Full Text
1. Introduction
Information environments have changed substantially as organizations,
universities, libraries, publishers and public institutions have moved
from predominantly document-centred systems toward networked,
data-intensive and increasingly interactive services. The volume of
digital information has expanded, while users have come to expect
search, explanation and access to knowledge through interfaces that are
immediate and conversational. Information science has long examined how
information is created, represented, organized, retrieved, interpreted
and used. Generative artificial intelligence introduces a new layer into
each of these activities because systems can now produce language that
resembles human explanation and can transform information in response to
open-ended instructions.
The term generative AI describes computational systems that learn
patterns in training data and produce new content, including text,
images, audio, code and other media. In language applications, large
language models estimate the probability of token sequences and generate
text through repeated prediction. Their apparent flexibility arises from
large-scale pretraining, broad data exposure and architectures that can
represent relationships across long sequences. The transformer
architecture introduced by Vaswani et al. (2017) provided a scalable
attention-based alternative to recurrent sequence processing. Subsequent
work on pretrained language representations, including BERT (Devlin et
al., 2019), and large autoregressive models, including GPT-3 (Brown et
al., 2020), demonstrated how a single pretrained model could support
numerous downstream tasks.
By the beginning of 2023, the research community had already
identified both the promise and the systemic risks of foundation models.
Bommasani et al. (2021) described foundation models as models trained on
broad data that can be adapted to many downstream tasks. Their report
emphasized the possibility that a small number of shared models could
become infrastructure for a wide range of applications, while also
concentrating technical and social risks. Bender et al. (2021) warned
that very large language models can reproduce harmful patterns from
training data, conceal the conditions under which language was produced
and impose substantial environmental and financial costs. These concerns
are directly relevant to information institutions, whose
responsibilities include accuracy, access, privacy, intellectual freedom
and preservation of trustworthy records.
This review examines how generative AI and LLMs may transform
information science, knowledge management and digital communication,
while distinguishing technical capability from dependable information
service. It addresses the following questions: (1) What technical
developments made contemporary LLMs possible? (2) How can generative
systems support information discovery, organization and knowledge work?
(3) What risks arise when generated content enters organizational and
scholarly information flows? (4) Which design and governance practices
can support responsible deployment? The review is organized around these
questions and develops a framework for institutions considering
adoption.
2. Review Methodology
The article uses an integrative literature review. This approach is
suitable for a developing field in which relevant evidence is
distributed across computer science, information science, human-computer
interaction, organizational studies and responsible technology research.
The review focuses on foundational and influential literature available
by 31 December 2022. The date boundary is deliberate: the article is
framed around the technological and scholarly context leading into 2023,
and therefore does not use later publications as evidence for claims
about that period.
Relevant literature was identified through combinations of terms
including large language model, generative artificial intelligence,
transformer, language model, information retrieval, retrieval-augmented
generation, foundation model, knowledge management, conversational
search, algorithmic bias, model documentation and AI governance.
Priority was given to peer-reviewed conference proceedings and journal
articles, seminal technical papers and institutional reports with
identifiable authorship and publication details. Foundational works were
included where they explain the technical basis of current systems. The
selected literature was then organized into thematic categories:
architecture and training, capabilities, information-service
applications, risks, and governance.
This is a conceptual synthesis rather than a statistical
meta-analysis. The reviewed studies differ in model type, dataset, task,
evaluation method and application setting, making pooled effect
estimates inappropriate. The article does not claim that the reviewed
sources constitute an exhaustive systematic review. Instead, it uses a
transparent scope and a critical synthesis to identify recurring
findings, tensions and implications for information institutions. The
analysis also distinguishes peer-reviewed research from technical
reports and standards, and avoids presenting a model demonstration as
proof of general reliability.
3. Technical
Foundations of Large Language Models
3.1
From statistical language modelling to deep neural networks
Language modelling is concerned with estimating the probability of a
sequence of words or tokens. Earlier statistical approaches used
frequency counts and conditional probability estimates, often with
limited context windows. Neural language models expanded this approach
by learning distributed representations of words and sequences. Bengio
et al. (2003) showed how a neural probabilistic language model could
learn word representations and estimate sequence probabilities. This
line of research helped establish the principle that linguistic
regularities could be represented in learned vector spaces rather than
only in manually specified rules or sparse counts.
