Bias Feedback Loops, Dangerous Speech, and AI in Public Discourse
- Dec 17, 2025
- 5 min read
Yes, I believe that public discourse is essential for a free society! But at the risk of offending Aristotelian scholars, everything but moderation is best in moderation!
Though public disagreement is far from new, what is more recent is the speed, scale and selectivity with which modern information systems can intensify disagreement; sometimes into dehumanization; and in rare but catastrophic cases, into real-world violence. In this post I want to explain (1) what "bias feedback loops" are, (2) why "dangerous speech" is a narrower and more analytically useful concept than many everyday uses of "hate speech," and (3) how artificial intelligence can function as both an accelerant and a corrective technology.
1) What a bias feedback loop looks like
A feedback loop is a mechanism in which outputs of a system become inputs that reinforce the system's future behavior. Most people have heard the screeching sounds from microphones being set up by a band, and most everyone knows that it is caused by the speakers and microphones creating a signal loop... called feedback. It can hurt one's ears! In contemporary public discourse, the loop is often driven by attention and engagement:
Content that triggers strong emotions (outrage, fear contempt) tends to generate more engagement.
More engagement causes the content to be shown to more people (or to be more prominent in a feed).
Increase reach produces more engagement, which further increases reach.
This isn't merely a story about "algorithms,' nor is it a story about "people are gullible." It is the interaction: human cognitive tendencies plus platform incentives plus automated ranking and recommendation.
The most important point is conceptual: when a system rewards engagement, it will preferentially amplify content that reliably produces engagement—even if that content is misleading, unfair, or socially corrosive. (The precise magnitude of these effects varies by platform, context, and time.)
Empirically, this picture is consistent with multiple strands of research: for example, large-scale analyses of online diffusion show that false stories can spread faster and farther than truthful ones on social platforms, driven largely by human sharing behavior rather than bots. Science
At the same time, responsible analysis should resist monocausal explanations. Polarization trends do not map cleanly onto internet use in a simple “more internet ⇒ more polarization” way across demographics, which cautions against single-factor narratives. NBER
2) "Dangerous speech" is about risk of violence, not merely offensiveness
The term dangerous speech is commonly defined as expression that increases the risk that an audience will condone or participate in violence against members of another group. Crucially, its danger depends on context; who is speaking, to whom, under what social conditions, not merely on the words in isolation. Dangerous Speech Project linked citation
This framing matters for two reasons:
It focuses attention on a specific harm (heightened risk of group-targeted violence), rather than treating "offense" and "incitement" in the same category.
It encourages risk assessment rather than reflexive censorship: the same phrase can be innocuous in one setting and dangerous in another.
In practice, dangerous speech dynamics can be strengthened by feedback loops: repeated exposure to dehumanizing narratives can normalize contempt; normalized contempt can lower inhibitions; lowered inhibitions can make calls for coercion, or 'justified' violence feel plausible.
Emotionally charged language is a common tactic to generate group attitudes and group think. As an example, comparing someone to a generally defamed or 'evil' individual creates emotional receptiveness or rejection of unrelated ideas depending upon the agenda of the person levying the comparison. The fact that the purpose of using emotionally charged language to support or oppose an argument is an obvious ploy does not seem to negate the effective result, and the comparison will be repeated without thought.
3) The statistical trap: correlation, causation, and the "composition effect"
A recurring failure mode in public argument is drawing causal conclusions about groups from data that cannot support them. This happens in several predictable ways:
Ecological fallacy: inferring individual behavior from aggregate statistics.
Base-rate neglect: focusing on raw counts without denominators (per-capita rates, exposure, opportunity).
Selection and concentration effects: crime, disorder, and social breakdown often concentrate geographically and socially; observers then misattribute concentrated harm to group identity rather than to local conditions, enforcement patterns, or network effects.
A concrete example (without reducing any group to a caricature): if a city has neighborhoods with high poverty, weak informal social control, and entrenched gang disputes, violence can be disproportionately concentrated there. If those neighborhoods are also demographically distinctive (for historical, economic, and housing-policy reasons), it becomes easy—yet analytically wrong—to turn “where the violence is concentrated” into “what kind of people are violent.” The statistical move is illegitimate unless one has causal identification and controls that justify it.
This is not a minor technicality. In polarized environments, the correlation/causation mistake becomes rhetorical fuel: it turns tragedy into propaganda and local pathology into group condemnation, precisely the kind of move that can drift toward dangerous speech.
4) Where AI makes things worse -- and where it might help
AI changes the discourse environment in two asymmetric ways:
AI as an accelerant
Scale: automated tools can generate persuasive text, images, and video at minimal cost.
Plausibility: synthetic media can look “real enough” to bypass ordinary skepticism.
Targeting: micro-tailored messages can exploit local grievances and personal vulnerabilities.
International bodies have warned about deepfakes and the need for detection, verification, and content provenance standards to maintain trust and reduce manipulation risks. reuters.com
AI as a corrective
AI can also be deployed to reduce harm:
Provenance and authenticity tooling (e.g., watermarking, cryptographic provenance, and standardized metadata) to help users distinguish authentic media from fabricated content. AI for Good
Detection systems for coordinated manipulation and synthetic media.
Friction and “accuracy nudges” (design choices that slow sharing, prompt reflection, or foreground source context) that can reduce impulsive amplification without suppressing lawful speech.
Provide citations so readers can fact check and discover nuances to interpretation that are omitted because they do not suit a narrative.
At the governance level, UNESCO has argued for multi-stakeholder platform governance approaches that simultaneously protect freedom of expression and address disinformation, polarization, and incitement to violence. UNESCO
5) Practical recommendations
For readers (and writers)
Treat viral claims as hypotheses, not facts—especially when they flatter your side. Be careful about what you share.
Ask “what would change my mind?” and then actually look for that evidence.
When statistics are involved, insist on denominators, definitions, and scope conditions. The truth is that important details are often glossed over or simply omitted in the preparation of statistics.
Separate “this is offensive” from “this is factual;" they are not the same claim.
Ask if it passes "the smell test," or if it is something you just really want to believe?
For platforms and product teams
Audit ranking and recommendation for downstream harms, not only engagement.
Add friction for high-velocity resharing of low-provenance content.
Invest in provenance standards and clear user-facing authenticity signals. AI for Good
For educators and civic institutions
Teach causal reasoning and statistical literacy as core civic competencies.
Teach the difference between disagreement and dehumanization.
Use the “dangerous speech” lens as a practical bridge between liberty and safety. Dangerous Speech Project
Closing thought
A healthy society does not require unanimity, but it does require restraint: restraint in inference, restraint in speech, and restraint in the incentives we build into our information systems. AI will not solve the moral problem of public discourse for us. But with careful design, transparent governance, and disciplined reasoning, it can help reduce the worst feedback loops rather than intensify them.
This posting is a distillation of what I've written in a longer article with over thirty references to peer reviewed studies on the subject. If you would like a copy of that paper, just use the contact information on this site.
Comments