What “AI in advertising” includes
| Layer | What teams use it for | Typical human responsibility | Typical failure |
|---|---|---|---|
| Brand strategy | Mining language, mapping competitors, exploring positioning, stress-testing a proposition | Choose the cultural or commercial problem; validate with actual research | Synthetic consensus, shallow strategy, or invented evidence |
| Consumer research | Summarizing interviews, clustering open text, finding themes, modeling questions | Own sampling, interpretation, privacy, and what cannot be inferred | Bias in the source set or over-reading weak signals |
| Audience analysis | Segmentation, lookalikes, propensity, media allocation, intent signals | Define legitimate use, exclusions, and fairness constraints | Proxy discrimination, over-targeting, and opaque optimization |
| Creative concepting | Routes, headlines, naming, moodboards, visual references, story prompts | Set the brief, taste, point of view, and kill criteria | Idea sameness, style imitation, and endless options without a decision |
| Copywriting | Outlines, variants, localization drafts, SEO, captions, scripts, CTAs | Verify claims, tone, law, voice, and cultural meaning | Hallucinated claims, generic voice, accidental plagiarism |
| Image creation | Concepts, backgrounds, product worlds, retouching, resizing, composites | Lock the product, likeness, brand system, and final craft | Wrong product geometry, hands, text, logos, or unlicensed similarity |
| Video production | Storyboards, previsualization, image-to-video, motion tests, editing, cleanup | Direct the story, camera, performance, continuity, and final cut | Inconsistent characters, impossible physics, uncanny motion, unclear provenance |
| Music and sound | Temp tracks, sound effects, adaptive music, cleanup, dubbing, voiceover | Clear composition, performer/voice rights, emotion, and mix | Voice or music similarity, blandness, or misleading synthetic performance |
| Digital humans and characters | Spokespeople, avatars, virtual influencers, interactive agents | Obtain consent, define persona, moderate behavior, disclose the fiction | Impersonation, deceptive endorsement, parasocial manipulation |
| Product visualization and 3D | Product-in-context scenes, pack shots, virtual sets, 3D variations | Verify materials, dimensions, claims, and manufacturability | A visually attractive product that does not exist |
| Personalization | Dynamic creative, offer, copy, product, voice, or location variants | Set lawful rules, frequency caps, data minimization, and approval | Creepy inference, discrimination, privacy breach, ad fatigue |
| Localization | Translation, dubbing, lip-sync, culturally adapted layouts and scripts | Native review, local legal review, talent/voice clearance | Literal translation, cultural error, or global-scale propagation of one mistake |
| Media buying | Bid optimization, placement, creative rotation, audience expansion, budget pacing | Define objectives, exclusions, incrementality, and stop conditions | Optimizing to cheap clicks, opacity, or spending drift |
| Customer engagement | Chatbots, shopping assistants, lead qualification, service agents | Grounding, escalation, disclosure, logging, and claims policy | Confidently wrong answers or unreviewed commitments |
| Measurement | Creative scoring, attribution, experiment design, sentiment, fatigue detection | Choose meaningful outcomes and counterfactuals | Mistaking correlation, platform attribution, or click volume for value |
Definitions that should not be collapsed
- AI-assisted advertising: A human-led advertisement where AI helps at one or more stages, such as research, rough concepting, cleanup, translation, editing, or versioning.
- AI-generated advertising: A campaign or asset whose material output was generated by a model. This does not imply that the concept, selection, edit, or production was automated.
- AI-personalized advertising: Creative or offers vary based on audience, location, behavior, product feed, or user input. The personalization may use machine learning without generating the image or copy.
- Generative advertising: A broad category in which a model creates or transforms copy, image, video, audio, code, or interactive responses.
- Automated media buying: Algorithmic selection of bids, inventory, audiences, placements, or budget allocation. It is not the same as generative creative.
- AI-created brand assets: Logos, packaging, product images, characters, voices, music, or design components generated or materially altered with AI. Rights and authorship should be documented asset by asset.
- AI-generated spokesperson: A synthetic or digital human presented as a voice, face, or character associated with a message. A stock avatar is not automatically a paid-ad spokesperson; some providers restrict that use.12
- Fully synthetic campaign: A campaign in which the concept may be human-authored but the visible, audible, and interactive outputs are predominantly generated or simulated. This is the highest-disclosure and provenance burden, not a quality category.
A simple landscape map
Human strategy / cultural insight / offer
↓
Creative direction + constraints
↓
Models and tools: text · image · video · audio · 3D
↓
Edit · verify · clear rights · localize · label
↓
Media delivery · personalization · conversation · measurement
↓
Learn, document, refresh, or stop the campaign
The “verify” step is not a small box at the end. It should recur whenever the output changes: new model, new market, new likeness, new claim, new product version, new audience rule, or new distribution channel.
Follow the evidence
This chapter comes from the September research notebook. Linked sources and qualifications remain attached to the claims.
Open all sources and footnotes →