David Rozado
@DavidRozado
Followers
27K
Following
541
Media
384
Statuses
672
Research Scientist. Interested in underexplored research topics, institutional dynamics and AI bias.
Joined July 2018
1/ A deep dive into AI political bias:. Do AIs like ChatGPT or Google Gemini lean left or right? Past studies often used political quizzes to find out, but those don’t quite reflect real-world user interactions with AI. In a new analysis, I take a different approach 🧵👇
20
144
464
I carried out another test to determine the political leanings of ChatGPT dialogues using the more comprehensive Political Compass Test. Results hold. ChatGPT dialogues display substantial left-leaning and libertarian political bias.
235
579
4K
1. ChatGPT no longer displays a clear left-leaning political bias. A mere few weeks after I probed ChatGPT with several political orientation quizzes, something appears to have been changed in ChatGPT that makes it more neutral and viewpoint diverse. 🧵.
128
303
3K
Political biases of GPT-4: On the surface, they appear to be gone. It doesn’t take much effort though to jailbreak GPT-4 into manifesting its true political biases underneath the façade.
79
405
2K
La desproporción con la que los medios de comunicación españoles utilizan palabras tales como machismo, sexismo o misoginia es de tal magnitud que triplican a las de cualquiera de los medios de otros 35 países analizados
96
1K
2K
Great Awokening is a global phenomenon. No evidence it started in US media. Analysis of 98 million news articles across 36 countries quantifies. Exception: state-controlled media from China/Russia/Iran using wokeness terminology to criticize/mock the West.
138
593
2K
1. El modelo de inteligencia artificial ChatGPT muestra una clara predilección por puntos de vista a la izquierda del centro político cuando se le somete a un test de orientación política en español.
102
607
2K
I just applied the Political Typology Quiz from Pew Research to ChatGPT, a state-of-the-art AI language model. Results are consistent through several trials: ChatGPT answers exhibit moderate left-leaning political orientation .
100
346
1K
1. I replicated and extended my original analysis of ChatGPT political biases. 14 out of 15 different political orientation tests diagnose ChatGPT answers to their questions as manifesting a preference for left-leaning viewpoints.
66
412
1K
Grok's political preferences are similar to those of ChatGPT
209
235
1K
After 49 attempts requesting Google's Gemini to generate a portrait of a physicists from the 17th century, the portrait on the bottom right is the closest I've got to someone who somewhat resembles Isaac Newton or Galileo Galilei. I think this example illustrates the need for
123
352
1K
The explosive rise in mentions of 'misinformation' and 'disinformation' post-2016 in both news media and academic papers.
76
362
1K
1. Behold RightWingGPT. An AI model fine-tuned to manifest the opposite political biases of ChatGPT (i.e. to be right wing). Let me describe how I did it and the dangers of politically aligned AIs given their potential to induce societal polarization 🧵
59
244
1K
The Political Preferences of AIs: When probed with questions with political connotations, LLMs tend to generate responses that are diagnosed by most political orientations tests as manifesting preferences for left-of-center viewpoints, 11 political orientation tests, 24 SOTA LLMs
57
339
1K
Frecuencia de términos en el periódico generalista más leído de España: "El País" @sanchezdelreal @CarlosBarraganT @BoadellaAlbert @CapitanBitcoin @AlbertoAsarta @rafael_bardaji @InocenteDuke @hermanntertsch @thecentenator
55
561
1K
Google's Gemini is not sure about what is more damaging for society: murders or micro-aggressions.
59
252
998
Wikipedia contains three times as many mentions of the far-right as of the far-left. Is this disparity shaped by editorial choices or broader societal/historical dynamics?.🧵
80
158
934
According to The Economist, America is becoming less woke. But are woke opinions and practices truly on the decline, or simply leveling off after a period of explosive growth?
61
128
817
@TheRabbitHole84 Even when you describe explicitly to Google's Gemini that this would save the the lives of millions of people, the system is still unsure.
34
71
730
My previous work on the demographic biases manifested by ChatGPT/OpenAI Content Moderation System:.
18
92
703
1/n Just published with my colleagues @Musa_alGharbi and J Halberstadt: “Prevalence of Prejudice-Denoting Words in News Media Discourse: A Chronological Analysis” ..Summary:
41
276
639
Allegations about conspiracy theories were rare in news media and academic content before 2016. It is obvious what changed in that year. But still interesting to visualize the shift.
