyers themselves, as noted in the introduction) have already acted on bad output from the ChatGPT base model, resulting in professional consequences.
We next consider why Expertise Fog occurs on the GPT Store. On the one hand exist straightforward financial incentives: OpenAI sells a $20.00 monthly subscription to ChatGPT Plus, and access to many GPTs rather than a single ChatGPT base model may motivate purchases. Moreover, though builders cannot yet monetize their GPTs within the Store, they can direct users to purchase their products or interact with their data via GPT Actions. However, Expertise Fog on the GPT Store may also arise from more complex motivations than financial considerations tied to the use of individual models. Consider that, unless a builder adds proprietary data, or defines a programmatic action, a GPT is nothing more than a text string describing how the GPT-4 base model should behave - which is by default written by a custom version of GPT-4 known as the GPT Builder. GPT-4 can itself employ all of the Capabilities with which a GPT can be equipped. Were a GPT’s instructions revealed, it could be trivially replicated by an end user. One can imagine a very different design for the GPT Store, which provided the most appropriate instructions to an end user, similar to open repositories of prompts for image generators like Stable Diffusion and Midjourney. However, transparently providing instructions rather than presenting a GPT would likely command less of a user’s time and attention than an interaction with a humanlike expert. For example, Instructions for a GPT like Legal+ could be as straightforward as “You are a lawyer”; however, such a prompt would likely generate less user activity than the Legal+ GPT.
Given the minor differences between many GPTs and the GPT-4 base model, we observe that the inability to review GPTs may reflect OpenAI’s concerns that user feedback would implicate not only a builder-created GPT, but OpenAI’s base model. From February 22, 2024, OpenAI has allowed users to rate GPTs on a vfie-star scale. However, users cannot leave reviews, wherein criticisms might reflect shortcomings both in individual GPTs and in GPT-4 itself. This differs from app stores like those maintained by Apple and Google. Where apps published to those stores contain flaws, they reflect a shortcoming of the app’s developer, rather than the store. Moreover, maintainers were motivated to ensure that reviews highlighted untrustworthy apps to maintain the quality of the store. However, identifying flaws in GPTs could also harm the brand of GPT-4.
Finally, OpenAI’s motivations may differ from builders. We encountered a deceptive aspect of the GPT Builder interface: when uploading files as Knowledge, or defining a new Action, the interface inserts a new “Additional Settings” dropdown at the bottom of the page, usually out of sight. Expanding this dropdown reveals a previously absent checkbox labeled “Use conversation data in your GPT to improve our models” - checked by default. OpenAI’s motivations may thus have more to do with collecting data for training models, whether builders are aware of it or not.
We propose a preliminary roadmap for addressing deceptive design patterns in GPTs by 1) transparently describing GPT components; 2) establishing expertise through systematic evaluation; 3) clearly labeling GPTs as tools; and 4) foregrounding shortcomings as a human expert would.
Transparent presentation of a GPT’s components could better communicate its functionality. For example, the GPT creation interface might permit builders to describe the Knowledge accessible by a GPT, and while making a GPT’s Instructions more transparent may conflict with intentional vagueness that renders these text strings valuable, the GPT Store could allow builders to indirectly describe instructions, and set standards for communicating a GPT’s intended behaviors. If a legal GPT is instructed to review drafts, but not to produce documents, describing these distinctions might mitigate improper use. Clearly communicating Actions can also help establish their value. For example, the Kayak GPT presents hotel options by retrieving them from the Kayak website. However, rather than describing what Action is occurring, the interface simply notes that the GPT is “talking to kayak.com” next to a progress bar, providing little information while anthropomorphizing model outputs. Simply stating that the model is querying Kayak’s database of hotel listings would provide clarity. Finally, GPT builders also deserve transparency in the use of the data with which they equip their GPTs. The dropdown obscuring a default enabling OpenAI to use conversation data from knowledge-equipped GPTs inhibits the agency of builders.
Generative AI shifted away from the pretraining and task-specific fine-tuning paradigm of early transformer models [32, 33]. Models like ChatGPT and DALL-E respond to user inputs in a “zeroshot” setting - with no examples of a task provided by the user, and no updates to the weights [34]. Such models improved the usability of AI and broadened its appeal. However, increased usability has come at the expense of evaluation. Most chat-based models are assessed on world knowledge [35] and the degree to which their responses are preferred by human users [36]. This is useful for assessing model usability and general knowledge, but it provides no measure of how well-suited specialist GPTs like those on the GPT Store are for the task for which they are specialized. Legal GPTs are not assessed on legal tasks, nor medical GPTs on medical tasks, nor financial GPTs on financial tasks. Adopting rigorous standards for evaluating GPTs, especially those positioned as experts, can mitigate deceptive design. Such efforts might build on prior work to adequately document machine learning models using Model Cards [37], or datasets using Datasheets [38]. Domain-specific GPTs could undergo a battery of evaluations, with the results presented to potential users. Popular models might be red-teamed [39] for problematic in-domain content, and potentially certified by human experts. Such evaluations would not relieve the responsibility to note that GPTs can make mistakes, but it would provide more specific, contextual knowledge to users.
