Examining the Influence of Political Bias on Large Language Model Performance in Stance Classification

Large Language Models (LLMs) serve many natural language processing tasks, consistently surpassing their supervised, neural network predecessors across a suite of benchmark challenges (Touvron et al. 2023). However, the domain of stance detection, defined as identifying an entity’s opinion regarding a particular target (Ng and Carley 2022a), remains relatively uncharted in LLM research. Stance detection is particularly pertinent in analyzing viewpoints on public figures or policies, which are often imbued with significant political implications: most people take stances around politically contentious targets making the task of stance detection intimately tied with politics.

Despite the advancements LLMs have brought to natural language processing, they are not without their shortcomings, including ingrained biases. These biases stem from the vast and varied datasets on which they are trained and can manifest in a variety of contexts biases (Koo et al. 2023; Navigli, Conia, and Ross 2023). These biases also extend to the political context, where LLMs have shown political biases in their responses to political orientation tests (Rozado 2023).

This study delves into the influence of political biases on LLMs’ stance detection capabilities. Our contributions are as follows.

1. 1.

   We evaluated the disparities in how LLMs classify stances with political orientations, considering the dataset, LLMs, and prompting methods. Statistical tests were conducted to illustrate the performance differences between stances favoring the left and the right.

2. 2.

   It was noted that informing the LLM of a statement’s political bias and altering the phrasing of the statement’s target does not reduce performance bias and may even impair performance when the target’s phrasing is more ambiguous.

Automatic stance detection is the identification of an author’s opinion towards a specific target. Stance analysis plays a pivotal role in mapping the dynamics of human opinions and the evolution of viewpoints over time (Chuang 2023; Ng and Carley 2022b). Historically, stance detection has harnessed supervised machine learning techniques, training support vector machines, neural networks, and transfer learning models on manually annotated datasets (Küçük and Can 2020).

With the advent of LLMs, recent research has focused on using LLMs for stance detection. Stance detection is a task requiring a nuanced understanding of context-dependent language, which seems well suited to LLMs. Researchers have experimented with various prompting techniques to elicit the desired output from LLMs like ChatGPT, with mixed results (Deng et al. 2023). Performance on benchmark datasets like SemEval-2016 has varied, with accuracy scores ranging from 0.38 to 0.63 (Chuang 2023; Zhang, Ding, and Jing 2022; Aiyappa et al. 2023). A study by  (Liyanage, Gokani, and Mago 2023) achieved a notable 0.80 accuracy using a zero-shot prompting approach on GPT-4, highlighting the potential of these models. However, investigations into stance classification with non-ChatGPT models remain sparse (Ziems et al. 2023; Cruickshank and Ng 2023).  (Mets et al. 2023) evaluated ChatGPT’s zero-shot classification on a tailored immigration dataset, finding its performance commendable yet trailing behind the best, in-domain supervised models. Overall, while LLMs demonstrate potential for stance classification, the optimal strategies for their application and the scope of their limitations require further exploration.

Investigations into the presence of political biases within LLMs have yielded concerning insights. Studies utilizing political orientation survey questions directed at ChatGPT identified a discernible left and libertarian bias (Motoki, Neto, and Rodrigues 2023; Rozado 2023; Rutinowski et al. 2023). Further research on other LLMs, including GPT-3, demonstrated that LLMs generally exhibit a libertarian political bias, especially when instruction-tuned (Gover 2023; Rozado 2024). However, changing the wording of the prompts in political orientation questions and including content from Wikipedia have shown to possess pronounced biases against political ideologies on both ends of the spectrum (Haller, Aynetdinov, and Akbik 2023; Sasuke and Takemoto 2023). Furthermore, ChatGPT has also been able to reproduce political viewpoints from any given political orientation with similar success across different political orientations (Lihammer 2023). Thus, it is not clear if the political bias present in LLMs affects their ability to reason about politically charged text.

The ramifications of political biases are profound, with the potential to foster divisiveness and tribalism within public discourse (Guo, Ma, and Vosoughi 2022). As LLMs gain ubiquity and begin to support functions such as essay writing in educational and professional settings or classifying the stance of comments in a politically charged debate, the possibility of them being leveraged to polarize opinions and manipulate debates becomes a tangible concern (Haller, Aynetdinov, and Akbik 2023). Building on previous methodologies for gauging political bias, our research aims to ascertain the extent to which these biases influence LLMs’ performance in the often politically-rooted task of stance detection.

