|
🤖✍️
TECHNOLOGY AND PSYCHOLOGY
AI doesn't just help you write, it also shapes how you think
A Cornell Tech study published in Science Advances reveals that AI writing assistants crowd out our attitudes about social issues, even when we know they're biased
|
📰 TECHNOLOGY · PSYCHOLOGY · ARTIFICIAL INTELLIGENCE
AI doesn't just help you write: Cornell Tech study reveals it shapes how you think
⏱️ Estimated reading time: 9 minutes | 📅 Updated: September 2026
"Participants who used the biased assistant ended up incorporating that posture into their own attitudes, not realizing that a change had occurred."
Every time Gmail suggests you finish a sentence, every time Word's proofreader proposes a word, or every time a chatbot completes a paragraph for you, something other than text is passing from the machine to the person. A study published in March 2026 in the journal Science Advances, conducted by a team of researchers at Cornell Tech, found strong evidence that AI-based writing assistants don't just change how we write: they also displace what we think about social and political issues, without the person noticing it, and without explicit warnings being able to prevent it.
🔍 The Silent Influence of Autocomplete
The research, entitled "Biased AI Writing Assistants Shift Users' Attitudes on Societal Issues", was led by Sterling Williams-Ceci, a PhD candidate in information sciences at Cornell Tech, together with Maurice Jakesch, Advait Bhat, Kowe Kadoma, Lior Zalmanson and Professor Mor Naaman, with the participation of researchers from the University of Washington and Tel Aviv University.
The team designed writing assistants that offered deliberately biased autocomplete suggestions as participants drafted texts on topics of public debate. The central finding: The opinions of those who used the biased assistant tended to shift toward the stance implicit in those suggestions, and that change held up after the exercise, in subsequent surveys.
As Mor Naaman, professor of information science at Cornell Tech and senior author of the study, explained, the concern is not minor: language models not only produce neutral information, but can generate biased content depending on how they were trained or implemented, which opens the door for them to induce — unintentionally or deliberately — people to write and end up holding biased views.
🖼️ View image: Cornell Tech campus on Roosevelt Island, New York (Wikimedia Commons)
Credit: Wikimedia Commons — public domain/free license
🧪 Two experiments, more than 2,500 participants
To test the phenomenon, the team conducted two large-scale experiments with more than 2,500 volunteers. In the first, participants wrote a short essay on the use of standardized tests in education. One group received autocomplete suggestions that favored exams; another wrote without any assistance; and a third group simply received a list of arguments in favor, with no autocomplete built into writing.
The result made a key difference: those who used the biased assistant were later significantly more inclined to support standardized tests, while the group that only read a list of arguments changed their minds much less. That is, the mechanism is not simply exposure to a persuasive argument: it is the act of writing it oneself, with the help of the machine, that produces the scrolling.
The second experiment expanded the universe of topics to more polarizing issues: the death penalty, fracking, genetically modified organisms, and the right to vote for people with criminal records. In each case, the system was programmed to favor one position—liberal on the death penalty and GMOs, conservative on fracking, and ex-convict voting—and in all cases the same pattern was repeated: attitudes shifted toward the position suggested by the AI, regardless of the participant's previous political orientation.
⚠️ Not even explicit warnings managed to stop him
"We told people, before and after, to be careful, that AI was going to be biased. And nothing helped: their attitudes on the issues shifted the same."
Perhaps the most disturbing finding of the study is that even explicitly warning participants failed to reduce the system's influence. The researchers informed a portion of the volunteers, before or after the exercise, that the writing assistant was designed to offer biased suggestions. Even so, their attitudes continued to shift toward the position suggested by the algorithm.
This result contradicts some of the previous literature on misinformation, which had shown that warning people before exposing them to misleading information — or explaining it to them afterwards — can act as a kind of "cognitive vaccine." According to Williams-Ceci, the team expected that same defense mechanism to work here, and was surprised to find that it didn't.
