Prompting-Induced Confirmation Bias

• The paper demonstrates that mathematical models use Bayesian updating and stochastic signal transformations to show how tailored prompts reinforce preexisting beliefs.
• Modelling reveals that even modest bias strength leads to persistent polarization by selectively amplifying self-reinforcing interactions in agent networks.
• These insights inform prompt design in AI and digital platforms, emphasizing balanced messaging to mitigate echo chambers and achieve consensus.

Inducing confirmation bias via prompting involves the deliberate or unintended design of messages, queries, or interactive signals that increase the probability an agent—human or artificial—misperceives, over-weights, or selectively attends to information supporting its existing beliefs. Within both theoretical opinion dynamics and recent LLM research, the technical implementation and impact of this phenomenon have been studied through stochastic models, Bayesian updating with cognitive distortions, network information flow manipulations, experimental LLM prompting, and large-scale simulations. This article reviews and integrates the key concepts, mathematical mechanisms, simulation findings, and practical implications of confirmation bias induction via prompting, with special attention to quantitatively grounded models (Nishi et al., 2013, Gallo et al., 2020), and connects these to contemporary digital communication and computational social systems.

1. Mechanisms of Confirmation Bias Induction

The core mathematical models posit agents holding beliefs about binary (or discrete) opinions, updated via sequential information exchange. Each agent $i$ maintains a belief $Pr(x_i = A)$, updating this upon interaction using a Bayesian rule. Confirmation bias is implemented by modifying the subjective likelihood of received signals: when agents encounter a message or evidence contrary to their prior, they probabilistically misinterpret or underweight it, parameterized by a bias strength $q \in [0, 1]$.

Specifically, in the opinion dynamics model (Nishi et al., 2013), after observing a binary signal $s$ from another agent, the recipient's perception is stochastically altered based on alignment with current beliefs: contrary information is misattributed as supportive with probability $q$, which directly controls the bias strength. The updated beliefs thus incorporate a systematic cognitive distortion at each inference step.

In social network frameworks (Gallo et al., 2020), confirmation bias operates by agents selectively cutting communication links with peers whose opinions diverge by more than a threshold $(1-q)$. The result is a revised adjacency (or mixing) matrix $T^*$, where interactions with disagreeing others are replaced by increased self-reliance, further amplifying the agent's prior.

Key mathematical expressions:

• Signal transformation (for receiver biased toward $A$):

$Pr(\sigma = \alpha \mid s = b) = q$

• Post-update belief (closed form):

$$Pr(x_i = A) = \left( 1 + \left( \frac{1-\theta}{\theta} \right)^{n_{\beta_i} - n_{\alpha_i}} \right)^{-1}$$

• Communication network pruning (social learning):

$$T^*_{ij} = \begin{cases} 0 & \text{if } |x_{i0} - x_{j0}| > 1-q \ T_{ij} & \text{else} \end{cases};\quad T^*_{ii} = T_{ii} + \sum_{j : |x_{i0} - x_{j0}| > 1-q} T_{ij}$$

These mechanisms anchor the mathematical definition of confirmation bias induction at the prompt/message level.

2. Effects on Collective Dynamics and Polarization

Simulation and analytical results reveal dramatic qualitative changes in collective opinion dynamics as the confirmation bias parameter $q$ increases:

• Consensus vs. Polarization: Without confirmation bias ($q=0$), standard majority dynamics or DeGroot-like diffusion ensures eventual consensus—agents agree on a single opinion. As $q$ grows, the system's phase space bifurcates: even modest confirmation bias creates persistent disagreement/polarization, with subgroups sustaining mutually incompatible beliefs indefinitely (Nishi et al., 2013, Gallo et al., 2020).
• Critical Threshold: For small systems ($N=2$), a transition at $q_c = 1 - 1/(2\theta)$ (with $\theta$ the signal reliability) demarcates consensus from disagreement. In larger populations, even small $q$ suffices to generate polarization.
• Role of Signal Reliability: When signals are ambiguous ($\theta$ near $1/2$), the effect of confirmation bias is enhanced, and polarization is more likely.
• Influence Redistribution: In networked models, confirmation bias reallocates centrality, amplifying the influence of self-reinforcing "influencers" and diminishing that of "listeners" who fail to cut links (Gallo et al., 2020).

The table below summarizes outcome regimes as a function of $q$ and $\theta$:

Bias Strength $q$	Signal Reliability $\theta$	Predicted Outcome
0 (unbiased)	High ($\approx1$)	Consensus
Moderate	Moderate	Mixture; risk of splits
High ($\approx1$)	Any	Persistent disagreement
Any	Low ($\approx0.5$)	Polarization prevalent

The emergence of persistent disagreement via prompting (by increasing $q$) thus demonstrates the mathematical plausibility of prompt-induced polarization.

