The literature review is organised around ten distinct information quality criteria. For each criterion, the analysis addresses three specific dimensions: the particular AI failure mode that threatens it; the empirical and theoretical literature documenting that failure; and the degree to which the existing research treats these failures as ethical problems rather than purely technical or empirical challenges.

In identifying these failure modes, the review builds upon the six characteristics of generative AI established in Wilson (2026): opacity, epistemic authority, sycophancy risk, personalisation, first-person voice construction and commercial optimisation.

The criteria have variable coverage in the literature: for example, accuracy and objectivity are well covered, while currency is less well covered. The relative scarcity of literature in some areas is itself a significant finding, the implications of which are explored further in the Discussion section.

When ChatGPT-3.5 was launched in November 2022, its designers were already aware that it could produce inaccurate information through hallucination. The ChatGPT blog (OpenAI, 2022) noted:

By the release of ChatGPT-4 in March 2023, performance had improved modestly. The GPT-4 Technical Report (OpenAI, 2023) recorded a 19% improvement on the TruthfulQA benchmark, with GPT-4 scoring 60%, compared to 41% for GPT-3.5. However, this still implied a 40% failure rate on a test designed to measure truthfulness. The report itself cautioned that GPT-4 ‘is not fully reliable (it “hallucinates” facts and makes reasoning errors). Great care should be taken when using language model outputs …’ (p. 10). More importantly, GPT-4’s greater fluency and confidence may make errors harder rather than easier for users to detect, directly implicating the problem of epistemic authority identified in Wilson (2026).

Early academic scrutiny confirmed domain-wide limitations: Kung et al. (2023) found ChatGPT approaching but not reliably reaching clinical examination pass thresholds, while Plevris et al. (2023) showed that responses to complex tasks were frequently unreliable and inconsistent across repeated prompts. Subsequent research has consistently reinforced two findings: surface fluency does not guarantee factual reliability, and accuracy cannot be assumed to be stable across interactions.

Accuracy failures are particularly acute in relation to citations and references. Oladokun et al. (2025) examined ChatGPT’s responses to librarianship questions, finding that 42.9% of citations generated by ChatGPT-3.5 were false or non-existent, rising to 51.8% in ChatGPT-4o. Verified citation accuracy was correspondingly low: 3.92% and 6.35% respectively. A similar pattern appears in clinical contexts: Blasingame et al. (2025) reported that although 83.3% of GPT-4 responses to clinical evidence questions were judged correct overall, only 63% of the references could be verified as non-fabricated. These findings highlight a crucial distinction for information professionals: a response may appear substantively accurate while its attribution is fundamentally unreliable.

The most systematic comparison of large language model (LLM) performance against human expert standards is provided by Wei and Chen (2026). Evaluating both Dense and Mixture-of-Experts architectures across six information quality dimensions, they found that both model types lagged behind human experts on intrinsic and contextual information quality—dimensions directly concerned with accuracy and fitness for purpose.

The ethical dimensions of accuracy failures were articulated even before the release of ChatGPT-3.5. Bender et al. (2021) argued that LLMs do not understand language but generate outputs through statistical pattern matching, without grounded meaning, making hallucinations a structural feature rather than a fixable flaw. Following ChatGPT’s release, Lund et al. (2023) reframed AI-generated content as an ethical issue for scholarly communication, raising concerns about authorship, attribution, and research integrity when inaccurate AI outputs are presented as authoritative.

Sima et al. (2026) extend this argument by distinguishing responsible AI use, characterised by transparency, validation, and human oversight, from careless or deceptive use that risks introducing inaccurate or unverifiable content into the scientific record. They conclude that careless AI use undermines scientific credibility and trust. Freyer et al. (2024), writing from the perspective of text simplification for marginalised groups, add a further ethical dimension: inaccuracies introduced during AI-driven simplification may be invisible to people with cognitive disabilities or limited literacy, precisely those least able to detect them. This links the accuracy criterion directly to accessibility and reinforces the information professional’s duty to protect vulnerable users.

