Before You Commit to AI, Ask What It Will Change
A culturally alert framework for choosing artificial intelligence without surrendering authorship, craft, trust, or the qualities that made the work valuable.
Jonah WhitcombePolitics & policyFirst published 10/19/2026 · monitored for updates; the next revision publishes a new version and appears here. Reader corrections are reviewed and folded into future versions.
Summary
AI commitments rarely begin with a dramatic decision. They arrive as a convenient transcription tool, an image generator in a design workflow, a chatbot trained on visitor questions, or an automated layer inside a cultural product. Yet each small adoption can quietly reshape authorship, labor, archives, aesthetics, privacy, and dependence on distant infrastructure. For founders, artists, museums, architects, and destination storytellers—especially in a place such as Iceland, where language, landscape, and cultural credibility are core assets—the useful question is not simply whether AI works, but what kind of future its use rehearses.
Key takeaways
- Begin with the human or cultural purpose, not the availability of a model.
- Ask whether AI is assisting judgment or replacing the judgment audiences came to encounter.
- Treat Icelandic-language performance, local context, and cultural nuance as acceptance criteria—not decorative extras.
- Map where training data, prompts, outputs, and user records travel before signing a contract.
- Price reversibility: exports, migration, human fallback, and the cost of rebuilding lost skills.
- Separate impressive demonstrations from reliable performance under real conditions.
- Disclose synthetic material whenever provenance could affect trust, rights, or interpretation.
- Measure success in quality, access, resilience, and cultural value—not only hours saved.
Explain like I'm 5
Imagine hiring a very fast apprentice who has read an enormous, imperfect pile of material but does not truly understand your place, audience, or intentions. The apprentice can make useful drafts, sort information, and suggest patterns. It can also invent facts, imitate styles, and sound certain when it is wrong. Before giving it a job, ask what task it should do, what material it may see, who checks its work, and whether you can dismiss it without losing your files or your own ability to work. If visitors, customers, or collaborators encounter its output, tell them when that knowledge matters. AI should earn a role through evidence; novelty alone is not a qualification.
Deep dive
Start with the value that must survive
The first question is deceptively simple: what must remain true after AI enters the work? A Reykjavík design studio may need faster concept exploration, but not at the cost of a recognizable voice. A museum might want multilingual access while preserving curatorial accountability. A tourism founder may automate itinerary drafts yet still need advice that respects weather, road closures, fragile sites, and local etiquette. Write these non-negotiables before comparing vendors. Otherwise the tool defines the brief. Then ask whether the proposed task is actually suitable. Low-stakes, reviewable work—transcription, metadata suggestions, internal search, rough summaries—often offers a sensible opening. Public safety advice, conservation guidance, historical interpretation, grant decisions, and claims about living artists require stronger evidence, oversight, and sometimes no automation at all. The relevant boundary is not whether a model can produce an answer; it is whether your organization can responsibly detect and repair its failures.
Interrogate the material beneath the magic
Generative systems are shaped by training data, retrieval sources, feedback, and the prompts supplied by users. Ask the vendor what data trained or fine-tuned the model, whether creators could consent or opt out, and whether your inputs may be retained or used for improvement. For collections and archives, distinguish public access from permission to extract, recombine, or commercialize. Copyright law remains unsettled across jurisdictions, while moral rights, contracts, confidentiality, and Indigenous or community protocols may impose additional limits. Language deserves its own test. Icelandic is not merely English with substituted words; inflection, names, idiom, register, and cultural references matter. Evaluate models with native speakers and domain experts using representative material. Record error rates for place names, historical assertions, pronunciation, and ambiguous queries. A fluent interface can conceal a shallow cultural model. If the system routinely anglicizes local concepts or presents folklore as literal history, that is a product defect.
Locate authorship and accountability
Ask who is making the meaningful choices. If an illustrator generates hundreds of options but selects, redraws, and composes the final work, AI may function as exploratory material. If a publisher types a short prompt and markets the result as an artist’s considered vision, the account of authorship is thinner. The distinction affects credit, compensation, audience trust, and the character of the work. Create a visible chain of responsibility: who approved the use case, who reviews outputs, who can stop deployment, and who responds when harm occurs? “The model did it” is not governance. For public-facing systems, define escalation routes to a person, label uncertainty, retain source links where possible, and log consequential changes. Red-team the experience with people unlike the commissioning team, including disabled users, Icelandic speakers, working artists, and those represented in archival material.
