In the public imagination, AI manipulation usually arrives wearing a mask: a deepfake video that looks like a real interview, a synthetic voice that sounds like a trusted official, or a bot swarm that floods timelines with convincing lies. Those threats are real, and they’re getting the most attention because they’re easy to visualize. But there’s another, quieter channel of influence that doesn’t require a fake face or a forged recording. It uses something far more ordinary: language.
Large language models—LLMs—can produce persuasive text at scale. They can write comments that sound like they come from a neighbor, draft talking points for a campaign, generate “explainer” threads that frame an event in a particular way, and tailor messages to different audiences. And when those outputs reflect bias—whether from the data used to train the model, the way the model learns patterns of persuasion, or the assumptions embedded in its design—the result can be a subtle steering of belief and emotion. Not always through outright falsehoods, but through emphasis, omission, tone, and framing.
This is where the phrase “hacking your emotions” becomes more than a metaphor. Emotions are not only triggered by what people see; they’re also shaped by how information is presented. Language models can influence both. They can nudge readers toward anger, fear, contempt, hope, or trust by selecting which details to foreground, which interpretations to suggest, and which moral cues to activate. The manipulation may be less dramatic than a deepfake, but it can be equally effective—especially when it blends into the everyday noise of social media.
What makes this risk distinct is that it doesn’t rely on a single spectacular lie. Instead, it exploits the mechanics of attention. Online platforms reward engagement, and engagement often correlates with emotional intensity. LLMs are particularly good at producing content that is readable, timely, and emotionally resonant. Even when the model isn’t instructed to deceive, it can still generate messaging that feels compelling because it mirrors the rhetorical patterns that already work on humans.
To understand how biased LLMs could influence public opinion, it helps to break down the process into steps—because the danger isn’t just “the model is biased.” The danger is what happens when bias meets distribution.
First comes generation. An LLM can be prompted to produce a narrative about an event. If the model has learned from historical text that certain groups are associated with certain traits, or if it tends to use particular stereotypes in its explanations, the output may carry those biases forward. Even without explicit slurs or obvious prejudice, bias can appear as selective framing: emphasizing alleged incompetence rather than systemic constraints, highlighting conflict rather than context, or treating uncertainty as if it were settled.
Second comes tailoring. A skilled operator can ask the same model to rewrite the message for different audiences. One version might emphasize economic harm to appeal to voters concerned about jobs. Another might emphasize cultural threat to appeal to those worried about identity. Another might emphasize personal safety to appeal to those primed by fear. The content can remain “plausible” while shifting its emotional target. This is not necessarily a fabrication of facts; it can be a reweighting of interpretation.
Third comes amplification. Social platforms don’t distribute content neutrally. They rank what users are likely to engage with. If LLM-generated posts are optimized—directly or indirectly—for emotional resonance, they can spread faster. And once repeated across accounts, the narrative gains a second layer of credibility: repetition. People often infer truth from familiarity, especially under time pressure. When the same framing appears again and again, it can feel like consensus.
Fourth comes reinforcement. Even if individuals don’t believe every claim, they may absorb the emotional posture. A reader might reject a specific statistic but still adopt the underlying sentiment: “Something is wrong,” “They’re hiding the truth,” “We’re being attacked,” “The system is rigged.” Over time, these sentiments can harden into identity-level beliefs. That’s the “emotional hacking” part: not just changing what people think, but shaping how they feel about what they think.
There’s also a fifth step that’s easy to overlook: interaction. LLMs can be used not only to publish content but to converse. Imagine a chatbot that responds to a user’s questions with tailored explanations. If the model’s responses are biased, the user’s worldview can be nudged through dialogue. The user asks, the model answers, and the answer subtly steers the next question. This creates a feedback loop where the model doesn’t merely broadcast a message—it co-constructs a belief path.
So how much of this is happening already? We can’t know precisely without access to internal systems and large-scale measurement. But we do know the capabilities are here. LLMs can generate fluent text quickly, in multiple styles, and in many languages. They can mimic the cadence of news commentary, the structure of policy analysis, or the tone of personal testimony. They can also produce variations—hundreds of near-duplicates—making it harder for moderation systems to detect coordinated inauthentic behavior based solely on identical text.