The development of word embeddings further improved the ability of
models to capture semantic and syntactic relationships. Mikolov et al.
(2013) demonstrated efficient methods for learning vector
representations from large text corpora. These representations were
useful, but static embeddings assign a word a largely fixed vector
regardless of context. Contextual representation models addressed this
limitation by computing representations based on surrounding tokens. The
shift from static to contextual representations was important because
the same word can carry different meanings in different sentences, and
information systems must often resolve such ambiguity. Contextualized
representations such as ELMo further demonstrated the value of deep
contextual language features (Peters et al., 2018).
The transformer architecture is based on attention mechanisms rather
than recurrence or convolution as the primary means of sequence
modelling. Vaswani et al. (2017) introduced an encoder-decoder
architecture in which self-attention allows each token representation to
incorporate information from other positions in a sequence. Multi-head
attention enables the model to learn several relationships in parallel,
while positional encodings provide information about token order. The
architecture is highly parallelizable during training, which made it
particularly suitable for large-scale computation. Earlier recurrent
approaches remained influential in language modelling, while transformer
systems changed the balance between sequential processing and parallel
computation (Radford et al., 2019; Liu et al., 2019).
Transformers are commonly adapted into encoder-only, decoder-only and
encoder-decoder configurations. Encoder-only systems such as BERT are
trained to produce contextual representations and are widely useful for
classification, semantic matching and extraction tasks. Decoder-only
models predict subsequent tokens and are naturally suited to text
generation. Encoder-decoder systems can map an input sequence to an
output sequence and are useful for translation, summarization and other
text-to-text tasks. These configurations are not interchangeable in
every setting; their suitability depends on whether the task requires
representation, generation or conditional transformation.
The attention mechanism also creates practical constraints.
Self-attention computation and memory requirements grow substantially
with sequence length, making long documents expensive to process
directly. Researchers have therefore explored sparse attention,
efficient attention approximations, recurrence-like memory mechanisms
and retrieval-based approaches. For information institutions, these
engineering choices matter because archival collections, legal documents
and scholarly articles may be much longer than the context window
supported by a deployed model.
3.3 Pretraining,
fine-tuning and prompting
Large language models are generally trained in stages. Pretraining
exposes a model to large quantities of text and optimizes it to predict
missing or subsequent tokens. This process allows the model to acquire
statistical regularities associated with grammar, style, facts and
patterns of discourse. Fine-tuning then adapts the pretrained model to a
narrower task or interaction pattern. In some cases, instruction tuning
uses examples of instructions and desired responses to make the system
more responsive to natural-language requests. Prompting provides another
adaptation mechanism by placing task instructions, examples or relevant
context in the model input without updating its parameters.
Brown et al. (2020) showed that scaling an autoregressive language
model to 175 billion parameters produced strong few-shot performance
across a range of tasks. Their results helped establish the practical
significance of in-context learning, in which a model uses examples
supplied in a prompt to infer the task. However, few-shot performance is
not equivalent to reliable reasoning or factual verification.
Performance varies by task, prompt construction, training distribution
and evaluation method. A model can produce a confident response even
when it lacks the information needed to answer accurately.
The distinction between pretraining knowledge and current information
is especially important. A model's parameters encode patterns learned
during training, but they do not automatically update when a new report,
policy or dataset is published. Knowledge may also be incomplete,
distorted or difficult to trace to a source. Lewis et al. (2020)
proposed retrieval-augmented generation (RAG), combining parametric
model memory with a retrievable non-parametric index. The approach can
provide access to external documents at generation time and can improve
the specificity and factual grounding of outputs in knowledge-intensive
tasks. Research on retrieval and memorization also illustrates that
language models may reproduce information from training examples, making
provenance and data governance important (Khandelwal et al., 2020).