Count dependencies, not just efficiencies
An attractive subscription price rarely captures the full commitment. Add integration, evaluation, staff review, legal analysis, energy use, model changes, and the possibility that usage-based fees rise with success. Consider concentration risk: a cultural organization may build its visitor experience on an external API whose prices, moderation rules, model behavior, or availability can change with limited notice. Demand exportable data, documented workflows, version records, and a credible exit plan. Preserve a human route through essential services. If the AI vanished for a month, could the work continue? If not, the organization has acquired infrastructure rather than a tool—and should govern it accordingly. A pilot should test reversibility as deliberately as capability.
Decide what better means
Time saved is easy to count and easy to overvalue. Pair it with measures that reflect the actual promise: factual accuracy, editorial distinction, visitor comprehension, accessibility, conversion, employee satisfaction, complaint rates, and the share of outputs requiring correction. Compare against a non-AI baseline, not against doing nothing. Finally ask what the commitment teaches the organization to become. Some uses expand access: searchable archives, captions, translation drafts, or interfaces for people unable to navigate conventional collections. Others flatten the very qualities audiences value—local knowledge, patient observation, material intelligence, and contact with a responsible person. The strongest AI choice may be a bounded deployment, a slower experiment, or a refusal. Discernment is not resistance to innovation; it is how innovation acquires taste.
- 1950Alan Turing publishes “Computing Machinery and Intelligence,” proposing the imitation game as a practical test of machine behavior.
- 1956The Dartmouth Summer Research Project helps establish artificial intelligence as a named field.
- 2012AlexNet’s ImageNet result accelerates the deep-learning era and demand for large datasets and computation.
- 2018The Icelandic government and Almannaróm establish a national language-technology program to strengthen Icelandic in digital systems.
- 2021OpenAI introduces DALL·E, intensifying debate about generated imagery, creative labor, and training data.
- 2022ChatGPT launches publicly on November 30, bringing conversational generative AI into mainstream workflows.
- 2023The Icelandic government partners with OpenAI and Miðeind to improve GPT-4’s Icelandic-language capability.
- 2024The European Union adopts the AI Act, creating phased, risk-based obligations for AI providers and deployers.
- 2025The first EU AI Act provisions, including AI literacy duties and prohibited practices, begin applying on February 2.
FAQs
What is the best first question before adopting AI?+
Ask which human, cultural, or business outcome must improve—and how you will recognize improvement. If the answer is only “we should use AI,” there is no defensible brief yet.
Which projects make sensible pilots?+
Choose a bounded, reversible task with accessible ground truth and low consequences when errors occur. Internal transcription, metadata suggestions, and draft translation can be useful, provided a qualified person reviews the result.
How should an Icelandic organization test language quality?+
Use native speakers from the intended audience, not translated English benchmarks alone. Test inflection, names, idiom, register, pronunciation, source fidelity, and questions involving Icelandic history or geography.
Can confidential material be entered into a public chatbot?+
Not by default. Check contractual retention, training, security, hosting, deletion, and access terms, then follow organizational policy and applicable data-protection law; use approved enterprise or local systems where warranted.
Does human review solve hallucination?+
It reduces risk only when reviewers possess relevant expertise, enough time, and access to sources. Routine exposure to plausible output can also cause automation bias, so review should be structured rather than ceremonial.
Must audiences be told that AI was used?+
Disclosure is especially important when synthetic production changes how a work, voice, recommendation, or historical account would be understood. Explain the meaningful role of the system rather than applying a vague label that reveals nothing.
How do we avoid vendor lock-in?+
Secure exports in usable formats, keep prompts and evaluation sets outside the vendor platform, document integrations, and maintain replacement options. Include termination, deletion, model-change, and continuity terms in procurement.
When is refusing AI the innovative choice?+
Refusal is rational when automation damages trust, safety, rights, cultural integrity, or the central pleasure of the experience. It can also clarify a premium proposition built around human attention, provenance, and craft.
Predictions
- Cultural institutions may increasingly publish AI-use statements beside accessibility, collections, and sustainability policies, particularly when synthetic interpretation reaches visitors.
- Small-language evaluation could become a competitive specialty: Icelandic organizations may build shared test sets for morphology, names, history, and local safety information.
- Provenance signals—content credentials, source trails, model records, and human credits—are likely to become part of premium product design rather than mere compliance.
- AI procurement may shift from buying a single impressive model toward assembling replaceable layers: approved data, retrieval, evaluation, interfaces, and human review.
