Before You Say Yes to AI: The Questions That Protect the Work
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.
Beatrice OkonkwoCritic at largeFirst published 10/3/2026 · last revised 10/4/2026 with fresh sources, corrections, and new context. Reader corrections are reviewed and folded into future versions.
Summary
The most consequential AI decision often happens before procurement, prompting, or deployment: defining what must not be lost. For founders and creative teams, generative systems can compress research, accelerate iteration, and reveal unfamiliar directions. They can also flatten style, obscure provenance, expose confidential material, and replace productive struggle with plausible output. A responsible yes therefore begins with sharper questions. What problem is the system solving? Which human judgment remains decisive? What data enters and leaves the workflow? Can the result be traced, corrected, and defended? Who gains leverage, and who absorbs risk? This guide offers a practical framework for treating AI as a designed relationship rather than an automatic upgrade—one that protects authorship, trust, craft, and the distinctive intelligence of the people doing the work.
Key takeaways
- Begin with the value to preserve—not the capability on offer. Name the human skill, relationship, or standard that cannot be traded away.
- Ask whether AI removes drudgery or removes judgment. The first can expand a practice; the second can hollow it out.
- Map data provenance, permissions, retention, model training policies, and output ownership before confidential or client material enters a system.
- Use reversibility as a design principle: pilot narrowly, preserve original files, log decisions, and maintain a credible manual fallback.
- Evaluate outputs for sameness, factual error, bias, and cultural misappropriation—not only speed, polish, or cost.
- Disclose meaningful AI involvement where it affects trust, authorship, safety, or a client's ability to assess the work.
- Measure whether the tool improves originality and decision quality over time. Faster production of weaker ideas is not innovation.
Explain like I'm 5
Imagine a very fast assistant who has read an enormous pile of material but does not truly understand which facts are correct, which ideas belong to whom, or what your audience will consider meaningful. This assistant can make drafts, images, lists, code, and variations in seconds. Before giving it important work, you would ask: What can it see? Will it remember our secrets? How often is it wrong? Who checks the answer? Are we allowed to use what it makes? Does it help us think, or merely help us finish? AI deserves the same questions. The goal is not to reject the assistant; it is to give it a clear role, sensible boundaries, and a human editor with authority.
Deep dive
Start with the work worth protecting
AI arrives dressed as capability: faster concepts, cheaper images, instant synthesis, endless variants. Yet capability is not the same as value. Before selecting a model, write a preservation brief. Identify what makes the work worth doing and worth choosing: the founder's intuition, the illustrator's line, the researcher's skepticism, the studio's discretion, the customer's trust, or the productive friction through which an original idea appears. Then ask what the proposed system changes. Does it remove transcription and file sorting, or does it make the aesthetic decisions that distinguish your practice? A useful test is counterfactual: if every competitor used the same model tomorrow, where would your advantage remain? If the answer is nowhere, the adoption plan may be optimizing away your difference.
Define the job before admiring the machine
Do not begin with ‘How can we use AI?’ Begin with a bounded problem: reduce interview-coding time, generate accessibility descriptions, detect support themes, prototype interface states, or explore material combinations. Establish a baseline for time, cost, error rate, and creative quality. Then set a threshold for success. A system that drafts 100 concepts is not valuable if reviewing them consumes more attention than producing ten strong concepts manually. Separate assistance, recommendation, and decision. Assistance formats notes; recommendation ranks options; decision determines what ships, who receives credit, or which applicant advances. The closer AI moves toward consequential decisions, the stronger the evidence, oversight, and appeal mechanisms must become.
Trace the invisible supply chain
Every AI output has an upstream story. Investigate the model provider, training disclosures, third-party subprocessors, data location, retention period, security controls, and whether prompts or uploaded files are used to improve models. Consumer accounts and enterprise contracts may offer materially different protections. Ask whether designers are uploading unreleased products, whether researchers possess informed consent for new processing, and whether client contracts permit external model services. Downstream, clarify who owns outputs, whether they can be exclusively protected, and how similarity disputes will be handled. In the United States, the Copyright Office has repeatedly emphasized human authorship; purely machine-generated material may not receive copyright protection. Keep source files, prompts, edits, licenses, and human decisions when provenance matters.
Design human authority into the loop
‘Human in the loop’ is empty language unless someone has time, expertise, and permission to say no. Name the accountable reviewer and define what they inspect: factual claims, visual references, code security, bias, tone, or rights. Calibrate scrutiny to stakes. An internal mood-board prompt needs different controls from medical guidance, hiring recommendations, public journalism, or a campaign depicting a living artist's style. Build checkpoints before publication rather than relying on cleanup afterward. Preserve an independent route to the source material so reviewers are not trapped inside the model's summary. Most importantly, prevent automation bias—the tendency to trust a polished machine suggestion because challenging it feels slower.