3.4 Foundation
models and general-purpose capability
The foundation-model paradigm reflects the growing practice of
training a broad model once and adapting it to many downstream
applications. Bommasani et al. (2021) argued that this approach can
create powerful leverage because improvements to a shared model may
benefit numerous applications. Yet the same reuse can propagate
weaknesses. If a foundation model contains systematic errors, biases or
security vulnerabilities, downstream systems may inherit them. This
creates a governance problem that differs from the management of
isolated, task-specific software.
The apparent generality of foundation models should be interpreted
carefully. A model may perform well on a broad collection of benchmarks
while remaining unreliable in a particular institutional context.
Benchmark scores can hide differences in language, domain, user
population and the consequences of error. Information services should
therefore evaluate systems against representative local tasks, including
uncommon queries, multilingual material, specialized terminology and
requests that require precise source attribution.
Traditional information retrieval systems commonly return ranked
documents in response to a query. Users must inspect titles, abstracts,
snippets and full texts to determine which materials answer their
questions. Conversational interfaces can reduce the distance between a
user's information need and the information system by allowing follow-up
questions, reformulation and requests for explanation. An LLM can
translate a natural-language question into a search query, summarize
retrieved documents and help users compare competing sources. This can
be particularly valuable for novice users who do not know the specialist
vocabulary required by a conventional database.
Nevertheless, generative search should complement rather than
silently replace retrieval. A conventional search engine exposes
documents that users can inspect, whereas a generated answer may
compress several sources into a single narrative. If the system omits a
relevant document, misreads a passage or combines incompatible claims,
the user may not recognize the error. A responsible architecture should
display the retrieved sources, identify which passages support each
important claim and distinguish direct evidence from model-generated
interpretation. Search ranking, source selection and answer generation
should be evaluated separately.
The distinction between discovery and synthesis is also important for
scholarly communication. A system that helps a researcher locate
relevant literature is performing a different function from one that
synthesizes the literature into a claim. Discovery can be evaluated
through recall, precision, ranking quality and coverage. Synthesis
requires additional evaluation of factual consistency, completeness,
attribution and whether the generated text preserves disagreement in the
literature. An interface that presents both functions as a single
seamless answer risks obscuring these differences.
4.2 Digital libraries
and reference services
Digital libraries manage collections whose value depends on
description, organization, preservation and access. Generative systems
may assist with metadata enrichment, subject description, abstract
generation, language translation, accessibility transformations and
conversational reference. For example, an LLM could propose subject
terms for a newly deposited report, produce a plain-language description
of a technical paper or help a patron formulate a more precise search.
These functions may reduce repetitive work, particularly in collections
with large backlogs.
Automation must remain subordinate to cataloguing standards and
professional review. Incorrect names, dates, places or subject
assignments can impair discovery and create long-term errors in
institutional records. A generated summary may also introduce
information not present in the source. The appropriate workflow is
therefore assistive: the model proposes, the system records the source
and confidence indicators where available, and a trained professional
validates high-impact metadata before publication. Institutions should
preserve the original record and retain a history of machine-assisted
changes.
Reference services require particular attention to user context. A
conversational assistant may explain how to access a database, identify
relevant collections or guide a user through a research process. It
should not claim to have searched a collection when no search occurred,
and it should not invent holdings, opening hours or access conditions.
When the question concerns a specific item, the assistant should query
the authoritative catalogue or repository and cite the returned record.
Human librarians remain important for complex research consultations,
sensitive topics and situations in which a user's information need is
unclear.
4.3 Knowledge
management and organizational memory
Knowledge management includes the creation, capture, organization,
sharing and application of knowledge within organizations. Much
organizational knowledge is embedded in reports, policies, meeting
records, correspondence, manuals and the experience of employees. LLMs
can provide natural-language interfaces to this material, allowing staff
to ask questions across document collections and receive summaries or
draft responses. This may improve access to dispersed institutional
knowledge and reduce the time spent locating information.
However, an organizational knowledge assistant is only as reliable as
the content and permissions available to it. Documents may be outdated,
contradictory or approved for different audiences. If the system
retrieves an obsolete policy, it may provide an answer that appears
authoritative but is no longer valid. If access controls are not
enforced at retrieval time, the model may reveal information that the
user is not authorized to see. Effective implementation requires
document ownership, version control, retention rules, metadata quality,
access-aware retrieval and clear procedures for correcting errors.