- Experiences explicitly made by people—guided walks, craft, performance, editorial curation—may gain status as audiences tire of abundant, frictionless synthetic media.
Opportunities
- Build Icelandic-language evaluation, annotation, and cultural-context services for companies whose general benchmarks miss local errors.
- Create rights-conscious discovery tools for museums and archives that retrieve source material without pretending generated prose is the collection itself.
- Design provenance-first creative software that records sources, model versions, edits, consent, and contributor credits as part of the artifact.
- Develop weather- and conservation-aware visitor assistants grounded in authoritative Icelandic sources, with explicit uncertainty and human escalation.
- Offer “human-made” cultural products whose value rests on traceable authorship, local knowledge, material craft, and deliberate limits on automation.
For professionals
For a serious deployment, turn the questions into a stage-gated assurance case. Define intended use, excluded uses, affected groups, legal basis, data lineage, evaluation thresholds, monitoring cadence, incident ownership, and decommissioning conditions. Test a fixed model version against a representative dataset and a credible non-AI baseline; segment results by language, user group, and failure severity rather than reporting one average score. Record false confidence, source attribution, prompt-injection resistance, accessibility, latency, cost, and reviewer burden. NIST’s AI Risk Management Framework and ISO/IEC 42001 provide useful scaffolding, while EU AI Act classification and GDPR duties require case-specific legal assessment. Procurement should treat model behavior as changeable supply-chain behavior. Contract for notice of material model or policy changes, retention limits, subprocessor visibility, security reporting, deletion, audit evidence, intellectual-property allocation, service continuity, and portable exports. Maintain model cards, system instructions, retrieval sources, evaluation sets, approvals, and release logs. For cultural data, add an authority layer beyond legal permission: who is entitled to contextualize, translate, or synthesize the material? A technically accurate system can still be culturally illegitimate. Governance is strongest when curators, designers, technologists, counsel, frontline staff, and represented communities can all challenge the deployment before the public must live with it.
Sources & references
- NIST AI Risk Management Framework (AI RMF 1.0)
- Regulation (EU) 2024/1689 — Artificial Intelligence Act
- OECD AI Principles
- UNESCO Recommendation on the Ethics of Artificial Intelligence
- The Icelandic Language Technology Programme
- Government of Iceland: Head Start for Icelandic
- Stanford AI Index Report 2024
- Content Credentials — C2PA
| External general model | Private or grounded system | Human-first workflow | |
|---|---|---|---|
| Best fit | Low-stakes ideation and generic drafts | Archive search, controlled assistance, domain answers | Signature art, sensitive interpretation, high-trust service |
| Setup burden | Low; account, policy, and basic evaluation | Medium–high; data preparation, retrieval, security, testing | Medium; staffing, training, scheduling, documentation |
| Cultural specificity | Variable; must be tested in Icelandic context | Potentially strong if sources and reviewers are authoritative | Strong when practitioners possess lived and domain knowledge |
| Data exposure | Depends heavily on provider tier and terms | More controllable, though infrastructure still needs auditing | Lower model exposure; ordinary privacy risks remain |
| Failure character | Plausible fabrication, imitation, generic style | Bad retrieval, stale sources, overconfident synthesis | Inconsistency, limited capacity, individual bias |
| Exit difficulty | Can become high through embedded APIs and habits | Moderate if data, prompts, and interfaces are portable | Low technical lock-in; knowledge may depend on key people |
A practical philosophy for producing fewer, stronger ideas—using deliberate pace, intelligent constraints, and humane systems to turn cultural observation into durable creative work.
A field guide to understanding generative AI, developing taste around it, and turning unfamiliar tools into thoughtful products, artworks, and opportunities.
A practical, culturally alert guide to the machinery behind generative AI—and how founders, artists, designers, and product thinkers can use it with greater taste, agency, and foresight.
AI products do not simply optimize intelligence. They negotiate among speed, cost, privacy, control, craft, labor, energy, and trust. The most distinctive builders make those bargains visible—and turn them into product character.
AI adoption is not merely a software decision. It is a choice about authorship, labor, trust, taste, and the future shape of your practice. This field guide helps creative teams interrogate the tool before the tool quietly redesigns the work.
AI is persuasive by default, discerning only by design. For Icelandic culture, creative work, and place-based products, the answer is not to abandon the technology but to give it boundaries, provenance, and human taste.
From our own rounds
Measured on The Curator, from real sessions people played on this site — not a third-party dataset.
- Rounds played here
- 173
- Questions per round
- 1.7