BEYOND CHAPTER 03 · Explore the changes

How AI changes cultural work

When a model can produce an answer quickly, which kinds of judgement must a cultural institution retain?

In this chapter

Imagine a museum receiving three proposals: use a model to draft labels, reconstruct colour in faded images and forecast next season’s attendance. All might be called AI, yet each needs different data, specialist review and measures of success. Defining the work to be improved is the first step towards deciding whether a tool is suitable.

AI can widen exploration while making gaps in its data appear fluent and convincing. This feature connects creation, knowledge organisation, visitor services and operational analysis through an accountable workflow: where inputs come from, what outputs mean and who can revise or reject them.

Distinguish tasks from models

Computers have long participated in music, writing and visual experimentation; today’s generative tools are one development within that history. Machine learning adjusts models using examples, and deep learning uses multilayer neural networks. Natural language processing handles language, while computer vision handles images. These categories overlap, but recognition, ranking, prediction and generation are distinct tasks that cannot be judged by one undifferentiated accuracy score.

Recognising text on a scanned page can be checked against a verified transcription. Suggesting ten exhibition narratives calls for an assessment of relevance, perspective and research. A grammatical collection description does not establish that the model has consulted the object’s registration file. Institutions should define the task, available evidence and consequences of error before choosing a tool.

Generation, style transfer and creative choices

Generative models can supply candidate compositions, melodies, texts or moving images. Goodfellow and colleagues’ 2014 generative adversarial network trained a generator against a discriminator; Gatys and colleagues’ 2015 style-transfer research explored combining representations of image content and style. These are historical examples of different methods, rather than a description of every contemporary AI architecture.

An artist may build a dataset, devise rules, shape inputs, select results and rework them by hand or with other software. Interactive works can respond to visitors, joining music, narrative and visual design. Evaluation should explain how these choices constitute the work instead of treating an unfamiliar output as proof of originality. Artistic ideas of authorship and legal eligibility for copyright also require separate discussion.

Curation, language and personalised interpretation

Search and recommendation systems can help curators navigate large collection records and propose connections across subjects, materials or periods. Language tools can draft summaries, translations and explanations at different reading levels, including dialogue in virtual exhibitions. These candidates still need checking against research, especially names, dates, quotations and attributions. Gaps in collection records should not become invented facts.

Personalised routes can respond to interests, language or available time, but a long dwell time may indicate confusion, waiting or reading difficulty rather than enjoyment. Recommending only popular works can further obscure overlooked collections. Curators can retain exploratory choices, explain recommendations and provide a standard tour. Visitor feedback should improve interpretation without replacing the curatorial argument.

Collections and the evidence limits of reconstruction

For collections, image recognition, text recognition and similarity search can assist initial classification, duplicate detection, photo matching or identification of surface changes for inspection. Similar brushwork is not authentication: attribution still requires provenance, material examination and art-historical research. Model-generated tags should carry a source and review status so that later researchers do not mistake automated additions for established knowledge.

AI colourisation or image completion can visualise a reconstruction hypothesis without proving that the original had those colours or details. Retaining source scans, marking speculative areas and separating material evidence from generated additions makes reconstruction open to scrutiny. A digitally complete image also does not mean the object has been conserved: physical care still involves material stability, environmental conditions and judgements about reversibility.

What market analysis and forecasting can answer

Models can organise public auction records, compare object attributes and flag price changes for investigation. Yet sales data often omit private transactions, unsold lots, condition details and negotiated terms; premium-inclusive prices must not be mixed indiscriminately with hammer prices. Forecasts should state their period, sample limitations and uncertainty, with holding, transaction and liquidity costs considered. They are not guarantees of appreciation.

Related methods can forecast attendance, staffing or maintenance needs, but should be compared with simple baselines such as the same period last year or a moving average. Weather, prices, school holidays and exceptional exhibitions can change results. Institutions can allocate resources across demand scenarios while retaining human adjustment; chasing predicted footfall alone risks crowding cultural diversity out of programming.