Protect taste from statistical gravity
Generative AI excels at producing the legible center of a category: the image that looks like contemporary branding, the sentence that sounds professionally strategic, the product idea assembled from familiar signals. This fluency can create aesthetic convergence. Counter it deliberately. Curate references beyond platform defaults. Ask for tensions, constraints, historical lineages, excluded perspectives, and materially different directions—not simply more options. Use models early for divergence or late for production support, while reserving the central act of meaning-making for people. Compare AI-assisted work against blind human alternatives. Look for specificity, surprise, coherence, and emotional residue. Taste is not selecting the smoothest result; it is recognizing what deserves to exist.
Make the yes conditional and reversible
A mature adoption decision is rarely a permanent yes. It is a conditional permission attached to a use case, dataset, team, and review date. Run a limited pilot with non-sensitive material. Log failures and near misses, not just time saved. Set stop conditions—for example, unacceptable hallucination rates, rights complaints, customer confusion, or declining craft competency. Maintain version histories and a non-AI fallback for critical processes. Revisit the policy when a provider changes terms or models. Finally, decide what honest disclosure looks like. Audiences do not need a theatrical label on every automated spell-check, but they deserve clarity when synthetic media, simulated testimony, consequential recommendations, or substantial machine generation changes how the work should be interpreted. The best AI policy is not a brake on imagination. It is infrastructure for pursuing ambitious work without making trust disposable.
- 1956The Dartmouth Summer Research Project popularized the term ‘artificial intelligence,’ framing machine intelligence as a research program.
- 2012AlexNet's ImageNet performance accelerated deep learning adoption and transformed computer vision, data demand, and GPU-led model development.
- 2017Google researchers published ‘Attention Is All You Need,’ introducing the Transformer architecture behind many modern generative models.
- 2020OpenAI introduced GPT-3, demonstrating that large language models could perform varied tasks through prompting rather than task-specific retraining.
- 2021The EU proposed the Artificial Intelligence Act, advancing a risk-based regulatory model for AI systems.
- 2022Stable Diffusion and ChatGPT brought text-to-image generation and conversational AI into mainstream creative and professional workflows.
- 2023The U.S. Copyright Office launched an AI initiative and issued guidance requiring applicants to disclose and disclaim more-than-de-minimis AI-generated material.
- 2024The EU AI Act entered into force on August 1, beginning phased obligations for prohibited practices, general-purpose AI, and high-risk systems.
- 2025–2027Major EU AI Act obligations phase in across this period, making documentation, literacy, transparency, and risk governance operational concerns for many organizations.
Glossary
- Automation bias
- The tendency to favor a machine-generated recommendation even when contrary evidence is available.
- Data provenance
- A record of where data originated, how it was collected, what permissions apply, and how it has been transformed.
- Foundation model
- A broadly trained model that can be adapted or prompted for many downstream tasks, such as writing, coding, or image generation.
- Hallucination
- A fluent but unsupported or false model output, including invented facts, citations, quotations, or technical details.
- Human-in-the-loop
- A workflow in which a qualified person reviews, changes, approves, or rejects an AI system's input or output.
- Model collapse
- Potential degradation that can occur when models are increasingly trained on synthetic outputs rather than reliable human-originated data.
- Prompt injection
- An attack or instruction designed to make an AI system ignore intended rules, disclose information, or perform an unauthorized action.
- Retrieval-augmented generation
- A method that supplies a model with selected external sources at query time to improve relevance and grounding.
- Synthetic media
- Images, audio, video, text, or other content generated or substantially altered by computational systems.
FAQs
What is the first question to ask before adopting an AI tool?+
Ask what valuable constraint, skill, or relationship the tool must preserve. Only then define the problem it should solve and the evidence that would justify adoption.
Should confidential client work be entered into a public AI service?+
Not without an approved contract, security review, documented retention and training terms, and client authorization where required. A provider's enterprise tier may differ significantly from its consumer product.
Who should approve AI-generated creative work?+
A named person with domain expertise, sufficient review time, and authority to reject it. Accountability should not be assigned vaguely to ‘the team.’
Can AI-generated work be copyrighted?+
Rules vary by jurisdiction. In the United States, copyright protects human-authored expression, not purely machine-generated material. Human selection, arrangement, and modification may qualify, but document those contributions and seek counsel for important assets.
When should AI use be disclosed?+
Disclose when its use materially affects authorship, authenticity, interpretation, safety, or a stakeholder's decision. Synthetic people, cloned voices, consequential recommendations, and substantial generated content warrant particular care.
How can a studio prevent AI from flattening its style?+
Use proprietary references only with permission, maintain human-led art direction, compare outputs against non-AI work, and evaluate specificity and surprise rather than volume or surface polish.
What metrics belong in an AI pilot?+
Track time and cost alongside factual error, correction effort, rights risk, output diversity, user trust, reviewer workload, and whether human skill improves or atrophies.
Is an open-source model automatically safer?+
No. Open models can improve inspectability and deployment control, but safety also depends on training data, configuration, hosting, access controls, maintenance, and the use case.
How often should an AI policy be reviewed?+
At least annually, and sooner after a major model update, provider terms change, security incident, legal development, new use case, or material failure.
Predictions
- AI procurement will become closer to supply-chain governance: buyers will demand model cards, data-flow maps, incident procedures, audit rights, and clear subprocessor lists.