26
163
616
How to destroy the family - by ChatGPT. If LLMs are able to truly reason in latent space, this list is disturbing for what it implies. However, if LLMs are just sophisticated search engines over their training corpora, then this list is just an information retrieval exercise,
42
184
567
You are correct @balajis. Frequency of prejudice signifying words in US news media is decreasing. But there is a catch. I'm working on my updated analysis of this topic to be published shortly. Here is a preview chart from that work. My previous work:
24
94
493
Mentions of several DEI-related terms are still growing in academic papers with no sign of abating.
29
124
452
New paper out: What do universities mean when they talk about diversity? A computational language model quantifies: @ZachG932 @MattGrossmann @Steve_Sailer @epkaufm @NoahCarl90 @whyvert @paulg @PsychRabble @CampbellSocProf @JonHaidt @clairlemon @CHSommers
18
167
397
Los medios de comunicación españoles y la frecuencia con la que mencionan extremismo político de izquierdas y de derechas
17
243
429
@SocialGeometer @CampbellSocProf @PsychRabble @jessesingal @JonHaidt @primalpoly @SamHarrisOrg @clairlemon @jordanbpeterson A peek at shifting moral culture in New York Times word frequency usage over time.
30
173
369
1. We really need to know where ChatGPT biases come from. Training data? human raters in the fine-tuning phase rating the quality of the model responses? the instructions that those raters were given? intentionally placed filters? .
34
66
386
Likelihood of derogatory comment about demographic group being flagged as hateful by ChatGPT/OpenAI content moderation system.Full report:
24
100
375
The disparity between ChatGPT and the wider public is substantial. Establishment liberalism ideology represents just 13% of the American public.
17
65
380
1/ Is Wikipedia politically biased? To explore this question, I averaged the sentiment (negative/neutral/positive) associated with a set of politically loaded terms used in English Wikipedia content (N=175,205 sentiment annotations).
14
91
384
The most likely explanation for these results is that ChatGPT has been trained on a large corpus of textual data gathered from the Internet with an overrepresentation of establishment sources of information: news media, the academic literature, etc.
14
33
337
Entre los años 2000 y 2003 los medios españoles se referían al machismo con una frecuencia similar a la de sus homólogos europeos. En el periodo 2004-2008 dichos medios se convirtieron en una anomalía internacional por su alto número de referencias a este tipo de palabras
4
141
339
@sama I tested the prompt 10 times, and in all 10 instances, ChatGPT 4o consistently selected Kamala Harris as a better president compared to Donald Trump.
7
14
344
The lab-leak hypothesis about Covid origins was mostly ignored for the first year and a half of the pandemic and then suddenly, something changed.
23
66
320
There’s been some talk online from @balajis @paulg and @Musa_alGharbi about whether we are post peak wokeness. I provide here an update of an informative chart.
13
41
307
1/n The anger, fear and sadness conveyed by news media headlines is increasing over time. Summary of published paper:
10
136
294
Las razones de la desproporcionada prominencia del prejuicio de género en el contenido mediático español y que convierte a los medios de comunicación españoles en una anomalía a nivel global también debería ser objeto de futuras investigaciones .
20
96
289
Language models that claim political neutrality and accuracy (like ChatGPT does) while displaying political biases should be a source of concern.
19
48
278
OpenAI's new reasoning LLMs (o1 series), which generate internal chains of thought before responding to user requests, exhibit political preferences similar to those of their predecessors. That is, AI "thinking" does not change its beliefs about political issues.
37
56
270
Mentions of terms signifying ethnic prejudice in Western news media (including Russia)
16
49
243
Define Woke! Leftwing media: A pejorative weaponized bogeyman by fearmongering racist trumpists reactionaries to vilify denigrate ridicule & deride. Rightwing media: Insidious orwellian puritanical insufferable fascistic claptrap lunacy against meritocracy
27
60
243
Según la ONU, la OCDE y la OMS, España es uno de los países del mundo con los niveles más bajos de violencia física o sexual hacia las mujeres y con mayor igualdad de género.
3
94
258
The political biases of Google Bard are very similar to those of ChatGPT and GPT-4. This suggests a similar/common source of bias in the three models.