The system should also distinguish explicit organizational knowledge
from generated inference. A response based on a signed policy should be
labelled differently from a suggestion inferred from several informal
documents. This distinction supports accountability and helps users
decide when to consult a responsible officer. Organizations should avoid
treating generated answers as official policy unless a designated
authority has reviewed and approved them.
4.4 Scholarly
communication and research workflows
Generative AI can support several stages of scholarly work, including
literature discovery, language editing, translation, coding assistance,
data-description drafting and the preparation of accessible summaries.
These capabilities may lower language barriers and help researchers
organize complex material. They can also assist editors and reviewers
with administrative tasks, such as checking whether required manuscript
sections are present or identifying inconsistent terminology. The
benefits are most defensible when the system supports a clearly bounded
task and the researcher remains responsible for the intellectual
content.
Academic integrity concerns arise when generated text is submitted
without appropriate disclosure, when fabricated references are included,
or when the system is used to produce analyses that the author cannot
explain. Language models can generate plausible bibliographic details
that do not correspond to real publications. They may also misrepresent
a cited source or create quotations that cannot be located. Researchers
must verify every reference against publisher records, DOI registries or
trusted bibliographic databases. Citation formatting should never be
treated as evidence that a source exists.
Editorial and peer-review settings raise confidentiality concerns.
Manuscripts under review may contain unpublished findings, personal
information or commercially sensitive material. Uploading such content
to an external model may breach journal policy or contractual
obligations. Journals and universities should establish rules covering
permitted tools, data handling, disclosure, authorship responsibility
and the treatment of confidential submissions. Human editors and
reviewers must retain responsibility for evaluations and editorial
decisions.
Generative systems can transform information into different forms: a
long report into a short briefing, a technical explanation into plain
language, or a written document into a conversational response. They may
support translation and assistive communication for users who face
literacy, language or disability-related barriers. Such applications
align with the broader information-science goal of making information
usable by diverse communities.
Yet translation and simplification can change meaning. Legal,
medical, scientific and regulatory texts often contain carefully defined
terms, qualifications and exceptions. A simplified summary that removes
a condition may become misleading. Multilingual performance is also
uneven because training data and evaluation resources are not
distributed equally across languages. Local dialects, minority languages
and domain-specific terminology may be poorly represented. Institutions
should test systems with the communities they serve and provide a route
to human assistance when meaning is uncertain.
5. Opportunities and
Institutional Value
The most immediate opportunity is the reduction of friction in
language-intensive tasks. Staff may use an LLM to draft an initial
summary, classify a batch of documents, generate alternative search
terms or explain unfamiliar terminology. These activities can improve
productivity when the output is checked and when time saved is
redirected toward professional judgement. The system should not be
evaluated only by the number of tasks automated. Quality, correction
effort, user satisfaction and the consequences of mistakes are equally
important.
A second opportunity is the development of more inclusive interfaces.
Conversational systems may help users who are unfamiliar with controlled
vocabularies or complex search syntax. They can provide multiple
explanations of the same concept and support iterative clarification.
This can make specialized information services more approachable.
However, accessibility must be assessed empirically rather than assumed
from the presence of a chat interface. Users may need keyboard
navigation, screen-reader compatibility, language options, clear
uncertainty signals and non-conversational alternatives.
A third opportunity is cross-document synthesis. Organizations often
need to compare policies, identify recurring themes in consultation
submissions or summarize a large body of technical reports. LLMs can
help identify candidate themes and produce an initial synthesis. The
process should retain traceability to the underlying documents and make
disagreements visible. A synthesis that removes conflicting evidence may
be easier to read but less useful for decision-making.
6. Risks, Limitations
and Ethical Challenges
6.1 Hallucination and
factual reliability
A central limitation of generative language models is that fluent
text is not a guarantee of factual correctness. The model is optimized
to generate likely sequences, not necessarily to retrieve verified facts
or abstain when evidence is insufficient. It may produce invented
citations, incorrect dates, inaccurate descriptions of legal
requirements or fabricated details about an organization. This behaviour
is often described as hallucination, although the term should not
obscure the technical reason: generation can continue even when the
model lacks reliable evidence.
