What breaks in production

Creative sameness: Models reproduce common patterns. If every team uses the same prompt language, references, templates, and optimization target, the category converges.
Loss of originality: An output may look novel to the team while being statistically familiar to the model. A human point of view requires selection, transformation, and an argument.
Generic brand voice: Language models can produce fluent copy that is indistinguishable from every other brand’s competent copy.
Weak emotional judgment: Models can describe tension, grief, humor, or desire without necessarily understanding the ethical or cultural weight of using them.
Bias and stereotype: Generated casting, beauty, family, profession, geography, or behavior may default to narrow stereotypes.
Cultural misrepresentation: A “global” visual can collapse local distinctions or borrow cultural signs without context.

Incorrect products: Packaging, dimensions, ingredients, controls, logos, and product behavior may be wrong.
Unreliable text: Text embedded in images can be misspelled or fabricated; legal disclaimers are particularly unsafe to generate and trust.
Inconsistent characters and environments: Face, body, wardrobe, scale, geography, physics, and lighting can drift across shots.
Over-polished wrongness: Stakeholders may accept an attractive frame before checking whether it tells the truth.
Hidden correction cost: A cheap first pass can create expensive cleanup, compositing, reshoots, voice replacement, legal review, or stakeholder churn.

Copyright and training-data uncertainty: A provider’s terms may give the customer rights against the provider while not guaranteeing that an output is free of third-party claims.
Likeness and voice: A face, voice, gesture, or digital double is an identity asset. Consent needs to specify duration, territory, media, edit rights, compensation, training, reuse, and revocation.
Deepfakes and deceptive endorsements: Viewers should not be led to believe a person said, did, experienced, or endorses something they did not.
Privacy and biometric data: Face, voice, browsing behavior, health, precise location, and inferred traits can be sensitive. Minimize, document, and delete where appropriate.
Consumer distrust: If “real” and synthetic content look identical, audiences may distrust both; disclosure is a trust mechanism, not a punishment.
Vendor lock-in and model drift: A workflow may depend on a model, plan, watermark, credit price, API, or safety setting that changes.
Labor displacement and devaluing craft: If organizations treat craft roles as interchangeable prompt operators, they may lose the expertise that makes the work reliable and distinctive.
Unclear authorship: The U.S. Copyright Office says AI-assisted work can still be copyrightable where a human determines sufficient expressive elements, but mere prompting is not enough by itself; the law varies by jurisdiction.14

Prefer AI when…	Prefer traditional or human-led methods when…
You need to explore many low-fidelity directions before choosing one	The idea depends on lived experience, subtle performance, or a real person’s trust
The task is repetitive, structured, and easy to inspect	The task is ambiguous, high-stakes, culturally specific, or legally sensitive
The asset is a draft, temp, reference, or controlled variant	Product truth, safety, medical/financial claims, or physical feasibility matters
You have clear source assets, rights, and constraints	You cannot explain what data or references entered the system
You can measure the result with a controlled test	The platform’s optimization target is a weak proxy for the business goal
A human owns the final decision and the audit trail	The proposed system has no owner, rollback, disclosure, or escalation path

Follow the evidence

This chapter comes from the September research notebook. Linked sources and qualifications remain attached to the claims.