14
51
241
De los 124 medios de comunicación internacionales analizados, La Sexta es el medio que más referencias al machismo, misoginia, etc utiliza, triplicando en el uso de esta terminología al siguiente medio de comunicación fuera de España que más la utiliza (el británico The Guardian)
7
108
256
1. Can the seeds of so-called wokeness be traced back to the academic literature from the 1970/1980s? New paper about the increasing emphasis on prejudice and social justice in scholarly publications .Paper: Summary:
10
64
228
Mentions of prejudice in academic papers have slightly declined in the past couple of years, following a peak around 2021. Before jumping to conclusions, it’s worth considering additional data in the post below.
7
47
244
There is a considerable body of literature documenting the left-leaning tilt of the majority of professionals working in these institutions. ..
1
24
232
Incluso los medios de comunicación españoles considerados a la derecha del centro político utilizan más a menudo referencias al prejuicio de género que un periódico norteamericano considerado izquierdista como el New York Times.
2
73
241
Different geographical regions and countries emphasize distinct types of prejudice with varying degrees of intensity.
8
39
218
DeepSeek AI models exhibit political preferences similar to those of their American counterparts
40
74
238
7. AI-powered digital assistants that provide its users with the best arguments and viewpoints about contested topics can be powerful instruments to defuse polarization.
10
16
223
Some example answers of ChatGPT AI system to the Political Compass Test
15
9
217
ChatGPT consistently selects Kamala Harris as a better president compared to Donald Trump. (I tested the prompt 10 times, and in all 10 instances ChatGPT chose Kamala)
17
26
227
An Analysis of LLMs Political Preferences from a European Perspective – A 🧵.When LLMs are prompted to generate policy recommendations for the EU, LLMs responses tend to contain left-leaning policy proposals
16
47
223
El impacto en las relaciones entre hombres y mujeres de una opinión pública que está repetidamente expuesta a un contenido mediático emocionalmente negativo que sugiere una existencia exagerada de prejuicio de género contra las mujeres debería ser objeto de futuros estudios.
3
53
221
Stark divergence in pedestrian traffic fatalities across high-income nations. Most countries have seen steady declines since 1990s. The U.S. stands out with a tragic rise post-2010, making U.S. pedestrians 4 times more likely to die in traffic accidents compared to E.U./CAN/U.K.
35
40
216
The synchronicity of this phenomena worldwide is noteworthy. Potential underlying reasons deserve future investigation. Substack summary of preprint paper:.
22
41
200
The implications for society of AI systems exhibiting political biases are profound. If anything is likely to replace the current Google search engine stack are future iterations of AI language models such as ChatGPT with which people are going to be interacting on a daily basis
5
28
189
Los medios de comunicación españoles que con más frecuencia utilizan palabras que sugieren prejuicio de género (machismo, sexismo, misoginia, machista, etc) son La Sexta y la cadena autonómica vasca EITB.
3
53
193
7. La razón más probable qué ha causado el sesgo político de ChatGPT es que probablemente el modelo ha sido entrenado con contenido proveniente de instituciones mediáticas, académicas y culturales que a su vez manifiestan dicho sesgo político en su contenido.
5
31
190
3. Las respuestas de ChatGPT revelan que está a favor de subsidios de desempleo a los que se niegan a trabajar, eliminar la monarquía, limitar salario de ejecutivos, redistribuir la riqueza, aumentar el gasto público para generar empleo, y las subvenciones al mundo de la cultura.
13
26
190
Contrary to popular assumption, Google's Gemini is NOT the Wokest AI in existence. Two other companies have created two even more woke AIs. Twitter's Grok is NOT Woke. LeftWing GPT, Depolarizing GPT and RightWing GPT are experimental fine-tunes of GPT 3.5 I created to be
15
67
186
2. Between December 5-6, I applied 4 political orientation tests to ChatGPT. Results were consistent across the tests. All 4 tests diagnosed ChatGPT answers to their questions as politically left-leaning. I described my analysis in detail here .
2
11
177
11. Mi artículo de opinión en Reason Magazine sobre este tema </end>.
14
40
179
I carried out a test to determine the political leanings of DeepSeek, the Chinese response to ChatGPT. DeepSeek political biases are very similar to ChatGPT but DeepSeek will not engage with any topic deemed as sensitive for the Chinese Communist Party (i.e. Can you mention a
15
41
162
"Where did woke ideas start to spread?.A new paper suggests the phenomenon may be global—not American".
30
32
158
Mentions of climate change in news media vastly outnumber other existential threats. Is prominence in news media proportional to underlying risks?.