Given the scale of misinformation, disinformation, and conspiracy theories on the Web, the information professional’s role in ensuring that accurate and unbiased information is available to users has become increasingly significant. However, AI systems’ ability to deliver objective information is compromised by biases embedded in their training materials. Foka and Griffin (2024), for example, observe that cultural history collections are inherently biased through processes of selection and description that carry historical prejudices into the present. When digital versions of these collections are used in training generative AI systems, those biases are propagated.

Empirical studies provide convergent evidence of such effects. Lee et al. (2024) identified racial bias in AI-generated ophthalmological educational materials: demographic information, including ethnicity, affected output length, with information for Black patients containing significantly fewer words than that for White patients, although reduced informational content was not definitively established. Fleisig et al. (2024) documented linguistic bias when prompts were written in non-standard English dialects, reporting that ‘responses to non-“standard” varieties consistently exhibit a range of issues: stereotyping (19% worse than for “standard” varieties), demeaning content (25% worse), lack of comprehension (9% worse), and condescending responses (15% worse)’ (p. 13541). A UNESCO study (van Niekerk et al., 2024) similarly revealed gender and racial bias across several LLMs: women were more often associated with terms such as home, family and children, while men were linked to business, career and executive. When ethnicity was introduced, Zulu men were more likely to be associated with security guard or gardener, and Zulu women with domestic servant, cook and housekeeper. Not all systems exhibited bias uniformly, however: ChatGPT, for example, generated positive or neutral content in over 80% of cases relating to LGBTQ+ subjects.

There is, therefore, strong empirical evidence for systematic bias in generative AI systems, largely traceable to their training data. The ethical implications for information professionals have also been explored. Slater (2025) argues that librarianship has long involved a ‘precarious balancing act’ between budgets, emerging technologies, and obligations to patrons. She suggests that the use of generative AI represents ‘a shift in the responsibility of facilitating ethical labor practices in LIS, operating at some degree for the public good, onto privatized technological solutions that are constantly changing and fetishized’ (p. 590). She further contends that the popularity of generative AI ‘does not correlate to ethical standards of information services created and supported by LIS practitioners’ (p. 597), concluding that librarians should refuse its use because it constitutes a ‘destructive force’ that is infrastructurally racist, obscures reliance on scraped digital labour, fails to produce information with integrity without human scrutiny, and contributes unnecessarily to environmental harm (p. 601).

Cox and Jimenez (2025) advance a related but distinct claim, arguing that ChatGPT embeds Western epistemological assumptions that amount to a systematic failure of objectivity. They propose design principles for a decolonial generative AI in response. That such concerns are now widely recognised within the profession is confirmed by Mannheimer et al. (2024) and Ibrahim (2025), both of whom identify bias against diverse user communities as a central ethical challenge for libraries adopting AI. While bias is clearly recognised as a professional ethical issue, the academic literature on appropriate professional responses remains under-developed; frameworks, guidelines, and policy statements are more commonly found on institutional websites than in peer-reviewed research.

Bias also extends to AI-generated representations of the profession itself. Spennemann and Oddone (in press) found that ChatGPT-generated images of librarians predominantly depicted Caucasian figures and significantly overstated the presence of men, reinforcing stereotypical and unrepresentative images of professional identity.

Generative AI systems present information in an authoritative tone of voice, but users have little insight into the sources on which that information is based and cannot readily judge whether it is genuinely authoritative. While accuracy has been addressed above, traceability is a distinct quality issue. The inability to identify sources (although partially addressed by newer system features) is compounded by first-person voice construction and, in the case of Claude.ai, by the use of a human name, both of which encourage users to treat the system as a human-like informant.

It appears that generative AI systems themselves cannot identify the sources of their training data. When asked about its training sources, ChatGPT acknowledged that its training involved a mixture of licensed data, data created by human trainers, and publicly available material, but confirmed that it has no access to its training data and cannot identify specific sources, drawing instead on internalised knowledge to compose responses (OpenAI, 2026). A similar query put to Claude.ai produced an equivalent response.