Many ostensibly expert GPTs differ from the GPT-4 base model only in the Instructions provided by the builder. The learned weights of the underlying model do not differ, and the GPT may not have access to proprietary data or programmatic actions. We argue that such GPTs should not be positioned as experts, in that they have access to precisely the same knowledge available to the base model. If a user should not use ChatGPT to write legal documents for submission to a court, the user also should not use ChatGPT adopting the persona of a “personal lawyer” to write such documents. Where users have made the mistake of relying on ChatGPT for such tasks due to claims about the model’s general intelligence, the advertising of GPTs like Legal+ now means that users could make the same mistake - about the same underlying model - due to unwarranted claims about the model’s specific intelligence. Anthropomorphized names and descriptions exacerbate the issue by positioning models as doctors, lawyers, and financial planners. Addressing this may serve as a site for regulation: where a GPT is functionally the same as a base model but for its Instructions, its name and description should reflect that it is a tool, not a standin for a human.
Human experts often defer to the authority of other experts. For example, a general practitioner would refer a patient experiencing heart issues to a cardiologist. Human experts with general skills are nonetheless unqualified to address every specific situation, and trustworthy experts highlight this inadequacy, prioritizing the well-being of clients. Similarly, generalist AI will be unqualified in many cases - and this should be highlighted. Conversations with ostensibly expert GPTs like Legal+ show that the GPT attempts to answer whatever question is put to it, and then adds a brief disclaimer. However, this assumes the user will read to the end, and disregard a potentially false answer in light of the disclaimer. We propose instead that, unless a GPT is directly quoting from a qualified human expert, the GPT can highlight that 1) the question is best addressed by a human with specific expertise; and 2) the model is unreliable for use as an expert.
In this position paper, we argued that the GPT Store facilitates deceptive design related to the presentation of expertise, resulting in a deceptive design pattern we call Expertise Fog. We suggested four ways to address deceptive design in GPTs via transparency of GPT components, robust evaluation of GPTs, labeling GPTs as tools, and foregrounding the shortcomings of GPTs. We intend to continue our analysis a full study of deceptive presentations on the GPT Store and deceptive interactions with GPTs. normative considerations, and measurement methods, in: Proceedings of the 2021 CHI conference on human factors in computing systems, 2021, pp. 1–18. [25] L. Di Geronimo, L. Braz, E. Fregnan, F. Palomba, A. Bacchelli, Ui dark patterns and where to find them: a study on mobile applications and user perception, in: Proceedings of the 2020 CHI conference on human factors in computing systems, 2020, pp. 1–14. [26] J. Gunawan, A. Pradeep, D. Choffnes, W. Hartzog, C. Wilson, A comparative study of dark patterns across web and mobile modalities, Proceedings of the ACM on Human-Computer Interaction 5 (2021) 1–29. [27] C. M. Gray, J. Chen, S. S. Chivukula, L. Qu, End user accounts of dark patterns as felt manipulation, Proceedings of the ACM on Human-Computer Interaction 5 (2021) 1–25. [28] I. Chordia, L.-P. Tran, T. J. Tayebi, E. Parrish, S. Erete, J. Yip, A. Hiniker, Deceptive design patterns in safety technologies: A case study of the citizen app, in: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 2023, pp. 1–18. [29] A. Chaudhary, J. Saroha, K. Monteiro, A. G. Forbes, A. Parnami, are you still watching?: Exploring unintended user behaviors and dark patterns on video streaming platforms, in: Designing Interactive Systems Conference, 2022, pp. 776–791. [30] A. Monge Roffarello, K. Lukoff, L. De Russis, Defining and identifying attention capture deceptive designs in digital interfaces, in: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, 2023, pp. 1–19. [31] C. M. Gray, Y. Kou, B. Battles, J. Hoggatt, A. L. Toombs, The dark (patterns) side of ux design, in: Proceedings of the 2018 CHI conference on human factors in computing systems, 2018, pp. 1–14. [32] J. D. M.-W. C. Kenton, L. K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of NAACL-HLT, volume 1, 2019, p. 2. [33] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019). [34] W. Wang, V. W. Zheng, H. Yu, C. Miao, A survey of zero-shot learning: Settings, methods, and applications, ACM Transactions on Intelligent Systems and Technology (TIST) 10 (2019) 1–37. [35] D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, J. Steinhardt, Measuring massive multitask language understanding, in: International Conference on Learning Representations, 2020. [36] L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al., Judging llm-as-a-judge with mt-bench and chatbot arena, Advances in Neural Information Processing Systems 36 (2024). [37] M. Mitchell, S. Wu, A. Zaldivar, P. Barnes, L. Vasserman, B. Hutchinson, E. Spitzer, I. D.
Raji, T. Gebru, Model cards for model reporting, in: Proceedings of the conference on fairness, accountability, and transparency, 2019, pp. 220–229. [38] T. Gebru, J. Morgenstern, B. Vecchione, J. W. Vaughan, H. Wallach, H. D. Iii, K. Crawford,
Datasheets for datasets, Communications of the ACM 64 (2021) 86–92. [39] D. Ganguli, L. Lovitt, J. Kernion, A. Askell, Y. Bai, S. Kadavath, B. Mann, E. Perez, N. Schiefer, K. Ndousse, et al., Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned, arXiv preprint arXiv:2209.07858 (2022).