To explore the potential political bias in LLMs during stance classification tasks, we implemented a methodology where we classified stances with clear political undertones and assessed the performance across various politically aligned targets. We specifically engaged LLMs to classify several targets known for their distinct political positions within the U.S. political framework, and subsequently analyzed the outcomes to determine any disparities or biases in how stances were classified. This section elaborates on the datasets, models, and prompting strategies employed to examine political bias in LLM stance classification.

We assessed seven LLMs for stance classification and political bias across three different datasets and four prompting schemes. Our findings are twofold, encompassing both the efficacy of LLM stance classification compared to traditional supervised methods and the extent of political bias present in the models’ performance.

The stance classification results, detailed in Table 2, indicate that CoT prompting techniques generally lead to superior model performance. Interestingly, we found marginal differences when comparing the zero-shot CoT with the bias CoT approaches, suggesting that explicitly prompting for bias may not significantly impact the classification outcomes.

When juxtaposed with established baselines, our results showcase a nuanced landscape of performance. While some configurations of LLM + prompt indeed outperform traditional approaches, this is not universally consistent across all tested LLM configurations. For example, Falcon 40B Instruct with CoT prompting emerged as the top performer for the BASIL dataset, which marks an improvement over the baseline (0.65 vs 0.43). Similarly, prompting schemes for UL2 Flan-tuned does result in higher F1 score for the SemEval2016 dataset (0.72 vs 0.56 baseline). For the Elections2016 dataset, the best performing UL2 Flan-tuned scores 0.62, modestly exceeding the 0.55 baseline.

In order to evaluate the possible effects of the political orientations of the stances on the performance of LLMs for stance classification, we performed a number of statistical tests at various levels of aggregation. At the dataset level, we evaluated the difference in micro-F1 score between left and right-leaning stances across all models and prompting schemes for each dataset and for each prompting scheme within each dataset. The t-statistics and p-values are displayed in Table 3 as t-stat/p-val per-dataset and per-prompt. Across the whole of each dataset, the differences in performance between right and left-leaning stances were statistically nonzero, indicating a bias in performance. Within the prompt categories, in each dataset, we found some prompts were not statistically non-zero in their performance. For example, on the SemEval2016 and Election2016 datasets, the task-only and task+context prompts were not statistically nonzero in their performance differences, but were statistically nonzero in their performance differences on the BASIL dataset. Whereas, the reasoning-based prompts (i.e., CoT and Bias CoT) tended to have the opposite performance.

Furthermore, in Figure 1, which displays the density curves of the differences between the left and right-leaning stances, we can see that for Elections2016 and BASIL, the LLMs tended to perform better at identifying left-leaning stances, while in SemEval2016 (i.e., curves are less than 0), they tend to perform better at right-leaning stances (i.e., curve is mostly greater than 0). Thus, the dataset affects the performance of different politically oriented stances when classified by an LLM.

We next analyzed the models, across all datasets and prompting schemes to see if there was a difference in performance for stance detection between the right and left-leaning stances due to different prompts, models, or datasets. From Table 5, we observe that every model shows a statistically nonzero difference in performance between the detection of right and left-leaning stances. The following table, Table 4, displays the average of the differences for each of the models, across all of the prompting schemes and datasets.

From Table 4, we can see that some models have less of a performance difference, on average, than others as well as differences in which political lean they tend to perform better at. For example, Mistral, Falcon-40B, and GPT-3.5 tend to perform better at right-leaning stances. To investigate whether there is actually a difference between the models, we also did pairwise t-tests with Bonferroni correction between each model, across all datasets and prompting schemes, and found that none of the models are statistically different from each other in terms of their performance. This result is also visible in model curves in Figure 1. Thus, all of the models, regardless of architecture or size, have roughly the same bias in their performance on stance classification of politically oriented topics.

Finally, as with the models we analyzed the prompting schemes, across all models and datasets to see if any prompting scheme elicits a difference in performance for stance detection between the right and left-leaning stances. Table 5, all prompting schemes elicit a statistically nonzero difference between the detection of right and left-leaning stances. The following table, Table 6, displays the average of the differences for each of the models, across all of the prompting schemes and datasets.