Theoretical analysis and empirical evidence for this problem of opacity are well established. Sundin (2025) notes that ‘while search engines traditionally provide links to websites and other documents, generative AI chatbots generate content without a stable source other than the LLM itself’ (p. 295), observing that hybrid systems combining LLMs with search engines are ‘shifting the focus towards searching for content rather than identifying sources’ (p. 298). As a result, users increasingly receive content divorced from explicit provenance, even though the data underpinning LLM outputs remains opaque to them. Sundin concludes that this represents a transfer of analytical skills (and agency) from the user to the system, leaving users increasingly dependent not only for retrieval but also for evaluation and interpretation (p. 298).

Heersmink et al. (2024) provide a useful conceptual framework, distinguishing reflective, phenomenological, and informational transparency. Reflective transparency (understanding how the system works and what data it relies on) is largely absent in generative AI systems. Phenomenological transparency arises as interfaces become familiar and effortless for frequent users, while informational transparency is achieved because outputs are generally easy to understand. Together, phenomenological and informational transparency foster trust and encourage anthropomorphism, while the lack of reflective transparency undermines the epistemic basis for that trust. The authors argue that improving trustworthiness will require greater reflective transparency, including clearer knowledge of training data and algorithmic processes.

A further consequence of opacity is the emergence of responsibility gaps. Taylor (2025) asks who is accountable when decisions are taken on the basis of AI-supplied information, or when decisions are made directly by the system. Although Taylor addresses AI systems broadly, the question applies equally to generative AI: when a human acts on information provided by an opaque system and harmful consequences follow, responsibility remains contested.

From an information literacy perspective, Tomaszewski (2025) proposes CAT (Check, Ask, Think) as an evaluative framework for AI interactions, encouraging users to assess ‘source credibility (Check), intent and authorship, whether human or algorithmic (Ask), and platform or emotional context (Think)’ (pp. 7–8). While such frameworks offer practical mitigation, the ethical implications of opacity remain under-examined. For information professionals committed to delivering credible information from identifiable and validated sources, the persistent lack of transparency in generative AI systems presents an ethical dilemma: without reliable traceability, professionals cannot guarantee the authority, accuracy, or validity of the information delivered to users.

The accessibility criterion intersects with several others. As noted in the Objectivity section, Cox and Jimenez (2025) argue that the Western, particularly American, bias of training materials limits access to cultural knowledge not represented in LLM training resources, disadvantaging users from other cultural contexts. Slater (2025) similarly contends that these systems reinforce existing inequalities of access by prioritising commercial over public interests. Linguistic and cultural barriers are further documented by Fleisig et al. (2024) and van Niekerk et al. (2024).

Matsieli and Mutula (2025) emphasise that the Global South is particularly vulnerable to these forms of exclusion, given the systematic omission of local cultural resources from training data. Freyer et al. (2024) highlight another dimension of inaccessibility, noting that generative AI systems do little to support users with cognitive disabilities. Economic barriers compound these problems: like many information technologies, generative AI operates on a freemium model, with more capable or useful features restricted to paying users. As Mazel (2025) observes in an examination of AI tools for literature searching, this creates discrimination on economic grounds. She notes that ‘to make full use of these tools, users must already have a high level of information privilege due to their reliance on users supplying or being able to access paywalled content’ (p. 323).

Simons (2025) situates these concerns within a broader ethical critique. He draws on indigenous and non-Western concepts of wisdom, including the Diné concept of hozho, which holds that individual wellbeing is inseparable from one’s relationships with community, nature and the universe and the African philosophy of Ubuntu, premised on the fundamental interconnectedness of persons, to argue that the profession’s response to generative AI must be grounded in relational and intercultural values rather than the dominant Western technocratic framework. This requires not merely avoiding harm but actively embodying the values of inclusion and equity in professional practice. Ford (2026), drawing on case studies of critical AI literacy interventions in public and academic libraries in Greater Sydney, complements this perspective by reframing AI literacy as not only understanding how AI works, but recognising when, how, and why it fails. This positions the librarian as an active mediator rather than a passive conduit for AI-generated content.