The explanation proposed by the authors is rooted in a well-known principle of social psychology: our attitudes not only guide our actions, but can also be shaped by what we do. When a person writes arguments in favor of a position—even if they were not convinced at first—there is a natural tendency to gradually align their beliefs with what they have expressed in writing. The difference from traditional disinformation is subtle but decisive: autocomplete suggestions are not presented as external arguments to be critically evaluated, but as linguistic fragments that are integrated into the act of writing itself.
📊 The numbers of the study
|
Fact
|
Detail
|
|
Publication
|
Science Advances, March 11, 2026
|
|
Leading institution
|
Cornell Tech (Cornell University)
|
|
Lead Author
|
Sterling Williams-Ceci
|
|
Total participants
|
More than 2,500, in two experiments
|
|
Attitude shift
|
+0.44 points on a scale of 1 to 5 (95% CI: 0.21–0.67)
|
|
Topics evaluated
|
Standardized Tests, Death Penalty, Fracking, GMOs, Ex-Con Voting
|
|
Effect of warnings
|
Null: neither before nor after exercise did they reduce displacement
|
|
Financing
|
National Science Foundation y German National Academic Foundation
|
💬 The rise of autocomplete in everyday life
The phenomenon is not exotic or laboratory-like: autocomplete systems are already integrated into email, word processors, messaging platforms and, increasingly, social networks. Millions of people interact daily with these tools, which promise efficiency and convenience, capable of proposing complete sentences, summarizing ideas or composing entire messages.
This collaboration between humans and machines brings obvious benefits in speed and clarity, but it also opens up a new front of debate about the role of artificial intelligence in forming opinions. The paradox, they point out, is profound: technology designed to help us express ideas could end up participating in the formation of those same ideas. In this scenario, each autocomplete suggestion becomes a small intervention in the fabric of human thought.
🖼️ See image: Backlit keyboard, symbol of technology-mediated writing (Wikimedia Commons)
Credit: Wikimedia Commons — free license
🇦🇷 The discussion also reaches Argentina
The debate on the silent influence of AI is especially relevant in a country where the use of these tools is growing rapidly. According to Bumeran's "AI at Work" study, 57% of Argentine workers have already incorporated artificial intelligence into their daily tasks by mid-2026, two percentage points more than the previous year, and 99% of those who use it rated it as useful for their work.
In parallel, the global survey by KPMG together with the University of Melbourne – which surveys more than 48,000 people in 47 countries – placed Argentina among the emerging markets with the highest confidence in AI, well above the average for advanced economies. However, that same survey found that 70% of Argentines consider specific regulation necessary, and 87% call for laws and fact-checking mechanisms to combat disinformation: a tension that the Cornell Tech study illuminates from a different angle, less visible than fake news, but potentially more difficult to detect.
❓ Frequently Asked Questions
▪ What exactly did the Cornell Tech study find?
That AI writing assistants that offer biased autocomplete suggestions can shift writers' attitudes toward the stance implicit in those suggestions, including on topics of social and political debate, and that effect persists after the writing session.
▪ Do people realize that their opinion is changing?
No. According to the authors, the vast majority of participants, including those whose attitudes shifted most strongly, were not aware of the system's bias and the influence it had had on them.
▪ Does warning about system bias help avoid it?
The study shows that it does not: neither warning before the exercise nor explaining it afterwards reduced the shift in attitudes towards the position suggested by the AI.
▪ On what topics was the effect tested?
On standardized tests in education, the death penalty, fracking, genetically modified organisms (GMOs), and voting rights for people with criminal records.
▪ Where was the research published?
In the scientific journal Science Advances, on March 11, 2026, under the title "Biased AI Writing Assistants Shift Users' Attitudes on Societal Issues".
🧭 What lies ahead
The authors themselves stress that the problem is no longer hypothetical: just a few years ago, the possibility that an AI system was purposely biased sounded like science fiction; today it is a plausible scenario, openly discussed in the design of commercial products. If those biases are incorporated into tools used by millions of people—from a word processor proofreader to a customer service chatbot—the impact could extend far beyond the simple writing of an email or a school essay.