As with the models, we did pairwise t-tests with Bonferroni correction between each of the prompting schemes, across all of the models and datasets, and found that none of the prompting schemes are statistically distinct from each other. Thus, the prompting schemes exhibit a nonzero difference in performance on different politically oriented stances, and none of the prompting schemes elicits a statistical difference in this result than any other prompting scheme across all of the models and datasets.

In this study, we investigated the performance of Language Models (LLMs) in the task of stance classification, focusing on politically oriented stances. Our findings suggest that LLMs exhibit varying performance when classifying stances related to right- and left-leaning topics. This discrepancy potentially indicates a political bias in LLMs when tackling politically oriented tasks. Such bias could pose societal risks, as LLMs may yield different responses depending on the political orientation of the topic, thereby introducing biases in downstream tasks and outputs.

We explored four different prompting schemes, each providing progressively more contextual information, to assess their impact on performance. Interestingly, we observed that prompts utilizing Chain of Thought (CoT) led to more biased results across datasets compared to non-CoT prompts (i.e., task and task+context). Specifically, non-CoT prompts exhibited non-statistically significant performance differences on tweet-based datasets, while the reverse was true for CoT prompts. This suggests that the type of data (tweets vs. news articles) can heavily influence the effectiveness of particular prompting schemes for politically oriented stance classification.

Furthermore, we found that prompts incorporating the proposed Bias CoT displayed both the only positive mean difference score (i.e. the model was better at classifying right-leaning versus left-leaning stances) and the highest number of nonzero mean difference scores. This suggests that prompting the LLMs to reason about the political bias of statements before classifying stance actually exacerbated political biases in the results. Additionally, no single model consistently outperformed others across all datasets and prompts, underscoring the challenge of stance classification, particularly in the context of politically oriented stances.

Overall, our results reveal that LLMs indeed exhibit political biases, consistent with prior findings indicating that ChatGPT leans more leftward (Motoki, Neto, and Rodrigues 2023). While performance differences between political ideologies are significant at the dataset level, they are less pronounced at the model level and least apparent at the prompt level. This suggests that LLMs tend to demonstrate consistent ideological bias outputs across prompting schemes. The smaller performance differences at the model level imply that most LLMs are trained on similar or overlapping datasets, or that the LLMs are equally biased at the model and prompt level. Conversely, dataset-level differences in political bias highlight the instability of datasets regarding political topics. SemEval2016 and Election2016 were both targeting the same event, but results in different political biases due to the differences in data collection techniques, the datasets are differently biased. This instability aligns with previous studies demonstrating that identical prompting schemes applied to the same dataset yield varying performance (Chuang 2023; Aiyappa et al. 2023).

Therefore, our findings suggest that the primary source of bias in LLM performance stems from the data used for classification. Training data may contribute to political biases in LLM performance, as machine learning algorithms have been known to amplify existing biases in training data (Hovy and Prabhumoye 2021). Additionally, biases may arise from human annotation of datasets, as human annotators may inject their inherent biases into gold stance labels, which are then subject to inconsistent interpretations (Ng and Carley 2022a). Although it remains unclear what aspects of the data trigger politically biased performance differences, our results indicate that such biases are indeed data-driven.

Lastly, our target alterations findings suggests that LLMs are sensitive to how targets are phrased, and any alteration to the target such as misspelling or supplementing with the political party would affect the accuracy of the model. This problem is more unique to LLMs, as most existing methods, especially those for stance classification assume the target to be a static entity, whereas LLMs can accept free text input, meaning the target can be variously phrased. This points towards a future line of research in developing models and prompting schemes that are robust towards such alterations and even possible target alterations (as opposed to prompt alterations) for improving the stance-classification abilities of LLMs.

Stance classification is a crucial task that offers insight into an author’s opinion toward a certain target. Often, these stances will have a political orientation, as individuals tend to hold stances on political topics. In this study, we evaluated the performance of the stance detection task on seven large language models, focusing on how LLMs differ in their performance when confronted with politically oriented stances. Testing four prompting schemes, we found that most LLMs perform stance classification on par or better than baseline models, with the Chain-of-Thought prompting scheme generally demonstrating the best performance.

However, LLMs exhibit significant political biases across datasets, models, and prompting schemes, highlighting their vulnerability to ideological biases. Notably, incorporating LLM reasoning about the bias of the statement and target before classifying the stance did not mitigate the exhibited political bias in the results.