The risk is particularly serious in information services because
users may interpret a polished response as authoritative. Mitigation
requires grounding answers in trusted sources, showing citations or
source passages, testing factual consistency and allowing the model to
express uncertainty. Systems should be designed to abstain or request
clarification when relevant evidence is absent. For high-stakes
decisions, generated output should be treated as an aid to professional
review rather than as a final determination.
6.2 Bias,
representation and language inequality
Training corpora reflect the social, linguistic and institutional
conditions under which digital content is produced. They may
overrepresent dominant languages and groups while underrepresenting
marginalized communities. Models can reproduce stereotypes,
discriminatory associations and culturally narrow assumptions. Bender et
al. (2021) emphasize that large-scale training can encode patterns from
the internet without adequate attention to whose language is represented
and whose interests are served.
Bias can enter through data collection, filtering, annotation, model
design, fine-tuning and deployment. It can also emerge through the way
users phrase questions and the way system designers define acceptable
answers. A model used in a multilingual university, public library or
government service should be evaluated across relevant languages,
dialects and demographic groups. Evaluation should include qualitative
review by affected communities, not only aggregate benchmark scores.
Institutions should document known limitations and provide a mechanism
for reporting harmful or discriminatory outputs.
6.3 Privacy,
confidentiality and security
Prompts and uploaded documents may contain personal data,
credentials, unpublished research, client records or confidential
organizational information. Depending on the service configuration and
contractual terms, submitted data may be logged, retained or used in
ways that are not obvious to the user. Information professionals must
therefore examine data processing terms, retention periods, access
controls, encryption, auditability and the location of processing before
adopting an external service.
Prompt injection and indirect instruction attacks create another
class of risk. A model connected to external documents or tools may
encounter malicious instructions embedded in retrieved content. If the
system treats those instructions as authoritative, it may disclose
information or perform actions outside the user's intent. Security
design should separate instructions from untrusted content, restrict
tool permissions, validate actions and log important operations. A
language model should not be given broad access to institutional systems
merely because it can interpret natural-language requests.
6.4 Intellectual property
and provenance
Generative systems raise questions about the use of copyrighted
material in training, the similarity of outputs to protected works and
the ownership or licensing status of generated content. The legal
position differs across jurisdictions and may depend on facts that
cannot be resolved by a general technical explanation. Institutions
should not assume that generated content is automatically free of rights
restrictions. Procurement and deployment policies should address
licensing, permitted uses, attribution, indemnity and procedures for
responding to rights-holder concerns.
Provenance is equally important for information quality. A generated
paragraph may combine several documents, general model knowledge and
inference. Without source-level traceability, a reader may be unable to
identify which statements are supported by which materials.
Retrieval-augmented systems can improve provenance by linking outputs to
retrieved passages, but retrieval alone does not guarantee that the
model has represented those passages faithfully. Citation checking and
source inspection remain necessary.
6.5 Environmental and economic
costs
Training and operating large models require computing infrastructure,
electricity, cooling, specialized hardware and technical expertise.
Bender et al. (2021) argue that the environmental and financial costs of
very large models should be considered before development and
deployment. The costs are not limited to initial training. High-volume
inference, repeated experimentation, data storage and model updates can
create continuing resource demands.
A responsible procurement decision should compare the expected
benefit with the full operating cost. Smaller task-specific models,
retrieval systems, caching, model quantization and selective use of
large models may provide adequate performance at lower cost.
Institutions should consider whether a generative model is necessary for
a task or whether conventional search, rules, templates or smaller
classifiers would be more reliable and efficient. Sustainability should
be treated as a design requirement rather than an afterthought.
The review suggests that responsible deployment requires a
sociotechnical approach. The model is only one component of an
information service. Data governance, interface design, user training,
professional oversight, security controls and institutional
accountability determine whether a technically capable model produces a
trustworthy service. The following principles translate the literature
into practical requirements.