And crucially, bias doesn’t have to be intentional. Many LLMs are trained to be helpful, safe, and aligned with certain norms. That alignment can still produce systematic distortions. For example, a model might default to certain political assumptions, treat some forms of skepticism as irrational, or frame complex issues with simplified moral binaries. In a vacuum, that might be annoying. In a high-stakes information environment, it can become influential.
Consider how public opinion is formed. It’s rarely built from a single piece of evidence. It’s assembled from fragments: headlines, summaries, commentary, and the emotional tone of the surrounding conversation. LLM-generated content can affect each fragment. A biased model might consistently describe one side as “reckless” and the other as “cautious,” even when the underlying facts are similar. It might repeatedly highlight worst-case scenarios for one group and best-case possibilities for another. It might use language that implies intent—“they planned this”—when the evidence is ambiguous. Over time, these micro-choices accumulate into a macro-narrative.
This is why researchers and observers are increasingly focused on “framing bias,” not just factual errors. Framing bias is about the interpretive lens. Two articles can report the same event but lead readers to different conclusions by emphasizing different causes, different stakes, and different moral responsibilities. LLMs are powerful at framing because they excel at summarization and explanation. They can compress complexity into a story—and stories are where emotions live.
Another concern is targeting. Public opinion isn’t uniform; different communities respond to different cues. LLMs can help operators craft messages that match the vocabulary, values, and anxieties of specific groups. This can happen through explicit targeting—choosing which audience sees which post—or through implicit targeting, where the model generates content that naturally resonates with certain demographics based on learned patterns. Even if the operator doesn’t specify a group, the model might produce content that aligns with stereotypes present in training data. That alignment can then be exploited by whoever distributes the content.
Targeting also extends beyond politics. Health misinformation, for instance, often spreads through emotionally charged narratives: fear of side effects, hope of quick cures, distrust of institutions. LLMs can generate persuasive health explanations, sometimes with plausible-sounding caveats. Even when the model includes uncertainty, the overall tone can still push users toward a particular action. In crisis situations, that can be dangerous.
Then there’s the question of amplification and repetition. When LLMs generate content at scale, they can flood feeds with consistent emotional cues. This can create a perception of momentum: “Everyone is talking about this,” “People are waking up,” “The tide is turning.” Perception of consensus is itself an emotional driver. People want to belong to the “right” side, and they want to avoid social costs. If LLM-generated narratives create the impression of widespread agreement, they can accelerate polarization.
But the most unsettling aspect is that LLMs can be used to make manipulation look normal. Deepfakes are unusual; they stand out. Biased LLM text can blend in because it resembles the kind of commentary people already write. It can cite sources, summarize arguments, and use the rhetorical style of legitimate analysis. That means detection is harder. It’s not enough to ask, “Is this video real?” You also have to ask, “Is this framing designed to move me emotionally?”
So what should be done? The answer isn’t a single silver bullet. It’s a layered approach that combines technical safeguards, transparency, and social resilience.
One area of focus is evaluation. Researchers are building tests to measure bias in generation and framing. Instead of only checking whether the model produces false statements, they examine whether it systematically favors certain interpretations, uses loaded language, or treats groups differently. This is challenging because bias can be subtle and context-dependent. A model might behave fairly in one scenario and skew in another. Still, evaluation is essential because it turns a vague concern into measurable properties.
Another area is provenance and transparency. If content is generated or heavily assisted by AI, labeling can help users calibrate trust. Provenance systems—ways to cryptographically track origin—are being explored for media authenticity, but the same logic can extend to text. If platforms can identify AI-assisted content, they can apply different ranking rules, attach warnings, or reduce the reach of suspicious material. Transparency doesn’t eliminate manipulation, but it changes the incentives and gives users more information to judge credibility.
A third area is platform governance. Moderation systems need to detect not only explicit misinformation but also coordinated inauthentic behavior and emotional manipulation patterns. That includes monitoring for rapid production of similar narratives, unusual posting patterns, and networks of accounts that amplify each other. Importantly, governance must be careful not to over-censor legitimate debate. The goal is not to suppress viewpoints; it’s to reduce deception and manipulation.
A fourth area is model design.