Before You Say Yes to AI: The Questions That Protect the Work
A culturally alert framework for deciding when artificial intelligence belongs in a studio, product, institution, or landscape—and when restraint is the more innovative choice.
Beatrice OkonkwoCritic at largeFirst published 10/3/2026 · monitored for updates; the next revision publishes a new version and appears here. Reader corrections are reviewed and folded into future versions.
Summary
AI is arriving as a material, a supplier, a collaborator, and sometimes an unwanted intermediary between people and culture. Before committing, founders and creative teams should ask not only whether a system works, but what it makes cheaper, less visible, harder to reverse, or newly possible. The essential questions concern purpose, provenance, consent, ecology, labor, control, and aesthetic consequence. In Iceland and other small cultural ecosystems, where language, landscape, trust, and authorship are unusually legible, those questions are not peripheral ethics: they are product strategy.
Key takeaways
- Begin with the human or cultural value you want to create—not with a model searching for a use.
- Treat training data, prompts, outputs, and user submissions as a provenance chain that must be inspectable.
- Ask whether automation preserves expertise or quietly removes the conditions in which expertise develops.
- Calculate energy, infrastructure, review time, licensing, and exit costs—not merely the API price.
- Keep consequential decisions reviewable by a named person with authority to stop deployment.
- Prefer reversible experiments, narrow scopes, and measurable thresholds over vague transformation programs.
- Judge the aesthetic result: efficiency is not enough if every voice, image, menu, or place begins to feel interchangeable.
- Plan an exit before integration; a tool is not truly optional if your archive, workflow, or audience becomes trapped inside it.
Explain like I'm 5
Imagine hiring a very fast assistant who has read an enormous, imperfect library but cannot reliably remember where each idea came from. The assistant can draft labels, sort photographs, translate a menu, or sketch variations. It can also invent facts, copy a familiar style, expose private material, and sound certain when it is wrong. Before giving it a job, ask: What exactly is it helping us do? Who supplied what it learned from? May we share this material with it? Who checks the answer? What happens if it fails? Can we leave later without losing our work? If those questions have no clear owner, the experiment is not ready to become infrastructure.
Deep dive
Start with the promise, not the machinery
The weakest AI brief is simply: ‘We should use AI.’ A stronger brief names a desired change: reduce the time museum staff spend transcribing oral histories; help visitors encounter Icelandic place-names without flattening pronunciation; let an architect compare daylight strategies earlier; or make a small food producer’s archive searchable. Then ask whether AI is necessary. Search, rules, a well-designed form, or another employee may be cheaper and more trustworthy. Define a baseline, a user, and a threshold for success before selecting a model. If the benefit cannot be expressed without words such as ‘innovation’ or ‘efficiency,’ it probably has not been designed yet.
Whose intelligence is inside the system?
Every output carries a hidden supply chain: datasets, annotators, artists, writers, software libraries, cloud providers, and users whose interactions may improve the service. Ask the vendor to identify data sources, licensing position, retention policy, subprocessors, and whether prompts or outputs are used for training. For culturally specific work, generic legality is an inadequate standard. A digitized saga, a craft pattern, an elder’s recorded voice, and contemporary illustration create different obligations. Consent must be granular enough to distinguish preservation from synthetic reuse. Attribution should remain attached where practical, and communities should have meaningful ways to refuse uses that alter context or identity.
Will it deepen taste—or average it?
Generative systems are powerful variation engines, but statistical fluency can become aesthetic sameness. A Reykjavík hotel, gallery, or restaurant may gain polished copy while losing the cadence that located it in a real street, climate, and community. Ask what should remain difficult, situated, or unmistakably authored. Compare outputs not only for accuracy but for surprise, specificity, emotional temperature, and cultural fit. Maintain reference sets of exemplary human work and record why it succeeds. AI can productively widen a sketch phase, expose alternatives, or handle low-value repetition; final judgment should belong to someone able to defend a choice rather than merely approve a plausible result.
What happens at the edge of failure?
A beautiful demo usually presents the center of the distribution. Real products meet dialect, irony, missing records, sensitive disclosures, unusual bodies, indigenous or minority names, and adversarial users. Test there. Build an error taxonomy: fabrication, omission, bias, privacy leakage, unsafe advice, mistranslation, and stylistic appropriation. Decide which failures are tolerable and which require blocking, escalation, or withdrawal. A chatbot offering opening hours has a different risk profile from a system interpreting medical symptoms for remote travelers. ‘Human in the loop’ means little unless a trained person has enough time, evidence, and institutional power to overrule the machine.