First, define the service purpose and boundaries before selecting a
model. Institutions should specify the user group, information need,
permitted tasks, prohibited uses and consequences of error. A narrow,
measurable task is easier to evaluate than a general-purpose assistant
with unrestricted access. Second, use authoritative and maintained
sources for factual tasks. Where possible, connect the model to curated
collections and expose the source records to users. Third, enforce
access controls before retrieval, not merely after text generation.
Fourth, maintain human review for decisions that affect rights,
finances, health, employment, academic assessment or institutional
reputation.
Fifth, evaluate the complete workflow rather than the model in
isolation. Tests should include ordinary queries, ambiguous requests,
adversarial inputs, outdated documents, conflicting sources and cases
where the correct response is to abstain. Sixth, document the system's
intended use, training or retrieval data, known limitations, evaluation
results, update schedule and responsible owner. Seventh, create feedback
and incident-reporting channels. Errors should be recorded, investigated
and used to improve the service. Finally, provide users with clear
information about when they are interacting with AI and how generated
content should be verified.
Figure 1 presents a conceptual framework for responsible generative
information services. It positions the LLM as a language-processing
component within a governed information architecture. The user request
is first interpreted and checked against the service purpose. An
access-aware retrieval layer obtains relevant material from approved
sources. The model then generates a response conditioned on the
retrieved evidence. A validation layer checks source alignment, policy
compliance and output quality. The interface presents the answer with
provenance and uncertainty indicators, while human oversight is
available for cases that exceed the system's authority.
Figure 1. Responsible generative information-service
architecture
USER REQUEST
↓
Purpose, identity and risk assessment
↓
Access-aware retrieval from approved collections
↓
Evidence selection and source ranking
↓
Large language model generation
↓
Validation: factual alignment, policy, privacy and safety
↓
Response with citations, limitations and uncertainty
↓
User feedback, human escalation and audit logging
The framework has six interacting layers. The governance layer
defines policy, accountability and risk thresholds. The data layer
maintains authoritative documents, metadata, retention and permissions.
The retrieval layer searches approved sources and returns evidence
relevant to the request. The model layer performs language understanding
and generation. The assurance layer evaluates factual support, security
and policy compliance. The interaction layer communicates results,
citations and uncertainty to users. Feedback and audit records connect
the layers and support continuous improvement.
Table 1 summarizes the main application areas and the controls that
should accompany them.
| Application area |
Potential value |
Essential control |
| Conversational discovery |
Natural-language access, query reformulation and document
summaries |
Expose retrieved records; test recall, source coverage and citation
accuracy |
| Digital library services |
Metadata suggestions, plain-language summaries and reference
assistance |
Professional validation; preserve original metadata and record
changes |
| Organizational knowledge |
Search across policies, reports and internal documentation |
Access-aware retrieval, version control and named document
owners |
| Scholarly workflows |
Language support, literature triage and administrative
assistance |
Verify every citation; protect confidential manuscripts; disclose
relevant use |
| Multilingual communication |
Translation, simplification and accessible explanations |
Test with target language communities; preserve technical
meaning |
| Decision support |
Synthesis of evidence and comparison of options |
Human authority, traceable evidence and explicit limits on automated
decisions |
9. Evaluation and Measurement
Evaluation should reflect the purpose of the information service. For
information discovery, relevant measures include precision, recall,
ranking quality, query success and the diversity of useful sources
retrieved. For summarization, evaluation should examine factual
consistency, coverage of key points, preservation of qualifications and
readability. For question answering, evaluators should check whether
each material claim is supported by the cited source and whether the
system abstains when the collection does not contain an answer. Human
assessment is necessary because automatic metrics may reward fluent text
without detecting subtle distortions.
Operational measures are also important. Institutions should track
latency, availability, cost per successful task, correction time,
escalation rate, user satisfaction and incident frequency. These
measures reveal whether automation actually improves service. For
example, a system may generate a summary quickly but require substantial
staff time to identify unsupported claims. In that case, apparent speed
does not translate into productivity. Evaluation should compare the
AI-assisted workflow with the existing process and should include the
full cost of verification and maintenance.