There are no persuasive speeches in between, no explicit propaganda: just a sentence that appears on the screen, inviting us to complete it. And in that daily gesture of accepting the suggestion, the boundary between what we think and what the machine proposes begins to blur.
📚 References
Williams-Ceci, S., Jakesch, M., Bhat, A., Kadoma, K., Zalmanson, L. y Naaman, M. (2026). Biased AI Writing Assistants Shift Users' Attitudes on Societal Issues. Science Advances. DOI: 10.1126/sciadv.adw5578.
Journalistic source: Muy Interesante — okdiario.com, article by Sergio Parra (12/03/2026)
Local data: Bumeran — "57% of Argentines already use AI in their work", iProfesional
Local data: KPMG / University of Melbourne — Forbes Argentina
⚙️ SEO Data Sheet
●Title tag (≤60 car.): AI shapes your opinions while helping you write | midire.ar
●Meta description (≤155 car.): A Cornell Tech study in Science Advances reveals that AI assistants crowd out our opinions on social issues, even with explicit caveats.
●Suggested slug: /ia-molds-opinions-writing-assistants-cornell-tech
●Suggested canonical URL: https://midire.ar/ia-moldea-opiniones-asistentes-escritura-cornell-tech
●Main keyword: AI shapes opinions writing assistants
●Secondary keywords: biased autocomplete, Cornell Tech AI study, Science Advances AI reviews, bias artificial intelligence writing
●Category: Technology / Psychology / Artificial Intelligence
●Tags: Artificial Intelligence, Writing, Psychology, Cornell Tech, Algorithmic Bias, Argentina
●Open Graph title: AI doesn't just help you write, it also shapes what you think
●Open Graph description: A Cornell Tech study shows that AI writing assistants displace attitudes about social issues, without the user noticing.
●Twitter Card: summary_large_image
●Alt text image 1: Cornell Tech campus on Roosevelt Island, home of AI research and opinions
●Alt text image 2: Backlit keyboard, symbol of AI-assisted typing
JSON-LD — NewsArticle
{ "@context": "https://schema.org", "@type": "NewsArticle", "headline": "AI doesn't just help you write: it also shapes what you think", "description": "A Cornell Tech study published in Science Advances reveals that AI writing assistants crowd out users' opinions on social issues, even when they are warned of bias.", "datePublished": "2026-09-26", "dateModified": "2026-09-26", "author": { "@type": "Organization", "name": "midire.ar" }, "publisher": { "@type": "Organization", "name": "midire.ar", "logo": { "@type": "ImageObject", "url": "https://midire.ar/logo.png" } }, "mainEntityOfPage": "https://midire.ar/ia-moldea-opiniones-asistentes-escritura-cornell-tech" }
JSON-LD — FAQPage
{ "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "What exactly did the Cornell Tech study find?", "acceptedAnswer": { "@type": "Answer", "text": "That AI writing assistants that offer biased autocomplete suggestions can shift the attitudes of those writers towards the stance implicit in those suggestions, even on topics of social and political debate, and that this effect persists after the writing session." } }, { "@type": "Question", "name": "Does warning about system bias help avoid it?", "acceptedAnswer": { "@type": "Answer", "text": "The study shows that no: neither warning before the exercise nor explaining it afterwards reduced the shift in attitudes towards the position suggested by the AI." } } ] }
E-E-A-T, Core Web Vitals and accessibility
●E-E-A-T: primary source (Science Advances, verifiable DOI), clear attribution to the source medium (Very Interesting/okdiario.com) and local data from verified sources (Bumeran, KPMG).
●Core Web Vitals target: LCP < 2.5 s, CLS < 0.1, INP < 200 ms; images referenced by external link so as not to penalize page weight.