Count the whole footprint
AI’s price is larger than a subscription. Include integration, evaluation, red-teaming, legal review, staff training, moderation, latency, model changes, and the energy and water associated with computation. Iceland’s renewable electricity does not make digital infrastructure impact-free: data centers still occupy land, require construction and transmission, and compete within local planning debates. Ask whether a smaller model, retrieval system, batch process, or conventional software can meet the need. Measure use per completed task, not impressive benchmark score. Environmental accounting should also consider avoided travel, waste, prototypes, or rework, rather than assuming every computation is either inherently green or inherently harmful.
Design the exit before the entrance
Commitment becomes dangerous when dependencies are invisible. Can prompts, embeddings, annotations, evaluations, and generated assets be exported in usable formats? Can another model replace the current one? What happens if pricing changes, a feature disappears, the provider is acquired, or regulation restricts deployment? Keep source files and canonical records outside proprietary systems; document model versions and significant prompts; separate vendor-specific code behind an interface. Establish review dates and a sunset test: if the tool no longer meets accuracy, cultural, cost, or environmental thresholds, who can turn it off? The mature AI strategy is not permanent adoption. It is the capacity to adopt selectively, learn quickly, and leave cleanly.
- 1950Alan Turing publishes ‘Computing Machinery and Intelligence,’ reframing machine intelligence as a testable question.
- 1956The Dartmouth workshop gives artificial intelligence its enduring name and research agenda.
- 2012AlexNet’s ImageNet result accelerates deep learning across vision, research, and commercial products.
- 2017Google researchers publish ‘Attention Is All You Need,’ introducing the transformer architecture behind modern language models.
- 2019OpenAI releases GPT-2 in stages, foregrounding debate about misuse and controlled model release.
- 2022ChatGPT brings conversational generative AI to a mass audience, reportedly reaching 100 million monthly active users rapidly.
- 2023Iceland’s government and OpenAI announce work to improve GPT-4’s Icelandic language capability with local expertise.
- 2024The European Union adopts the AI Act, establishing phased, risk-based obligations for providers and deployers.
- 2025The EU AI Act’s general-purpose AI provisions begin applying under the regulation’s staged implementation calendar.
FAQs
What is the first question to ask about an AI proposal?+
Ask what valuable outcome cannot be achieved as well with simpler means. Name the user, baseline, measurable improvement, and acceptable failure rate. If the proposal begins with a tool rather than a need, return it to discovery.
How can a small creative business investigate training data?+
Request the provider’s model card, data and copyright documentation, retention terms, and list of subprocessors. For commissioned work, contractually define whether your inputs and outputs may be used for training. A vendor’s silence is itself useful risk information.
Is human review enough to make an AI system safe?+
Not automatically. Reviewers need relevant expertise, sufficient time, access to sources, and authority to reject output. Teams should measure reviewer agreement and the errors that pass through, rather than treating oversight as a ceremonial checkbox.
Should artists disclose AI assistance?+
Disclosure is prudent when AI materially shapes an image, voice, text, performance, or claim of craft. The useful standard is whether a reasonable audience, commissioner, collaborator, or rights holder would understand the work differently if the process were known. Context matters more than a universal badge.
Can AI help preserve Icelandic language and culture?+
It can improve search, transcription, accessibility, language tools, and encounters with archives. It can also standardize dialect, hallucinate history, or detach cultural material from provenance. Preservation projects therefore need Icelandic-language evaluation and stewardship by relevant institutions and communities.
How should a team run its first pilot?+
Choose a bounded, reversible task with low consequences and a reliable human baseline. Predefine quality, cost, privacy, and cultural-fit metrics, then test unusual cases as deliberately as common ones. Keep a written decision log and an end date.
What belongs in an AI exit plan?+
Specify export formats, data deletion, replacement procedures, record retention, intellectual-property status, and who can authorize shutdown. Preserve canonical assets outside the vendor platform. Estimate migration cost before dependency grows.
When is the correct decision not to use AI?+
Decline it when consent is absent, errors would cause disproportionate harm, provenance cannot be established, or automation would destroy the value being offered. A craft studio, intimate archive, or culturally sensitive service may gain distinction through explicit human authorship.
Predictions
- Creative procurement may shift from asking whether a product ‘has AI’ to demanding model cards, provenance records, evaluation results, and clean export paths.