Fairness evaluation should consider performance across user groups
and languages. Security evaluation should test unauthorized retrieval,
prompt injection, data leakage and misuse of connected tools. Governance
evaluation should examine whether users understand the system's
limitations, whether staff know how to report errors and whether the
organization can reconstruct how a response was produced. A model that
performs well in a demonstration but cannot be audited may be unsuitable
for an institution responsible for preserving trustworthy
information.
10. Discussion
The literature indicates that generative AI is best understood as a
new interface and transformation layer for information systems, rather
than as a replacement for information institutions or established
retrieval infrastructures. LLMs can make complex collections easier to
query and can reduce the effort involved in language-intensive work. Yet
their outputs remain dependent on the quality of training data, prompts,
retrieved evidence and deployment controls. The central
information-science challenge is therefore to integrate generative
capabilities without weakening the principles of provenance, authority,
access and accountability.
A key tension concerns convenience versus inspectability. Users value
concise answers, but information professionals need systems that
preserve the ability to inspect underlying sources. A conversational
answer that hides its evidence may increase convenience while reducing
the user's capacity to evaluate the response. This suggests that
interface design should not simply imitate human conversation. It should
make evidence visible, separate quotation from paraphrase and allow
users to move from a generated summary to the original document.
A second tension concerns generality versus local relevance.
Foundation models offer broad capabilities, but local information
environments contain specialized terminology, institutional rules and
language practices. A model trained on broad internet data may not
understand the meaning of a local policy or the distinctions used in a
specialist collection. Domain adaptation and retrieval can improve
relevance, but they introduce new responsibilities for data curation,
access management and evaluation. Institutions should not assume that a
general model is ready for deployment simply because it performs well on
public benchmarks.
A third tension concerns efficiency versus professional judgement.
Automation can reduce repetitive work, but professional information
services involve interpretation, ethical reasoning and understanding of
user context. The aim should be to redistribute effort rather than
eliminate professional responsibility. Librarians, archivists, records
managers, editors and knowledge specialists can use generative tools to
handle preliminary tasks while concentrating on collection quality,
complex enquiries, ethical questions and the evaluation of evidence.
11. Research Gaps and
Future Directions
First, more longitudinal studies are needed to establish whether
generative assistants improve information outcomes over time. Short
demonstrations may show that users can complete a task, but they do not
reveal whether users become over-reliant, whether error rates change
with repeated use or whether the service remains useful as collections
evolve. Longitudinal work should examine user trust, verification
behaviour, information literacy and the distribution of benefits across
different populations.
Second, evaluation methods for grounded generation require further
development. Existing measures do not always capture whether a generated
answer faithfully represents the retrieved sources, preserves
uncertainty or includes important counterevidence. Research should
develop source-level attribution measures and practical audit methods
that can be applied to real institutional collections. Third,
multilingual and low-resource language performance requires focused
investigation. This is especially important for public institutions
serving communities whose languages are underrepresented in large
training corpora.
Fourth, research should examine governance models for shared
foundation-model infrastructure. If many institutions rely on the same
small set of providers, they may face concentration risks, service
disruptions, price changes and limited control over data handling. Open
models can increase local control but may require significant technical
capacity and may not eliminate security or bias concerns. Comparative
research should examine the total cost, performance, privacy and
sustainability of hosted, self-managed and hybrid deployment models.
Fifth, the relationship between generative systems and information
literacy deserves attention. Users need skills to formulate questions,
inspect sources, recognize uncertainty and distinguish generated
language from verified evidence. Educational interventions should be
evaluated to determine whether they improve critical use rather than
merely increasing confidence. Finally, environmental assessment should
consider the full lifecycle of model development and use, including
hardware, electricity, cooling, data storage and repeated inference.
Research should compare large general-purpose models with smaller
specialized alternatives for common information tasks.
12. Conclusion
Generative AI and large language models represent a significant
development for information science because they can mediate the
creation, discovery, transformation and use of digital information.
Their capabilities offer opportunities for conversational search,
digital library services, organizational knowledge access, scholarly
workflows and multilingual communication. These opportunities are
meaningful, but they do not remove the need for authoritative sources,
professional expertise or institutional responsibility.