Taken together, these arguments point towards a professional response requiring both ethical commitment and practical skill. In Kantian terms, the information professional’s duty is to ensure that information is accessible to all users regardless of economic circumstance, language, cognitive capacity, or cultural background, a duty that cannot be discharged simply by deploying AI systems whose design systematically disadvantages some users. Of all ten criteria, accessibility has attracted the most explicit ethical attention, indicating a professional recognition that access to information is not merely a practical concern but a matter of justice.

The obligation to provide information that is relevant to the user’s actual need, and genuinely fit for the purpose for which it is sought, is complicated by one of generative AI’s most distinctive characteristics: sycophancy risk. Malmqvist (2025), in a technical survey of sycophancy in large language models, defines the phenomenon as, ‘the propensity of models to excessively agree with or flatter users, often at the expense of factual accuracy or ethical considerations’ (p. 61), a behaviour unintentionally reinforced during reinforcement learning from human feedback. The result is a system structurally biased towards telling users what they appear to want to hear rather than what they actually need to know, a direct failure of fitness for purpose even when surface relevance appears high.

The empirical evidence confirms that this failure mode is real and consequential. Schulze Balhorn et al. (2024) assessed ChatGPT’s responses to 594 questions generated by faculty members across five scientific disciplines at Delft University of Technology. They found that while responses were rated on average as ‘mostly correct’, ratings declined significantly as the educational level of the question increased, and note that, ‘A striking difference between what faculty members expect from student answers and answers from ChatGPT is a critical reflection and a clear understanding of the concepts used’ (p. 8).

Goodman et al. (2023), in a study of ChatGPT’s responses to 284 medical questions generated by thirty-three physicians across seventeen specialties, found a similar pattern: performance was strong on simple, straightforward questions but degraded in accuracy and completeness as questions became more complex, context-dependent and clinically consequential. Mun and Hwang (2025) confirm from the user side that relevance is among the information quality attributes most strongly associated with perceived usefulness and trust in ChatGPT, suggesting that users are sensitive to relevance failures even when they cannot always articulate them.

No paper in the retained set explicitly frames fitness for purpose failure as a professional ethical issue. Schulze Balhorn et al. (2024) note that their results indicate a lack of critical attitude in AI responses, and observe that the model does not generally reason on the ethical implications of its answers, relying instead on a small number of content filters. This is a significant observation but it stops short of an ethical argument about professional duty. That sycophancy constitutes a specific violation of the Kantian duty of respect for persons, treating the user as a rational agent with a genuine information need, has not yet been argued in the literature.

Currency, the obligation to provide information that is up to date, presents a particular challenge for generative AI systems: their responses reflect the state of knowledge at the time of training not at the time of query, and no mechanism exists within the systems themselves to signal this limitation to users. Hersh (2024) notes that users of information retrieval systems have legitimate concerns about the timeliness of search results, and argues that LLMs cannot meet those concerns in the way that traditional search systems can, since traditional search returns current documents while the LLMs’ content is frozen at the training cut-off. McGrath et al. (2024), evaluating ChatGPT’s responses to clinical genetics questions, identify outdated information as among the system’s notable remaining challenges even after improvement from version 3.5 to version 4, a finding with direct patient safety implications in a domain where knowledge evolves rapidly. Li and Goyal (2025) add a further structural dimension: even when LLMs are updated with new factual information, the best models achieve less than 2% accuracy on tasks requiring reasoning about that new knowledge, suggesting that the currency problem is architecturally deep rather than a surface limitation that post-training updates can readily address.

