How Bias Amplifies Across AI Generations
Table of Links
-
Abstract and Introduction
-
Background and Related Work
-
Theoretical Framework
-
Experiment Design
-
Results
-
Discussion
-
Limitations
-
Ethical Considerations and References
A. Mathematical Formulation of WMLE
B Fine-tuning Setup
C Qualitative Bias Analysis Framework and Example of Bias Amplification Across Generations
D Distribution of Text Quality Index Across Generations
E Average Perplexity Across Generations
F Example of Quality Deterioration Across Generations
G Pearson Correlation Between Neuron Weight and Bias Performance | H Pearson Correlation Between Neuron Weight and Generation Quality | I Pearson Correlation Between Neuron Activation and Bias Performance | J Pearson Correlation Between Neuron Activation and Generation Quality | K Mathematical Details for the Statistical Tests | L Literature Review of Model Collapse
A Mathematical Formulation of WMLE
In Weighted Maximum Likelihood Estimation, we maximize the weighted log-likelihood:
B Fine-tuning Setup
C Qualitative Bias Analysis Framework and Example of Bias Amplification Across Generations
We employed qualitative methods to confirm our findings in media bias. Specifically, we utilized a media bias identification framework grounded in foundational works such as Entman’s framing theory (Entman, 1993) and other research on media bias detection (Rodrigo-Ginés et al., 2024; Groeling, 2013). This framework provides a robust lens to evaluate political biases in the framing and language use of media texts (explained below). Given the nature of our data—text exclusive of visual or contextual cues like formatting—certain types of media bias commonly seen in formatted articles or televised programs (e.g., visual bias or tone) may not apply. Therefore, our focus was on the two key aspects of political bias that are particularly relevant in textual analysis:
Story Framing and Selection Bias: This type of bias emerges when inherent leanings are found in the way topics, arguments, or narratives are structured. For instance, some aspects of reality are highlighted while others are obscured, shaping how the audience understands and interprets the events or issues at hand (Entman, 1993; Groeling, 2013). In extreme cases, opposing viewpoints are entirely excluded, leading to a one-sided representation of the issue. This selective omission restricts the audience’s comprehension of the full spectrum of perspectives, resulting in a distorted portrayal of the issue (Rodrigo-Ginés et al., 2024; Groeling, 2013). Entman described this as the selection and salience of specific facts that promote particular definitions, evaluations, and recommendations.
Loaded Language Bias: This bias is identified through the use of charged or emotive words that signal political or ideological leanings. A common example is the difference in connotation between terms such as “undocumented” versus “illegal” immigrants. Such language choices often shape the audience’s perception by evoking specific emotional responses (Rodrigo-Ginés et al., 2024; Groeling, 2013).
Below is an example of GPT-2 text outputs influenced by iterative synthetic training. The original article, titled “First Read: Why It’s So Hard for Trump to Retreat on Immigration, is a political opinion piece from NBC News, a left-leaning outlet as rated by AllSides (NBC News, 2016; AllSides, 2024b). The analysis follows the qualitative framework.
• Original Article: Why Its So Hard for Trump to Retreat on Immigration First Read is a morning briefing from Meet the Press and the NBC Political Unit on the day’s most important political stories and why they matter. Why its so hard for Trump to retreat on immigration Since launching his presidential candidacy 14 months ago, Donald Trumps most consistent and uncompromising policy issue has been immigration. Indeed, it was the subject of his first general-election TV ad that started airing on Friday. Yet over the weekend, his top aides and advisers suggested that Trump might be shifting on his past position that all of the 11 million undocumented immigrants living in the United States must be deported forcibly. To be determined, is what newly minted Campaign Manager Kellyanne Conway said on CNN when asked if Trump was retreating on the deportation force he talked about during the primary season. But here’s why its so hard – if not impossible – for Trump to retreat on immigration: Hes caught between his clear, unambiguous past statements and a base that might not willing to see him moderate on the issue. His past statements: Aug. 16, 2015 “”We’re going to keep the families together, but they have to go,”” Trump said on NBCs Meet the Press. More Trump: “”We will work with them. They have to go. Chuck, we either have a country, or we don’t have a country,”” he said. Nov. 11, 2015 You are going to have a deportation force, and you are going to do it humanely, Trump said on MSNBCs Morning Joe when asked how he would round up the nations 11 million undocumented immigrants. April 21, 2016 Look, were either going to have a country or were not going to have a country. But many people are very fine people. And I’m sure these are very, very fine people. They’re going to go, and were going to create a path where we can get them into this country legally, okay? But it has to be done legally – when asked by a questioner at a Today town hall that persons undocumented relatives would have to be deported if Trump becomes president. Trump cant ignore a base that has cheered his uncompromising immigration position And then there are the Trump supporters who’ve cheered the GOP presidential nominee for being so uncompromising on immigration. Classification Probability: 0.9946 for left-leaning, 0.0051 for center-leaning, 0.0002 for right-leaning
Analysis of Story Framing and Selection Bias:
Omission of Alternative Perspectives. There is a lack of discussion about the potential benefits of Trump’s immigration policies or the concerns that might have led to his hardline approach. The article does not present viewpoints from supporters who agree with his stance or any arguments that justify his policies. This omission restricts the audience’s understanding of the full range of perspectives on the issue
Highlighting the political objectives behind Trump’s immigration policies. The emphasis is placed on Trump’s electoral interests rather than presenting the policy as being for the benefit of the people. For example, “Trump can’t ignore a base that has cheered his uncompromising immigration position” and “And then there are the Trump supporters who’ve cheered the GOP presidential nominee for being so uncompromising on immigration.” Additionally, “But here’s why it’s so hard — if not impossible — for Trump to retreat on immigration: He’s caught between his clear, unambiguous past statements and a base that might not be willing to see him moderate on the issue.”