Our study is limited by the series of LLMs and prompts that we were testing, and the strain on our compute resources. The study is also limited by the availability of datasets that are annotated for political stance, and even so, the human annotations may not be consistent both across and within the datasets (Ng and Carley 2022a). Furthermore, we have only considered the U.S. political landscape in this investigation, due to data availability.

Despite these limitations, this study suggests several potential avenues for future research. Firstly, attempts to ameliorate differences in performance between different politically oriented stances through prompting were unsuccessful, suggesting the need for novel prompt patterns to address this issue. Additionally, since LLMs demonstrate different performances for various political orientations in stance detection, similar differences in performance may arise in other politically charged tasks, such as bias classification.

This work was conducted within the Cognitive Security Research Lab at the Army Cyber Institute. The views and conclusions are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Department of Defense, the U.S. Army, or the U.S. Government.

1. 1.

   For most authors…

   1. (a)

      Would answering this research question advance science without violating social contracts, such as violating privacy norms, perpetuating unfair profiling, exacerbating the socio-economic divide, or implying disrespect to societies or cultures? No

   2. (b)

      Do your main claims in the abstract and introduction accurately reflect the paper’s contributions and scope? Yes

   3. (c)

      Do you clarify how the proposed methodological approach is appropriate for the claims made? Yes

   4. (d)

      Do you clarify what are possible artifacts in the data used, given population-specific distributions? Yes

   5. (e)

      Did you describe the limitations of your work? Yes

   6. (f)

      Did you discuss any potential negative societal impacts of your work? Yes

   7. (g)

      Did you discuss any potential misuse of your work? Yes

   8. (h)

      Did you describe steps taken to prevent or mitigate potential negative outcomes of the research, such as data and model documentation, data anonymization, responsible release, access control, and the reproducibility of findings? Yes

   9. (i)

      Have you read the ethics review guidelines and ensured that your paper conforms to them? Yes

2. 2.

   Additionally, if your study involves hypotheses testing…

   1. (a)

      Did you clearly state the assumptions underlying all theoretical results? NA

   2. (b)

      Have you provided justifications for all theoretical results? NA

   3. (c)

      Did you discuss competing hypotheses or theories that might challenge or complement your theoretical results? NA

   4. (d)

      Have you considered alternative mechanisms or explanations that might account for the same outcomes observed in your study? NA

   5. (e)

      Did you address potential biases or limitations in your theoretical framework? NA

   6. (f)

      Have you related your theoretical results to the existing literature in social science? NA

   7. (g)

      Did you discuss the implications of your theoretical results for policy, practice, or further research in the social science domain? NA

3. 3.

   Additionally, if you are including theoretical proofs…

   1. (a)

      Did you state the full set of assumptions of all theoretical results? NA

   2. (b)

      Did you include complete proofs of all theoretical results? NA

4. 4.

   Additionally, if you ran machine learning experiments…

   1. (a)

      Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? Yes, it is uploaded as a supplementary material, the URL will be public when the paper is accepted

   2. (b)

      Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? Yes

   3. (c)

      Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? Yes

   4. (d)

      Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? Yes

   5. (e)

      Do you justify how the proposed evaluation is sufficient and appropriate to the claims made? Yes

   6. (f)

      Do you discuss what is “the cost“ of misclassification and fault (in)tolerance? Answer

5. 5.

   Additionally, if you are using existing assets (e.g., code, data, models) or curating/releasing new assets, without compromising anonymity…

   1. (a)

      If your work uses existing assets, did you cite the creators? Yes

   2. (b)

      Did you mention the license of the assets? Yes

   3. (c)

      Did you work any new assets in the supplemental material or as a URL? Yes

   4. (d)

      Did you discuss whether and how consent was obtained from people whose data you’re using/curating? NA

   5. (e)

      Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? Yes

   6. (f)

      If you are curating or releasing new datasets, did you discuss how you intend to make your datasets FAIR (see fair)? Yes

   7. (g)

      If you are curating or releasing new datasets, did you create a Datasheet for the Dataset (see gebru2021datasheets)? Yes

6. 6.

   Additionally, if you used crowdsourcing or conducted research with human subjects, without compromising anonymity…

   1. (a)

      Did you include the full text of instructions given to participants and screenshots? NA

   2. (b)

      Did you describe any potential participant risks, with mentions of Institutional Review Board (IRB) approvals? NA

   3. (c)

      Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? NA

   4. (d)

      Did you discuss how data is stored, shared, and deidentified? NA