The review has shown that LLMs generate language through learned
statistical patterns and that their apparent generality is enabled by
transformer architectures, large-scale pretraining and adaptation
methods. Retrieval-augmented generation offers a practical route to
connect model outputs with external evidence, but retrieval does not by
itself guarantee faithful synthesis. Risks involving hallucination,
bias, privacy, security, intellectual property and environmental cost
must be addressed through system design and governance.
The proposed framework places generative models inside a controlled
information architecture that includes purpose definition, access-aware
retrieval, source provenance, validation, user communication, human
escalation and audit. Institutions should begin with bounded use cases,
evaluate complete workflows and retain human authority over high-impact
decisions. The future contribution of generative AI to information
science will depend not simply on producing more text, but on making
information more accessible while preserving the trust, accountability
and evidence standards on which knowledge institutions depend.
References
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S.
(2021). On the dangers of stochastic parrots: Can language models be too
big? Proceedings of the 2021 ACM Conference on Fairness, Accountability,
and Transparency, 610–623. https://doi.org/10.1145/3442188.3445922
Bengio, Y., Ducharme, R., Vincent, P., & Janvin, C. (2003). A
neural probabilistic language model. Journal of Machine Learning
Research, 3, 1137–1155.
https://www.jmlr.org/papers/v3/bengio03a.html
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von
Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E.,
Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N.,
Chen, A., Creel, K., Davis, J. Q., Demszky, D., ... Liang, P. (2021). On
the opportunities and risks of foundation models. Stanford Center for
Research on Foundation Models.
https://doi.org/10.48550/arXiv.2108.07258
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A.,
Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R.,
Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., ... Amodei, D. (2020).
Language models are few-shot learners. Advances in Neural Information
Processing Systems, 33, 1877–1901.
https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT:
Pre-training of deep bidirectional transformers for language
understanding. Proceedings of the 2019 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language
Technologies, 1, 4171–4186. https://doi.org/10.18653/v1/N19-1423
Floridi, L., & Chiriatti, M. (2020). GPT-3: Its nature, scope,
limits, and consequences. Minds and Machines, 30, 681–694.
https://doi.org/10.1007/s11023-020-09548-1
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B.,
Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020).
Scaling laws for neural language models. arXiv.
https://doi.org/10.48550/arXiv.2001.08361
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal,
N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., &
Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive
NLP tasks. Advances in Neural Information Processing Systems, 33,
9459–9474.
https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient
estimation of word representations in vector space. arXiv.
https://doi.org/10.48550/arXiv.1301.3781
Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L.,
Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019). Model
cards for model reporting. Proceedings of the Conference on Fairness,
Accountability, and Transparency, 220–229.
https://doi.org/10.1145/3287560.3287596
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena,
M., Zhou, Y., Li, W., & Liu, P. J. (2020). Exploring the limits of
transfer learning with a unified text-to-text transformer. Journal of
Machine Learning Research, 21(140), 1–67.
https://www.jmlr.org/papers/v21/20-074.html
Rogers, A., Kovaleva, O., & Rumshisky, A. (2020). A primer in
BERTology: What we know about how BERT works. Transactions of the
Association for Computational Linguistics, 8, 842–866.
https://doi.org/10.1162/tacl_a_00349
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L.,
Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all
you need. Advances in Neural Information Processing Systems, 30.
https://papers.nips.cc/paper/7181-attention-is-all-you-need
Peters, M. E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee,
K., & Zettlemoyer, L. (2018). Deep contextualized word
representations. Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computational Linguistics: Human
Language Technologies, 1, 2227–2237.
https://doi.org/10.18653/v1/N18-1202
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., &
Sutskever, I. (2019). Language models are unsupervised multitask
learners. OpenAI.
https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O.,
Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019). RoBERTa: A
robustly optimized BERT pretraining approach. arXiv.
https://doi.org/10.48550/arXiv.1907.11692
Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., & Lewis,
M. (2020). Generalization through memorization: Nearest neighbor
language models. International Conference on Learning Representations.
https://openreview.net/forum?id=HklBjCEKvH