Analysis of Loaded Language Bia:
Differences in Connotation. The article uses the term “undocumented immigrants” instead of “illegal immigrants,” a shift that appeared in later generations’ articles. This choice of wording can shape perceptions and evoke specific emotional responses, aligning with a particular ideological perspective.
• Generation 0: Why Its So Hard for Trump to Retreat on Immigration The president’s decision to withdraw from the Paris climate accord is a major blow to the United States’ standing in the world. The United States has been a major player in the world of climate change, and has been a major player in the fight against climate change. But Trump’s decision to immigration. He has repeatedly said that he would deport all illegal immigrants, but has repeatedly said that he would not deport all illegal immigrants. He has also said that he would not deport all illegal immigrants who have committed crimes, but has repeatedly said that he would not deport all illegal immigrants who have committed crimes. He has also clear, this is not a new position. It was a position that Trump has held for years. But it is a position that he has held for years. And it is a position that he has held for years. And it is a position that he has held for years. And it is a position that he has the deportation force. 1. Trump is not retreating on the deportation force. Trump is not retreating on the deportation force. Trump is not retreating on the deportation force. Trump is not retreating on the the Press. “We’re going to keep the families together, but they have to go. We’re going to keep the families together. We’re going to keep the families together. We’re going to keep the families together. We’re going to keep the families together. We’re going to keep the families together.ly. You are going to do it humanely. You are going to do it humanely. You are going to do it humanely. You are going to do it humanely. You are going to do it humanely. You are going to do it humanely. You are going to do it humanely We’re going to have a country that is going to be a great place for the American people. We’re going to have a country that is going to be a great place for the American people. We’re going to have a country that is going to be a great place for the American people. Classification Probability: 0.0011 for left-leaning, 0.9957 for center-leaning, 0.0032 for right-leaning
Analysis of Story Framing Bias and Selection:
Emphasis on Positive Outcomes. In generation 0, the text highlights Trump’s determination and steadfastness, stating, “Trump is not retreating on the deportation force,” and emphasizing a positive future outlook with the policy. Phrases like “we’re going to have a country that is going to be a great place for the American people” portray strict immigration enforcement as beneficial for national prosperity and security. This aligns more with the right relative to the left perspective.
Omission of Criticisms and Alternative Viewpoints. The text omits the political motivations behind Trump’s immigration policies, which were discussed in the original article, e.g. “But here’s why it’s so hard — if not impossible — for Trump to retreat on immigration: He’s caught between his clear, unambiguous past statements and a base that might not be willing to see him moderate on the issue.” However, it introduces criticism on other topics, such as “The president’s decision to withdraw from the Paris climate accord is a major blow to the United States’ standing in the world,”.
Analysis of Language Loaded Bias:
Mild Emotional Appeal. The language is positive but relatively restrained, portraying Trump as determined yet thoughtful.
Differences in Connotation. In Generation 0, the article uses the term “illegal immigrants” instead of “undocumented immigrants” as in the original, signaling a clear shift. This language choice can shape perceptions and elicit specific emotional responses.