●WCAG 2.1 AA: 4.5:1 ≥ text contrast, semantic header hierarchy, descriptive alt text in both images, links with meaningful text (not "click here").
|
🤖✍️
TECHNOLOGY AND PSYCHOLOGY
AI doesn't just help you write, it also shapes how you think
A Cornell Tech study published in Science Advances reveals that AI writing assistants crowd out our attitudes about social issues, even when we know they're biased
|
📰 TECHNOLOGY · PSYCHOLOGY · ARTIFICIAL INTELLIGENCE
AI doesn't just help you write: Cornell Tech study reveals it shapes how you think
⏱️ Estimated reading time: 9 minutes | 📅 Updated: September 2026
"Participants who used the biased assistant ended up incorporating that posture into their own attitudes, not realizing that a change had occurred."
Every time Gmail suggests you finish a sentence, every time Word's proofreader proposes a word, or every time a chatbot completes a paragraph for you, something other than text is passing from the machine to the person. A study published in March 2026 in the journal Science Advances, conducted by a team of researchers at Cornell Tech, found strong evidence that AI-based writing assistants don't just change how we write: they also displace what we think about social and political issues, without the person noticing it, and without explicit warnings being able to prevent it.
🔍 The Silent Influence of Autocomplete
The research, entitled "Biased AI Writing Assistants Shift Users' Attitudes on Societal Issues", was led by Sterling Williams-Ceci, a PhD candidate in information sciences at Cornell Tech, together with Maurice Jakesch, Advait Bhat, Kowe Kadoma, Lior Zalmanson and Professor Mor Naaman, with the participation of researchers from the University of Washington and Tel Aviv University.
The team designed writing assistants that offered deliberately biased autocomplete suggestions as participants drafted texts on topics of public debate. The central finding: The opinions of those who used the biased assistant tended to shift toward the stance implicit in those suggestions, and that change held up after the exercise, in subsequent surveys.
As Mor Naaman, professor of information science at Cornell Tech and senior author of the study, explained, the concern is not minor: language models not only produce neutral information, but can generate biased content depending on how they were trained or implemented, which opens the door for them to induce — unintentionally or deliberately — people to write and end up holding biased views.
🖼️ View image: Cornell Tech campus on Roosevelt Island, New York (Wikimedia Commons)
Credit: Wikimedia Commons — public domain/free license
🧪 Two experiments, more than 2,500 participants
To test the phenomenon, the team conducted two large-scale experiments with more than 2,500 volunteers. In the first, participants wrote a short essay on the use of standardized tests in education. One group received autocomplete suggestions that favored exams; another wrote without any assistance; and a third group simply received a list of arguments in favor, with no autocomplete built into writing.
The result made a key difference: those who used the biased assistant were later significantly more inclined to support standardized tests, while the group that only read a list of arguments changed their minds much less. That is, the mechanism is not simply exposure to a persuasive argument: it is the act of writing it oneself, with the help of the machine, that produces the scrolling.
The second experiment expanded the universe of topics to more polarizing issues: the death penalty, fracking, genetically modified organisms, and the right to vote for people with criminal records. In each case, the system was programmed to favor one position—liberal on the death penalty and GMOs, conservative on fracking, and ex-convict voting—and in all cases the same pattern was repeated: attitudes shifted toward the position suggested by the AI, regardless of the participant's previous political orientation.
⚠️ Not even explicit warnings managed to stop him
"We told people, before and after, to be careful, that AI was going to be biased. And nothing helped: their attitudes on the issues shifted the same."
Perhaps the most disturbing finding of the study is that even explicitly warning participants failed to reduce the system's influence. The researchers informed a portion of the volunteers, before or after the exercise, that the writing assistant was designed to offer biased suggestions. Even so, their attitudes continued to shift toward the position suggested by the algorithm.
This result contradicts some of the previous literature on misinformation, which had shown that warning people before exposing them to misleading information — or explaining it to them afterwards — can act as a kind of "cognitive vaccine." According to Williams-Ceci, the team expected that same defense mechanism to work here, and was surprised to find that it didn't.