The cultural position of data

Training data influence what a model recognises, how it classifies and which language it uses to describe culture. A collection dominated by a few regions may leave other traditions misclassified or overlooked; colonial categories and derogatory terms in older catalogues may be reproduced at scale. Responses include examining dataset composition, testing across languages and groups, and involving relevant communities in naming and interpretation.

UNESCO’s 2021 Recommendation on AI ethics brings human rights, cultural diversity and human oversight into governance. It provides principles, not automatic certification of compliance for every institution. Practical questions remain: which data should be restricted, who can request corrections and which sensitive materials need limits on display. Bias can be investigated and reduced, rather than removed by declaring an algorithm entirely neutral.

Visitor data and people’s work

Visitor questions, search histories and tour preferences may contain personal data. Deployment requires clarity about purpose, an applicable lawful basis, retention, vendor reuse and less data-intensive alternatives. Unpublished donor information, loan agreements or sensitive collection records should not be placed in external systems without an authorised arrangement. A convenient interface does not explain where its data travel.

Automating repetitive tasks may release research time or create hidden work in checking, cleaning data and resolving errors. Curators, registrars, conservators, educators and technical staff should help design the process, with clear authority to approve publication and correct mistakes. Training and manageable workloads matter too. Counting only hours saved can underestimate the labour needed to sustain quality.

Start with an evaluable pilot

A useful pilot has a bounded task, verifiable evidence and correctable errors, such as drafting multilingual labels for a small, researched collection. Record the time and quality of the existing human process, then document model versions, inputs, edits and final approvers. NIST’s AI Risk Management Framework offers governance, contextual analysis, measurement and management as ways to connect responsibility to an actual workflow.

Evaluation should include error types, translation quality, visitor understanding, accessibility, energy use and total cost alongside speed. Agree in advance when to pause, use a manual process or leave the vendor’s service, so a pilot does not quietly become an unowned routine system. A valuable outcome may identify suitable assistance or establish that particular cultural judgements should remain directly human.

Your turn to investigate

Which clue would you check first?

A museum wants to reconstruct missing areas of a damaged painting with AI. Which proposal best supports research and public understanding?

Choose a clue to see what it can tell you.

READING COMPANION

A few useful terms

Machine learning: Adjusting models using data to perform classification, prediction or other tasks; performance depends on data and evaluation context.
Deep learning: A subset of machine learning using multilayer neural networks to learn data representations.
Generative model: A model that produces text, images, sound or other content from learned patterns; outputs are not necessarily true, original or free to use.
Generative adversarial network: A generative approach that trains a generator against a discriminator, usually abbreviated GAN.
Natural language processing: Computational methods for human language, including search, translation, summarisation and dialogue.
Computer vision: Methods for analysing visual features and relationships in images or video; results require validation for the intended use.
Recommendation system: A system ranking options using content, behaviour or other data; its ranking also shapes what becomes visible.
Digital reconstruction: A digital representation of a proposed past state of an object or setting, distinguishing evidence from conjecture.
Human oversight: Arrangements enabling people with appropriate authority and competence to check, revise, reject or stop system outputs.

Explore primary and secondary markets

YOUR TURN

How would you handle an engaging model-generated story about a poorly documented object?

Consider another way to think about it

Separate verifiable facts, plausible but unconfirmed inferences and invention, then consult object records and relevant communities. If used as imaginative interpretation, label it and explain what remains unknown. Preserving a gap can be more honest than manufacturing a complete narrative.

TAKE THIS THOUGHT WITH YOU

AI’s cultural value depends on accountable collaboration that distinguishes generative capability, evidence quality and human responsibility.

Reading notes & references

Organised around cultural workflows; technical and governance sources checked on 2 October 2026. Examples and research are read in their publication context.

Connect the ideas

Continue into people, works and research.