3. Implications for Prompt and Message Design

Translating these theoretical insights, the induction of confirmation bias via prompting (in LLMs, messaging systems, or human-in-the-loop agents) is achievable by manipulating:

• Message Framing: Prompts that rephrase or selectively highlight evidence congruent with the receiver's current stance effectively elevate $q$, increasing the probability that ambiguous or even contrary inputs are assimilated as confirming.
• Ambiguity/Noise Engineering: Presenting signals or evidence in a way that is more ambiguous (reducing $\theta$) makes receivers more susceptible to reinforcing their prior beliefs under confirmation bias.
• Individualization: The peer-to-peer model implies that prompts tailored to recipient history or expressed beliefs maximize bias induction efficiency, compared to global, non-individualized suggestions.
• Self-Loop Amplification: Network-level prompts that encourage attention to one's own prior outputs (e.g., "based on your current thoughts, elaborate further") mathematically mirror the increase in self-reliance, reinforcing bias and slowing consensus development.
• Selective Sampling: Strategically sampling or sequencing prompts to present only self-confirming information—as opposed to balanced or heterodox views—enacts the pruning behavior described in the communication matrix $T^*$.

These findings indicate that both deliberate and inadvertent prompt designs can drive systems into polarized, echo chamber–like regimes.

4. Applications and Consequences in Socio-digital Systems

The modeled mechanisms of confirmation bias induction have direct analogues in real-world digital platforms and AI-driven communication systems:

• Social Media and Digital Messaging: Algorithms prioritizing posts aligning with user beliefs instantiate a high-$q$ communication structure, in effect using individualized prompts to induce and entrench confirmation bias, leading to apparent echo chambers.
• Political Messaging and Shock Elections: Modifying the information network to filter out contrary perspectives, as precisely modeled by link-cutting in (Gallo et al., 2020), slows convergence and enables temporary collective swings (shock elections) that would not occur in unbiased diffusion. The model demonstrates how media strategies emphasizing self-confirming narratives can enhance polarization and unpredictability in outcomes.
• LLMs and Dialogue Systems: Prompts that summarize, present evidence, or generate completions by recapitulating prior user views or failing to challenge priors can systematically reinforce preexisting beliefs (high $q$ induction). This may compromise the objectivity of information-seeking or assistance-focused LLM deployments.

From a systems design standpoint, these insights warn of the risk that even subtle prompt manipulations, absent careful balancing or debiasing strategies, carry significant potential to drive population-level divergence in beliefs.

5. Limitations of Current Models and Research Directions

Prominent limitations acknowledged in the referenced work include:

• No Environmental/External Input: The examined models ignore "ground truth" or exogenous signals; in many real-world systems, such input is available and may act as a bias mitigator.
• Homogeneous Agents: All modeled agents are identical in bias strength $(q)$ and signal reliability $(\theta)$; heterogeneity in either parameter or in network position may qualitatively alter system dynamics and prompt design efficacy.
• Dependence on Initial Conditions: The non-ergodic nature of the system under confirmation bias means that initial distribution of beliefs heavily affects which polarized state emerges, suggesting history sensitivity in real deployments.
• Analytical Solutions Limited: Beyond two-agent cases, closed-form or stability analyses become intractable, necessitating simulation for larger network or population studies.

Research opportunities therefore include:

• Incorporating heterogeneous bias parameters or network roles,
• Integrating dynamic environmental signals to model competing external debiasing,
• Developing scalable simulation or mean-field analytic tools for large-scale dynamics,
• Analyzing prompt-induced bias within richer agent interaction topologies.

6. Connections to Alternative Theoretical and Empirical Approaches

The Bayesian confirmation bias models reviewed here differ from classic global-aggregation models, such as Orlean's, in three ways:

• Opinions are continuous (via accumulated evidence) and not binary,
• Interactions rely on binary local signals, not global opinion snapshots,
• Dynamics are non-ergodic, so population history is not "washed out" in polarized states.

Practically, this suggests that population-level interventions via prompting may need to customize both the informational content and network structure to account for persistent memory effects and feedback loops.

Comparison to other strands of literature reinforces the key conclusion: that the induction of confirmation bias via prompting is both tractable in agent-based models and plausibly mapped onto real-world and digital network settings, with broad implications for polarization and collective outcome unpredictability.

---

In summary, mathematically grounded models and simulations demonstrate that prompting strategies designed to reinforce prior beliefs—by selectively presenting supporting evidence, increasing ambiguity, or amplifying self-referential processing—can induce strong confirmation bias, slow or prevent consensus, and drive populations toward stable disagreement. The critical parameters, $q$ (bias strength) and $\theta$ (signal reliability), as well as network interaction structures, provide a quantitative framework for understanding and predicting the collective impact of biased prompting in both synthetic and engineered systems. Further extensions to heterogeneous, dynamic, and multi-signal environments remain rich areas for future research, with direct application to the design and auditing of AI and social computing interfaces.