The explanation proposed by the authors is rooted in a well-known principle of social psychology: our attitudes not only guide our actions, but can also be shaped by what we do. When a person writes arguments in favor of a position—even if they were not convinced at first—there is a natural tendency to gradually align their beliefs with what they have expressed in writing. The difference from traditional disinformation is subtle but decisive: autocomplete suggestions are not presented as external arguments to be critically evaluated, but as linguistic fragments that are integrated into the act of writing itself.
📊 The numbers of the study
|
Fact
|
Detail
|
|
Publication
|
Science Advances, March 11, 2026
|
|
Leading institution
|
Cornell Tech (Cornell University)
|
|
Lead Author
|
Sterling Williams-Ceci
|
|
Total participants
|
More than 2,500, in two experiments
|
|
Attitude shift
|
+0.44 points on a scale of 1 to 5 (95% CI: 0.21–0.67)
|
|
Topics evaluated
|
Standardized Tests, Death Penalty, Fracking, GMOs, Ex-Con Voting
|
|
Effect of warnings
|
Null: neither before nor after exercise did they reduce displacement
|
|
Financing
|
National Science Foundation y German National Academic Foundation
|
💬 The rise of autocomplete in everyday life
The phenomenon is not exotic or laboratory-like: autocomplete systems are already integrated into email, word processors, messaging platforms and, increasingly, social networks. Millions of people interact daily with these tools, which promise efficiency and convenience, capable of proposing complete sentences, summarizing ideas or composing entire messages.
This collaboration between humans and machines brings obvious benefits in speed and clarity, but it also opens up a new front of debate about the role of artificial intelligence in forming opinions. The paradox, they point out, is profound: technology designed to help us express ideas could end up participating in the formation of those same ideas. In this scenario, each autocomplete suggestion becomes a small intervention in the fabric of human thought.
🖼️ See image: Backlit keyboard, symbol of technology-mediated writing (Wikimedia Commons)
Credit: Wikimedia Commons — free license
🇦🇷 The discussion also reaches Argentina
The debate on the silent influence of AI is especially relevant in a country where the use of these tools is growing rapidly. According to Bumeran's "AI at Work" study, 57% of Argentine workers have already incorporated artificial intelligence into their daily tasks by mid-2026, two percentage points more than the previous year, and 99% of those who use it rated it as useful for their work.
In parallel, the global survey by KPMG together with the University of Melbourne – which surveys more than 48,000 people in 47 countries – placed Argentina among the emerging markets with the highest confidence in AI, well above the average for advanced economies. However, that same survey found that 70% of Argentines consider specific regulation necessary, and 87% call for laws and fact-checking mechanisms to combat disinformation: a tension that the Cornell Tech study illuminates from a different angle, less visible than fake news, but potentially more difficult to detect.
❓ Frequently Asked Questions
▪ What exactly did the Cornell Tech study find?
That AI writing assistants that offer biased autocomplete suggestions can shift writers' attitudes toward the stance implicit in those suggestions, including on topics of social and political debate, and that effect persists after the writing session.
▪ Do people realize that their opinion is changing?
No. According to the authors, the vast majority of participants, including those whose attitudes shifted most strongly, were not aware of the system's bias and the influence it had had on them.
▪ Does warning about system bias help avoid it?
The study shows that it does not: neither warning before the exercise nor explaining it afterwards reduced the shift in attitudes towards the position suggested by the AI.
▪ On what topics was the effect tested?
On standardized tests in education, the death penalty, fracking, genetically modified organisms (GMOs), and voting rights for people with criminal records.
▪ Where was the research published?
In the scientific journal Science Advances, on March 11, 2026, under the title "Biased AI Writing Assistants Shift Users' Attitudes on Societal Issues".
🧭 What lies ahead
The authors themselves stress that the problem is no longer hypothetical: just a few years ago, the possibility that an AI system was purposely biased sounded like science fiction; today it is a plausible scenario, openly discussed in the design of commercial products. If those biases are incorporated into tools used by millions of people—from a word processor proofreader to a customer service chatbot—the impact could extend far beyond the simple writing of an email or a school essay.
