What happens when the ad in your feed starts to use your favorite colors, your slang, even your tempo? For some brands, that�s no longer a hypothetical? Agencies are deploying large language models (LLMs) to infer personality traits from public posts and then tailor the copy, music, and visuals of each ad�one impression, one creative?
From segments to individuals
Cheil UK, working with startup Spotlight, is testing campaigns that mirror how people talk and the aesthetics they favor, based on their public online activity? It�s a shift from demographic targeting to psychological profiling, with ads adapting for introverts versus extroverts, or for calmer versus louder music preferences?
The pitch is measurable performance? A Northwestern University study led by marketing professor Jacob Teeny found that LLM-written iPhone ads tailored to a person�s personality were more persuasive than non-personalized versions? �All roads point to the fact that this will become the way [digital advertising is done],� Teeny said? And if AI can cut the roughly 15% of digital media spend that goes unseen or ignored, the ROI case gets stronger?
�Creepy slop� or creative breakthrough?
Not everyone is buying the promise of infinite micro-variants? �Congratulations�your AI just spent a fortune creating an ad only one person will ever see, and they�ve already forgotten it,� said Alex Calder of consultancy Jagged Edge? Brand strategist Ivan Mato warns that surveillance-driven personalization risks backlash: the question isn�t whether brands can personalize everything�it�s whether they should?
There are recent cautionary tales? As Microsoft touts Windows as an �agentic OS� that blends devices, cloud, and AI, user responses have been harsh? Former Microsoft engineer Dave Plummer argues that �ads, nags, and suggestions� inside the OS erode trust: �When the Start menu shows sponsored apps, you put a price on my attention on my machine?� His takeaway for platforms�and by extension, advertisers: �Trust is more valuable than any click-through metric?�
Data pipelines under pressure
Hyper-personalization depends on data�lots of it? Cheil describes models reading what people post on public platforms and layering insights over location, age, and purchase history? Some vendors also claim access to search activity and even prompts typed into AI tools via consented integrations, though that�s contingent on explicit user permissions and partnerships?
But the free buffet of web data is closing? The Wikimedia Foundation has urged AI developers to stop scraping Wikipedia and use its paid Enterprise API instead, noting that bots attempting to evade detection contributed to unusual traffic spikes while human pageviews fell 8% year over year? The direction of travel is clear: provenance, licensing, and attribution? For marketers, that means evaluating adtech partners� data sources and anticipating higher data costs that could flow through to CPMs?
The hidden cost: compute and energy
There�s also the physics of personalization at scale? Generating thousands of unique creatives on the fly pushes inference workloads up? The biggest bottleneck for AI isn�t money�it�s energy, according to analysis by the Financial Times and MIT Technology Review? Data centers are straining grids in some markets, and forecasts for power needs over five years range from less than 2x to 4x today�s demand?
One pragmatic solution: flexibility? A Duke University study cited in that analysis found that if data centers curtailed consumption about 0?25% of the time�roughly 22 hours a year�the grid could support 76GW of new demand? That matters for ad platforms deciding between real-time generation versus pre-rendered dynamic creative optimization: the compute and energy tab will increasingly shape media economics?
How to personalize without crossing the line
- Use consented first-party data and licensed sources; ask vendors to attest to provenance and opt-in rates?
- Cap depth: target mood, need state, or benefit�but avoid mirroring intimate traits that can feel invasive?
- Test lift versus cost: compare micro-variants against strong �big idea� creative across reach and memory metrics, not just CTR?
- Set creative guardrails so every variant still looks and sounds like your brand?
- Be energy-aware: pre-generate where possible and schedule heavy AI workloads during green or off-peak hours?
- Be transparent: if content is AI-generated or personalized, say so simply? Trust compounds; creepiness compounds faster?
AI can make ads feel like a service, not a stalker? But relevance turns into �creepy slop� the moment people sense overreach or manipulation? The next advantage in adtech won�t just be better targeting�it will be provably ethical data, efficient compute, and experiences that respect the audience�s attention?