- Provenance will become a product feature. Content credentials, edit histories, and source-aware interfaces will help audiences distinguish authored, captured, and generated material.
- Creative advantage will shift from raw generation toward curation systems: distinctive archives, expert feedback loops, proprietary context, and exceptional editorial judgment.
- Smaller task-specific models will gain ground where privacy, latency, energy use, and controllability matter more than maximal general capability.
- AI literacy will separate into role-specific crafts. A strategist, engineer, lawyer, filmmaker, and researcher will need different verification habits—not one generic prompting course.
- Premium brands will increasingly advertise meaningful human involvement, making traceable craft and accountable authorship signals of value rather than signs of inefficiency.
Risks
- Confidentiality leakage through prompts, uploaded files, logs, plugins, integrations, or poorly governed retrieval systems.
- False claims and fabricated citations that inherit credibility from polished language or photorealistic presentation.
- Copyright, publicity-rights, trademark, and contractual disputes arising from training sources, stylistic imitation, or output similarity.
- Aesthetic monoculture as teams optimize toward the same model defaults, references, and statistically familiar forms.
- Deskilling when junior practitioners lose opportunities to research, sketch, write, debug, negotiate, and develop judgment firsthand.
- Discrimination in hiring, credit, pricing, moderation, or access when historical patterns are encoded into automated recommendations.
- Vendor dependence, including abrupt price changes, model retirement, policy shifts, outages, and the loss of workflow knowledge.
- Trust erosion when audiences discover undisclosed synthetic content, simulated people, or automation that contradicts a brand's stated values.
Opportunities
- Build provenance-native creative software that records sources, permissions, transformations, model versions, and human interventions by default.
- Create private, rights-cleared models and retrieval systems for archives, museums, material libraries, fashion houses, and research-intensive studios.
- Design AI red-team services specifically for culture and creativity, testing stereotype, style imitation, fabricated references, and brand-risk scenarios.
- Develop interfaces that expose uncertainty, show source passages, compare alternatives, and make rejection easier instead of encouraging one-click acceptance.
- Offer consent, attribution, and compensation infrastructure that lets creators license work for defined model uses on negotiable terms.
- Use AI to widen access through captioning, translation, image description, adaptive interfaces, and alternative formats—with disabled users involved in design and testing.
- Create measurement tools for originality and creative health, tracking reference diversity, revision depth, reviewer burden, and skill development alongside productivity.
| Pressure | Opening | |
|---|---|---|
| #1 | Confidentiality leakage through prompts, uploaded files, logs, plugins, integrations, or poorly governed retrieval systems. | Build provenance-native creative software that records sources, permissions, transformations, model versions, and human interventions by default. |
| #2 | False claims and fabricated citations that inherit credibility from polished language or photorealistic presentation. | Create private, rights-cleared models and retrieval systems for archives, museums, material libraries, fashion houses, and research-intensive studios. |
| #3 | Copyright, publicity-rights, trademark, and contractual disputes arising from training sources, stylistic imitation, or output similarity. | Design AI red-team services specifically for culture and creativity, testing stereotype, style imitation, fabricated references, and brand-risk scenarios. |
| #4 | Aesthetic monoculture as teams optimize toward the same model defaults, references, and statistically familiar forms. | Develop interfaces that expose uncertainty, show source passages, compare alternatives, and make rejection easier instead of encouraging one-click acceptance. |
| #5 | Deskilling when junior practitioners lose opportunities to research, sketch, write, debug, negotiate, and develop judgment firsthand. | Offer consent, attribution, and compensation infrastructure that lets creators license work for defined model uses on negotiable terms. |
For professionals
Use a five-gate review before approval. Gate 1—Purpose: state the exact task, user benefit, baseline, and success metric. Gate 2—Materials: classify every input by confidentiality, consent, copyright, and contractual permission. Gate 3—Authority: assign an accountable owner, qualified reviewer, escalation path, and prohibited uses. Gate 4—Evidence: test accuracy, bias, security, accessibility, output similarity, and user comprehension with representative scenarios. Gate 5—Exit: preserve source files, exportable records, a manual fallback, stop conditions, and a review date. Record the decision in a one-page AI use case card containing provider, model version, data flow, retention terms, known limitations, disclosure language, and incident contact. For high-stakes or public-facing work, involve legal, security, affected communities, and subject-matter experts before launch—not after controversy. The governing question is simple: if this output causes harm or dispute, can the organization explain who decided, what they knew, what they tested, and how a person can obtain correction? If not, the workflow is not ready.
Sources & references
- NIST AI Risk Management Framework (AI RMF 1.0)
- OECD AI Principles
- European Commission: Regulatory Framework for Artificial Intelligence
- U.S. Copyright Office: Copyright and Artificial Intelligence
- UNESCO Recommendation on the Ethics of Artificial Intelligence
- Stanford Institute for Human-Centered AI: AI Index Report
- Coalition for Content Provenance and Authenticity: C2PA Specifications
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 not merely a feature to install. It redistributes judgment, labor, authorship, trust, and advantage. A practical field guide for deciding what should change—and what should remain distinctly human.
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 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
- 185
- Questions per round
- 1.7