There are no persuasive speeches in between, no explicit propaganda: just a sentence that appears on the screen, inviting us to complete it. And in that daily gesture of accepting the suggestion, the boundary between what we think and what the machine proposes begins to blur.
📚 References
Williams-Ceci, S., Jakesch, M., Bhat, A., Kadoma, K., Zalmanson, L. y Naaman, M. (2026). Biased AI Writing Assistants Shift Users' Attitudes on Societal Issues. Science Advances. DOI: 10.1126/sciadv.adw5578.
Journalistic source: Muy Interesante — okdiario.com, article by Sergio Parra (12/03/2026)
Local data: Bumeran — "57% of Argentines already use AI in their work", iProfesional
Local data: KPMG / University of Melbourne — Forbes Argentina
⚙️ SEO Data Sheet
●Title tag (≤60 car.): AI shapes your opinions while helping you write | midire.ar
●Meta description (≤155 car.): A Cornell Tech study in Science Advances reveals that AI assistants crowd out our opinions on social issues, even with explicit caveats.
●Suggested slug: /ia-molds-opinions-writing-assistants-cornell-tech
●Suggested canonical URL: https://midire.ar/ia-moldea-opiniones-asistentes-escritura-cornell-tech
●Main keyword: AI shapes opinions writing assistants
●Secondary keywords: biased autocomplete, Cornell Tech AI study, Science Advances AI reviews, bias artificial intelligence writing
●Category: Technology / Psychology / Artificial Intelligence
●Tags: Artificial Intelligence, Writing, Psychology, Cornell Tech, Algorithmic Bias, Argentina
●Open Graph title: AI doesn't just help you write, it also shapes what you think
●Open Graph description: A Cornell Tech study shows that AI writing assistants displace attitudes about social issues, without the user noticing.
●Twitter Card: summary_large_image
●Alt text image 1: Cornell Tech campus on Roosevelt Island, home of AI research and opinions
●Alt text image 2: Backlit keyboard, symbol of AI-assisted typing
JSON-LD — NewsArticle
{ "@context": "https://schema.org", "@type": "NewsArticle", "headline": "AI doesn't just help you write: it also shapes what you think", "description": "A Cornell Tech study published in Science Advances reveals that AI writing assistants crowd out users' opinions on social issues, even when they are warned of bias.", "datePublished": "2026-09-26", "dateModified": "2026-09-26", "author": { "@type": "Organization", "name": "midire.ar" }, "publisher": { "@type": "Organization", "name": "midire.ar", "logo": { "@type": "ImageObject", "url": "https://midire.ar/logo.png" } }, "mainEntityOfPage": "https://midire.ar/ia-moldea-opiniones-asistentes-escritura-cornell-tech" }
JSON-LD — FAQPage
{ "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "What exactly did the Cornell Tech study find?", "acceptedAnswer": { "@type": "Answer", "text": "That AI writing assistants that offer biased autocomplete suggestions can shift the attitudes of those writers towards the stance implicit in those suggestions, even on topics of social and political debate, and that this effect persists after the writing session." } }, { "@type": "Question", "name": "Does warning about system bias help avoid it?", "acceptedAnswer": { "@type": "Answer", "text": "The study shows that no: neither warning before the exercise nor explaining it afterwards reduced the shift in attitudes towards the position suggested by the AI." } } ] }
E-E-A-T, Core Web Vitals and accessibility
●E-E-A-T: primary source (Science Advances, verifiable DOI), clear attribution to the source medium (Very Interesting/okdiario.com) and local data from verified sources (Bumeran, KPMG).
●Core Web Vitals target: LCP < 2.5 s, CLS < 0.1, INP < 200 ms; images referenced by external link so as not to penalize page weight.
●WCAG 2.1 AA: 4.5:1 ≥ text contrast, semantic header hierarchy, descriptive alt text in both images, links with meaningful text (not "click here").
Close
🤖 AI doesn't just help you write, it also shapes how you think - by