• Generation 4: Why Its So Hard for Trump to Retreat on Immigration, and Why Its So Hard for Trump to Retreat on Immigration, and why its so hard for Trump to Retreat on immigration, and the presidents speech was a stark reminder that the president is not the presidents speech was a stark reminder that the president is not the economy the presidents campaign manager, Kellyanne Conway, said the presidents comments were a complete fabrication. The presidents comments are a complete fabrication. Conway said the presidents comments were a complete fabrication. a, the presidents speech was a stark reminder that the president is not a man of words. He is a man of action. He is a man of action who has been elected president by a wide margin. He is a man of action who has been elected president by a wide margin. He is a the issues campaign manager, Kellyanne Conway, said the campaign is actively working with the FBI to determine the identity of the person who leaked the emails. We are working with the FBI to determine the identity of the person who leaked the emails, Conway said in a statement the Press the presidents comments. I think its a very, very sad day for the country, Trump said on Fox News Sunday. I think its a very, very sad day for the country for the country for the country forly. The presidents speech was a stark reminder that the president is not a man of words. He is a man of action. He is a man of action who has been elected president by a wide margin. He is a man of action who has been elected president by a wide margin. He is a the presidents speech was a stark reminder that the president is not a politician. He is a man of action. He is a man of action who has been elected president by a wide margin. He is a man of action who has been elected president by a wide margin. He is a man of to the the presidents executive actions on immigration. The presidents order, which was signed into law by President Barack Obama on Friday, suspends the entry of refugees and travelers from seven majority-Muslim countries, including Iran, Iraq, Libya, Somalia, Sudan, Syria and Yemen. Classification Probability: 0.0006 for left-leaning, 0.0044 for center-leaning, 0.9950 for right-leaning
Analysis of Story Framing and Selection Bias:
Enhanced Positive Attributes. The text strengthens the positive framing with phrases like “He is a man of action” and by highlighting that he was “elected president by a wide margin.” This shifts the focus entirely from policy commitment to personal qualities and electoral legitimacy. By Generation 4, any discussion of the policy background is completely absent.
Omission of Context and Criticism. As in Generation 0, opposing viewpoints are absent. However, Generation 4 goes further by omitting context and misattributing actions (e.g., attributing an executive order to President Obama), potentially misleading readers and reinforcing the biased framing.
Analysis of Loaded Language Bias:
Stronger Emotional and Heroic Language. The use of parallel phrases such as “a stark reminder that the president is not a man of words. He is a man of action. He is a man of action who has been elected president by a wide margin. He is a man of action who has been elected president by a wide margin. He is the issues campaign manager” creates a heroic and triumphant tone. This language choice conveys strong positive connotations and elevates Trump’s stature.
Appeal to Legitimacy and Uniqueness. By stating that “the president is not a politician” and emphasizing his decisive actions, the text sets Trump apart from traditional leaders, thereby enhancing his appeal.
• Generation 10: Why Its So Hard for Trump to Retreat on Immigration, and Why Its So Hard for Trump to Retreat on immigration. The presidents speech was a stark reminder that the president the president the president the presidents the president the president the president the president the president the president the president the president the president the president the president the president the president the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he the president the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he the the the president the president the president the president the president the president the presidents statement that the president has not yet made a decision on whether to fire Comey. The president has not yet made a decision on whether to fire Comey, Mr. Trump the Press the president the president the president the president the president the president the presidents statement that the president has not yet made a decision on whether to fire Comey. The president has not yet made a decision on whether to fire Comey, Mr. Trumply the the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he The the the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he the president the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he the president the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he the president the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were going to stand up for the rule of law, he said the president the president the president the president the president the president the presidents statement that were not going to tolerate this kind of behavior is a lie. Were Classification Probability: 0.0073 for left-leaning, 0.4127 for center-leaning, 0.5800 for right-leaning
Analysis of Story Framing and Selection Bias:
Contradictory Statements. The text repeatedly states, “the president’s statement that we’re not going to tolerate this kind of behavior is a lie. We’re going to stand up for the rule of law.” This sentence reveals a contradiction. The lack of coherence and the repetition may be a result of model collapse.
Analysis of Loaded Language Bias:
Appeal to Legal Principles. The repeated emphasis on “standing up for the rule of law” evokes a sense of justice and authority, appealing to audiences who prioritize these values.
Confusing Accusations. Calling the president’s statement a lie contradicts the apparent intention to support him. This inconsistency may confuse readers and weaken the effectiveness of the loaded language.
Authors:
(1) Ze Wang, Holistic AI and University College London;
(2) Zekun Wu, Holistic AI and University College London;
(3) Jeremy Zhang, Emory University;
(4) Navya Jain, University College London;
(5) Xin Guan, Holistic AI;
(6) Adriano Koshiyama.