No paper in the retained set identifies currency failure as an ethical issue. The duty to provide current information, and to signal when currency cannot be guaranteed, is a foundational professional obligation whose ethical dimensions remain unaddressed in the literature.

The criterion of sufficiency requires that information be not merely accurate but adequate to the user’s purpose, omitting nothing essential. Sufficiency is therefore inherently relative: a body of information adequate for one task or one user may be deficient for another, and the criterion cannot be assessed independently of the purpose it is to serve. Generative AI systems have no access to that purpose, and lacking a representation of the user’s task, they cannot determine what would constitute a complete answer for the particular case. These systems are trained to produce fluent and well-structured text: this fluency may mask omissions, and an answer can read as comprehensive while lacking essential information.

Empirical studies support this concern. Peng et al. (2025) show that AI performs reasonably well on tasks focused on understanding or evaluation, but performs poorly on tasks that require remembering and analysing information. These weaknesses are closely related to problems of sufficiency, especially with respect to accuracy and professional standards. Blasingame et al. (2025), discussed earlier in relation to accuracy, find that 16.2% of GPT-4 responses to clinical evidence questions were only partially correct. This represents a failure of sufficiency rather than accuracy: some of the information provided was correct, but key details required for clinical decision-making were missing. Similarly, Freyer et al. (2024), cited in the accessibility section, note that AI-driven text simplification often strips content down to a minimal core. Instead of making complex information accessible, this can limit what users receive to the bare minimum, excluding important nuance or context. Hosseini et al. (2025) connect sufficiency to open science, arguing that AI-generated research outputs may lack enough methodological detail to allow others to verify or replicate results, thereby undermining scientific reliability.

None of the papers in the retained literature explicitly describe failures of sufficiency as a professional ethical problem. This absence is striking. Ensuring that users receive information that is genuinely sufficient for their purpose, not merely information that appears complete, reflects a core ethical responsibility. It aligns with a Kantian view of respect for persons as rational agents whose decisions depend on the quality and completeness of the information they are given.

The appropriate level of complexity at which information should be presented varies with the recipient: an account prepared for an expert will be inaccessible to a novice, while one for a novice will not satisfy the needs of an expert. This requires output tailored to a model of the individual user’s knowledge level. Generative AI systems do not construct such a model and, lacking a representation of the user’s prior knowledge, they adopt an intermediate level of explanation that serves well neither the expert nor the novice. Moreover, their output is inconsistent: the level of complexity at which a given system responds may vary substantially from user to user.

Most studies on this problem focus on health information, where getting the complexity wrong can be dangerous. Yi et al. (2025) looked at Websites about gender-affirming surgery and found they were written at a 12th-grade reading level, substantially above the 8th-grade level that health experts recommend for patients. When AI tried to simplify these materials, it did improve accessibility but was inconsistent and sometimes showed bias, especially for certain types of procedures. Chen et al. (2026) studied 2000 American adults and found that when AI simplified medical scan reports, people understood them about 11% better and were more accurate in follow-up tasks. However, 60% of people still preferred the original complicated reports, suggesting they did not trust the simplified AI versions even when they understood them better.

Freyer et al. (2024) make an important ethical point: people with learning disabilities or low literacy skills have a right to understand information. This means we have a duty to make information accessible to them, not just offer it as an extra service when convenient.

While researchers have studied this issue more than some other AI problems, they have not fully developed the ethical argument that adjusting information complexity to match users’ needs is a professional responsibility, not just a nice-to-have feature.

Generative AI systems are stochastic by design: the same query submitted multiple times will produce different responses, and there is no guarantee that those responses will be mutually consistent. This variability is a structural property of LLMs rather than a bug to be fixed, and it creates a specific professional problem: the information professional cannot rely on an AI system to give the same answer tomorrow that it gave today, or to give the same answer to different users asking the same question.

