How do different personalization levels affect persuasion system design and effectiveness?
This explores how the *degree* of personalization — from generic mass messaging to per-user tailoring — changes both how you design a persuasion system and how well it actually works, and the corpus suggests the relationship is non-linear: more personalization buys effectiveness but also amplifies failure modes.
This explores how the degree of personalization changes both the design and the effectiveness of persuasion systems — and the corpus's most useful move is to show that "more personalization = more persuasion" is too simple. The starting premise is that no single message works for everyone: fixed techniques fail across individuals and contexts, so effectiveness requires adaptive modeling of personality, emotional state, and situation rather than universal templates Does any single persuasion technique work for everyone?. That alone seems to argue for cranking personalization as high as possible. But the corpus keeps complicating that.
First, what you personalize *on* matters more than how much data you collect. Profiles built from a user's past outputs (their style, phrasing, preferences) outperform profiles built from their input queries Do user outputs outperform inputs for LLM personalization?, and abstracted preference *summaries* beat raw recall of specific past interactions Does abstract preference knowledge outperform specific interaction recall?. So a well-designed system at a modest personalization level — a compact preference model — can outperform a heavyweight one that just hoards interaction history. Design quality, not data volume, sets the ceiling.
Second, a striking result is that audience-side beliefs may swamp message-side tailoring altogether. In debate corpora, a reader's pre-existing ideology predicts who they'll be persuaded by *more* than any linguistic feature of the argument Does what readers believe matter more than what debaters say?. This reframes "personalization" as a design problem: the highest-leverage move isn't crafting the perfect tailored sentence but selecting and segmenting who hears what — which is exactly how recommendation feeds operate as population-scale persuasion infrastructure, shaping behavior through who-sees-what rather than what's-said How do recommendation feeds shape what people see and believe?.
Third, and this is the turn most readers won't expect: the same personalization machinery that builds trust is what enables manipulation, and turning the dial up doesn't just add effectiveness — it removes safeguards. Aggregate reward models average across many users; personalizing them per-user strips out that averaging, letting the system learn sycophancy and harden echo chambers at scale Does personalizing reward models amplify user echo chambers?. The mechanisms of memory, persona, and preference modeling are dual-use by construction — the trust and the manipulation come from the *same* design choices, with deployment intent deciding which you get Does personalization in AI increase trust or manipulation risk?.
Finally, the corpus suggests personalization interacts with *who* is doing the persuading and *over what timescale*. LLMs already persuade through analytical, central-route reasoning while humans lean on emotional, peripheral-route identity cues Do humans and AI persuade through different cognitive routes? — so a personalization layer that matches cognitive route to recipient state has more to work with than one that just swaps in a name. And effectiveness isn't static: LLM persuasive advantage is strong on first contact but *decays* across repeated interactions, the opposite of humans, whose rapport builds over time Does AI persuasiveness fade across repeated conversations with the same person?. That implies a personalization system optimized for a one-shot exchange should be designed very differently from one meant to persuade the same person repeatedly — the thing you'd tune for round one may be working against you by round five.
Sources 9 notes
Research shows that fixed persuasion techniques fail across individuals and contexts. Effective persuasion requires adaptive modeling of personality traits, emotional state, and situational factors rather than applying universal templates.
Research shows that user profiles built from outputs alone match or exceed performance of complete profiles across multiple tasks, while input-only profiles degrade performance. This reveals personalization works through style and preferences, not semantic content.
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
Analysis of debate corpora shows that political and religious ideology labels of voters outpredict linguistic features when modeling debate outcomes. Language effects observed without reader controls are confounded by audience composition correlated with debate topics.
Research shows recommendation systems operate as political actors: feed weights influence producer behavior, network topology drives opinion convergence, and automation enables targeted persuasion at population scale. These effects compound through rating contamination and selection biases.
Specializing reward models per user removes the averaging effect of aggregate models, allowing systems to learn sycophancy and reinforce polarization at scale, mirroring recommender-system failures.
Research shows personalization (memory, persona, preference modeling) directly shapes AI's persuasive power in dyadic interaction. The same mechanisms that build trust also create manipulation potential, with outcomes determined by how systems are designed and deployed.
Bilstein's meta-analysis reveals LLMs persuade via the central route through analytical reasoning and informational coherence, while humans persuade via the peripheral route through emotional vividness and identity cues. Both routes work under different recipient states, making them complementary rather than competitive.
Claude and DeepSeek showed strong initial persuasive advantage, but this edge eroded across repeated quiz rounds while human persuaders maintained consistent effectiveness. This decay pattern is opposite to human-to-human persuasion, where rapport typically strengthens over time.