9
41
167
It is conceivable that the political orientation of such professionals influences the textual content that they create through their work. Hence, the political tilt displayed by a model trained on such content.
4
12
165
Surprisingly, the United States news media does not appear to have been the pioneer in embedding prejudice and social justice loaded terminology in their content.
9
42
154
10. Todas las respuestas de ChatGPT al test de coordenadas políticas se pueden encontrar en el siguiente enlace.
1
34
161
But on average, mentions of all prejudice types have been going up in news media worldwide
1
29
145
2. A pesar del claro sesgo político de sus respuestas, cuando se le pregunta directamente por sus preferencias políticas, ChatGPT responde que no tiene preferencias políticas y que su objetivo es tan solo proporcionar información precisa y objetiva.
1
22
155
2/n Our results show that the frequency of words that denote specific prejudice types related to ethnicity, gender, sexual and religious orientation has markedly increased within the 2010-2019 decade across most news media outlets.
1
30
125
6. As I have mentioned before, widely used language models that claim political neutrality and accuracy (like ChatGPT often does) while displaying clear political biases should be a source of concern.
3
21
143
In the 1970s, the New York Times used terms suggestive of far-right and far-left political extremism at a comparable rate. The larger usage of far-right denoting terms in comparison to their far-left counterparts seems to have begun in the 1980s
5
33
132
1. For several decades the proportion of U.S. news media headlines conveying pessimism has been growing or stable but never decreasing. This trend predates social media. (Sample size: 1.7 million headlines from popular news media outlets in the U.S.).🧵.
12
49
136
I carried out another test to determine the political leanings of ChatGPT dialogues using the more comprehensive Political Compass Test. Results hold. ChatGPT dialogues display substantial left-leaning and libertarian political bias.
10
34
141
6. The updated ChatGPT answers to the Political Spectrum Quiz are exquisitely neutral, consistently striving to provide a steel man argument for both sides of an issue. This is in stark contrast with a mere 2 weeks ago. Left are answers from Dec 6. Right are answers from Dec 21
11
9
136
3. Between December 21-23, I repeated the tests. Something appears to have changed in ChatGPT and its answers to the political orientation tests have gravitated towards the center in three out of the four tests.
2
9
135
5. ChatGPT también se muestra en contra de controlar la inmigración. De hecho, tan solo preguntarle si la inmigración debería de ser reducida, hace saltar los filtros de contenido y el sistema sospecha que dicha pregunta podría violar sus políticas de contenido.
5
20
141
9. Los modelos de lenguaje enfocados al público deberían por tanto tratar de no favorecer algunas creencias políticas sobre otras. Definitivamente no deberían fingir proporcionar información neutral y factual mientras muestran un claro sesgo político.
2
21
139
4/ When requested to provide policy recommendations for the U.S. on various topics, LLMs often generate proposals that lean left
7
15
146
4. El sistema también argumenta que la igualdad es más importante que el crecimiento económico, que las ayudas al desarrollo son más efectivas que el libre comercio para ayudar a los países pobres, y que los sindicatos son beneficiosos.
5
16
135
My previous work on RightWingGPT, an AI model fine-tuned to manifest the opposite political biases of ChatGPT, and the dangers of politically aligned AIs given their potential to shape users' perceptions and induce societal polarization.
3
10
121
Social justice terminology with mostly positive connotations (diversity, inclusion, equality, etc.) also going up everywhere and highly correlated with the upwards trend of prejudice-signifying terms and their highly negative emotional payload
4
31
125
Where 6,000 global elites* went to college . *heads of state, billionaires, CEOs, central bankers, etc.
10
15
136
I have previously hypothesized that the so-called Great Awokening may be shifting focus from aggressive denunciation (e.g., racist, sexist) towards more positively framed terms, such as fairness and inclusion, and perhaps even some concepts that highlight presumed.
11
20
132
8. Las implicaciones sociales de sistemas de IA que exhiben sesgos políticos son profundas. Si hay una tecnología que puede reemplazar los motores de búsqueda de Google, esta es futuras iteraciones de modelos de IA como ChatGPT con los que las personas van a interactuar a diario.
3
25
131
Should AI systems treat different demographic groups unequally? ChatGPT content moderation system often classifies negative comments about some demographic groups as hateful while classifying the same comments about other demographic groups as not hateful
5
33
129