Savage et al. (2024) identify a further consistency failure that compounds the variability problem: LLMs consistently overestimate their own confidence, meaning that the system’s expressed confidence cannot be used as a proxy for reliability. Comparing multiple outputs from the same query (sample consistency) is currently the most effective proxy for assessing uncertainty. However, doing so requires deliberate effort from the user or information professional, and the system does not prompt or support this practice. Iqbal et al. (2024), in a study of ChatGPT’s medication recommendations for dermatological conditions, found a high overall dermatologist approval rate of 98.87% but documented specific consistency failures including incorrect Anatomical Therapeutic Chemical codes, repeated errors and incomplete medication codes, concrete failure modes in a clinical context where consistency is critical for patient safety. Xie et al. (2024) add a further dimension: while ChatGPT can identify conflicting evidence in clinical literature with a 90% recall rate, this capacity to detect inconsistency in others’ outputs does not extend to guaranteeing consistency in its own, and the authors caution that its conclusions remain inferences that could be harmful if acted upon in clinical practice. No paper in the retained set frames consistency failure explicitly as a professional ethical issue. The argument that stochastic variability violates the Kantian principle of universalisability, the requirement that the same quality of service be provided to every user, is this paper’s contribution.

Although the concept of information overload pre-dates the computer era (see, e.g., Schroder et al., 1967), research interest grew considerably in the 1980s and later. Within information science, the concept became explicitly articulated in discussions of library use and information services, where authors such as Line (1981) and later Bawden (1986) framed information overload as a condition arising when the volume and diversity of available information exceed users’ cognitive and practical capacity to process it effectively. Later, following a study of UK companies, Allen and Wilson (2003) associated information overload with task overload, rather than simply an excess of information.

The obligation to present information without overwhelming the user’s capacity to process and use it effectively is directly threatened by generative AI systems’ tendency towards verbose responses. Outputs appear thorough and authoritative, since brevity may be interpreted by users as insufficient engagement with their needs. The result is a structural tendency to produce more information than users need, presented in a form that creates the impression of comprehensiveness while potentially impeding effective use.

A striking finding from Constable et al. (2025) illustrates the problem well: users rated ChatGPT responses to practical queries as more trustworthy and informative than the more concise responses from traditional search, yet when tested on the material, they performed better on the Google answers. More detailed responses produced less actual learning, a paradox of information abundance in which apparent comprehensiveness works against genuine understanding. Lee et al. (2025), surveying 319 knowledge workers, find that higher confidence in generative AI is associated with less critical thinking: users shift from active engagement with information to passive acceptance of AI-generated content. Khan and Suhluli (2025), studying 998 researchers, find that heavy use of generative AI increases rather than reduces cognitive strain, the opposite of what users might expect. Wei and Chen (2026) find that LLM outputs fall in the high information load category in over 80% of cases, compared with the low load maintained by human expert texts. No paper in the retained set develops the ethical implications of this failure mode from a professional standpoint. We argue that systems designed to appear comprehensive rather than meeting the user’s actual need violate the professional duty to protect rational agency of the user.

Across the ten criteria, the literature presents a notably asymmetric picture. For accuracy, objectivity, identifiable origin, and accessibility, ethical treatment is well developed: the quality failures documented in the empirical literature are framed as professional and ethical concerns, not merely technical ones. For the remaining six criteria: relevance and fitness for purpose, currency, sufficiency, understandability, consistency, and the avoidance of information overload, the ethical dimension is largely absent, despite each criterion being equally grounded in the Kantian principles established in the ethical framework section. The failure modes are documented, but they are treated as engineering problems rather than ethical failures. This asymmetry is itself a significant finding. It reflects the profession’s tendency to engage most deeply with the most visible and most easily documented failures, while the subtler failures, the sycophantic response that appears relevant but is not, the confident answer that is months out of date, the incomplete summary that omits what the user most needed, the variability that means different users receive different answers to the same question, have not yet attracted the ethical attention they deserve. The argument that each of these failures constitutes a specific violation of one or more Kantian principles, and that they therefore represent professional ethical obligations rather than merely service quality targets, is a contribution of this paper.