New research shows why optimizing content for attention can quietly push it off strategy.
Listen to this article:
When businesses ask AI to write “on brand,” they’re activating far more than a style guide. Brand voice involves a nuanced set of choices about tone, target audience, and what the organization ultimately wants to stand for. LLMs are increasingly being asked to manage those decisions at scale, often without explicit guidance. What if strict adherence to your brand jeopardizes attracting a large audience? Or conversely, what if tasking the AI to gain as many clicks as possible puts your brand voice at risk? In “Balancing Engagement and Polarization: Multi-Objective Alignment of News Content Using LLMs,” a team of researchers including HBS AI Institute associate Elie Ofek tackles this challenge head on by looking at news organizations and media creation. What they found holds implications well beyond journalism: engagement and polarization are linked in language, and optimizing for one while maintaining the other requires a disciplined and technical approach.
Key Insight: Engagement and Polarization Move Together
“[T]he linguistic choices that enhance engagement, such as word choice, tone, or emphasis or omission of specific details, also tend to increase polarization.” [1]
Beginning from the idea that the most engaging stories are often the most divisive, the research team analyzed 13,508 New York Times articles, scoring each for both reader engagement and political polarization using GPT-4o mini. They established a clear positive correlation between how much the writing grabs attention and how polarizing it is, ratings they validated against human annotators. Then the researchers prompted GPT-4o mini to rewrite each article in a “more engaging” style, and found that the new versions were more polarizing as well. The magnitude of this effect was proportional, the more an article’s engagement score increased, the more its polarization score climbed alongside it. Running the experiment in reverse by prompting for less engagement produced a mirror image, with both metrics falling together.
Key Insight: How to Manage Tradeoffs
“MODPO thus allows firms and managers to specify how much they care about different objectives and develop LLM policies precisely aligned with those weights.” [2]
To address this problem, the researchers propose a solution using Multi-Objective Direct Preference Optimization (MODPO). Rather than trying to rely on a clever prompt from an off-the-shelf model, the authors use MODPO to tune the model itself around competing goals in order to optimize multiple objectives more systematically. Organizations that depend heavily on advertising revenue might weight engagement more heavily. Subscription-driven outlets concerned with reader loyalty and trust might weight polarization control more heavily. The implementation has four phases. They start with a pre-trained open-source LLM, Meta’s Llama-3-8B. Then they use supervised fine-tuning to teach the model to generate more engaging versions of articles. Next, they model preferences around polarization. Finally, they incorporate those preferences into language modeling so the resulting model can generate content that improves engagement without moving too far from the original editorial stance. The result is an AI that effectively captures attention while adhering closely to the original editorial spirit of the source material.
Key Insight: The Model Learns Which Language Levers Matter
“[T]he widespread use of LLMs, without safeguards to limit ideological slant, can exacerbate polarization in news content.” [3]
The paper also investigates how MODPO achieves its balance. The authors examine 12 content strategies, including strong openings, narrative structure, emotional language, urgency, provocative language, factuality, and balanced perspective. Their analysis shows that different language features matter for different objectives. Strong openings and narrative structure are especially important for engagement. Provocative language and balanced perspective are important for polarization. The MODPO model appears to learn this distinction, and the authors also tested robustness using alternative evaluators, human raters, a different base model, and a real-world AllSides dataset. Across those checks, the central finding holds: multi-objective alignment can help organizations generate more engaging content while controlling strategic drift.
Why This Matters
For business professionals and leaders, one primary takeaway is that using “off-the-shelf” LLMs may come with unintended consequences. If your marketing, communications, and execution strategies rely on simple prompting, focusing on increasing one strategic goal may come at the cost of another. But the presence of solutions like the MODPO framework demonstrate that these tradeoffs can be managed. What this requires is intentionality, treating AI content alignment as a strategic design problem. In a world of increasing digital noise and declining consumer trust, the ability to be both engaging and “on brand” could be your critical competitive advantage.
Bonus
As this research shows, AI is not a generic performance booster. It must be aligned to the right goals and deployed in the right context. For a look at this issue within the realm of AI service agents, check out The Fast-Talking Chat Agent.
References
[1] Cheng, Mengjie, Elie Ofek, and Hema Yoganarasimhan, “Balancing Engagement and Polarization: Multi-Objective Alignment of News Content Using LLMs,” Harvard Business School Working Paper No. 25-051, Harvard Business School Marketing Unit Working Paper No. 25-051 (April 17, 2025): 5. https://dx.doi.org/10.2139/ssrn.5226629
[2] Cheng et al., “Balancing Engagement and Polarization,” 16.
[3] Cheng et al., “Balancing Engagement and Polarization,” 35.
Meet the Authors
Mengjie (Magie) Cheng is a PhD candidate in Marketing at Harvard Business School.
Elie Ofek is the Malcolm P. McNair Professor of Marketing at Harvard Business School, and an Associate at the HBS AI Institute.
Hema Yoganarasimhan is Professor of Marketing and Michael G. Foster Faculty Fellow at the University of Washington Foster School of Business.
Watch a video version of the Insight Article here.