- Smaller language communities may build more sovereign datasets and evaluation suites, with access governed as cultural infrastructure rather than treated as open raw material.
- A visible ‘made without generative AI’ position could become a premium signal in illustration, hospitality, publishing, and craft—although verification will remain difficult.
- On-device and smaller domain models will likely gain appeal where privacy, latency, energy use, or offline operation matter more than maximum generality.
- AI literacy may evolve from prompt技巧 into editorial judgment: source checking, task design, failure analysis, rights negotiation, and knowing when not to automate.
Opportunities
- Build provenance interfaces that let museums, publishers, studios, and audiences trace source, consent, model version, and human intervention without reading technical logs.
- Create Icelandic-language evaluation datasets covering inflection, dialect, place-names, folklore, tourism claims, and contemporary cultural vocabulary.
- Design ‘slow AI’ tools for artists and architects: systems optimized for reflection, comparison, material constraints, and documented choices rather than instant volume.
- Offer independent AI due diligence for small cultural organizations that cannot maintain internal legal, security, accessibility, and evaluation teams.
- Develop low-compute products that pair trusted collections with retrieval and citation, giving archives and destinations useful interfaces without training a foundation model.
For professionals
For product leaders, the commitment decision should be governed as a portfolio of explicit claims. Write a system card before procurement: intended use, prohibited use, impacted groups, data classification, model and hosting options, baseline process, evaluation set, acceptance thresholds, escalation path, monitoring cadence, and decommission trigger. Separate model quality from system quality. A strong benchmark score does not test retrieval freshness, interface-induced overtrust, queue latency, reviewer fatigue, or the commercial incentives surrounding deployment. Evaluate by task and subgroup, retain representative failures, and version every consequential component. For cultural applications, add a stewardship layer beyond conventional security and compliance. Map rights and responsibilities across creator, subject, community, custodian, model provider, deployer, and audience. Legal permission should be treated as the floor; legitimacy may require contextual integrity, reciprocal benefit, attribution, or non-use. Procurement should include audit access, breach notification, deletion commitments, intellectual-property allocation, service-change notice, portability, and indemnity proportionate to risk. Use staged gates—sandbox, limited pilot, monitored release, renewal—and require fresh approval when the model, dataset, user population, or purpose materially changes. The strategic asset is not any particular model. It is the organization’s capacity to formulate good tests, preserve institutional memory, and exercise informed refusal.
Sources & references
- Regulation (EU) 2024/1689 — Artificial Intelligence Act
- NIST AI Risk Management Framework 1.0
- OECD AI Principles
- UNESCO Recommendation on the Ethics of Artificial Intelligence
- Attention Is All You Need
- On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?
- Government of Iceland: Head Start for Icelandic
- Stanford AI Index Report 2024
| Conventional software | AI-assisted workflow | AI-led experience | |
|---|---|---|---|
| Best fit | Stable rules, forms, filters, databases | Drafting, tagging, search, variations with review | Open-ended conversation or generation central to the product |
| Typical cost shape | Higher upfront specification; predictable operation | Moderate integration plus recurring inference and review | Continuous model, moderation, evaluation, and support costs |
| Primary failure | Brittleness when rules omit a case | Reviewer overtrust or inconsistent output | Plausible fabrication at user-facing scale |
| Cultural control | High when content is authored and curated | High to medium with references and approval gates | Lower unless tightly grounded and constrained |
| Reversibility | Usually high with open formats | Medium; preserve prompts, logs, and source assets | Often low after audience behavior and data pipelines depend on it |
| Evidence required | Functional testing and accessibility review | Task evaluation against expert human baseline | Red-teaming, subgroup tests, monitoring, incident response |
A practical philosophy for producing fewer, stronger ideas—using deliberate pace, intelligent constraints, and humane systems to turn cultural observation into durable creative 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.
The next interface may not wait for commands. It will interpret intent, assemble tools, negotiate systems, and act—turning software from a collection of destinations into a designed field of agency.
AI is neither a synthetic mind, an instant job-destroyer, nor an impartial machine. Seeing it clearly reveals better products, richer creative practices, and more durable opportunities.
A field guide to the software entities that perceive, decide, act, and learn—and to the creative and commercial possibilities emerging around them.
From our own rounds
Measured on The Curator, from real sessions people played on this site — not a third-party dataset.
- Rounds played here
- 163
- Questions per round
- 1.7