René Kizilcec

コーネル大学の調査によると、低所得層の高校生ほどAIを使って大学出願の小論文を書くようになっており、それが合格に与える”ダメージ”が大きいとのこと。 AIを使えば文章の質は表面上は上がりますが、審査官はそれを「この学生らしくない」と受け取ることがあるためと考えられています。 今回研究者らが8万1千件以上の大学入試の小論文を分析したところ、ChatGPT登場後、低所得層の学生ほどAIに頼る割合が高くなっていました。これは理にかなっており、家庭教師や塾よりAIのほうが安価だからです。 ところが、AIを使った低所得層の出願者と同じ学力・属性の高所得層を比較すると、審査結果の格差は以前より31%も広がっていました。 なぜそうなるのか、実際は研究者たちも明確な答えは出せていませんが、AIが生み出す「うまい文章」は、文化資本を持つ人々の言葉のクセに基づいており、その規範に沿って書けば書くほど、規範から遠い学生が「不自然だ」と判断されてしまう構造になっているためと推測されています。 ということは、格差はAIを使った結果を評価する側の眼差しに宿っていることになります。そこが深いポイントです。

📣Festival of Learning 2026 🎆 June 29 – July 3, 2026 • Seoul, South Korea Co-location of AIED @AIEDconf, EDM @EDMConf2025, and L@S @LearningAtScale w/ contributions from SoLAR @SoLAResearch and ISLS @islsnews Mark your calendars, stay tuned! Website: festival-of-learning-2026.info

Excited to be at #lak2025 check out Kimberly Williamson’s paper today at 10:20 on Algo Appreciation among Educators! Happy to chat about the National Tutoring Observatory— we are hiring a postdoc and Asst Research Prof. nationaltutoringobservatory.org

Excited for @LearningAtScale to be in Palermo next summer together with @AIEDconf and @EDMConf2025 and confirmed keynote by @chrispiech!

What do LLMs write like? Researchers at @CornellCIS, @Stanford and @Penn found that #AI-generated college admissions essays are most similar to those written by male students with high socioeconomic status and privilege. @whynotyet @aj_alvero news.cornell.edu/stories/2024/1…

Sharing a paper just published in the Journal of Big Data! Studies have shown strong associations with writing and demography, so here we ask "who do LLMs write like?" Answer: higher social status people (with some caveats) BUT with very little variance link.springer.com/article/10.118…

@michael_muller We have not tested this but yes, you could ask the model to answer like a person on one part of the country and see how much information is needed to get the responses aligned. A benchmark dataset would be needed to evaluate alignment.

@whynotyet Would this approach work for a region with high internal heterogeneity? I am very grateful to live in such a place.

In addition to showing the cultural bias in GPT’s responses, we demonstrate a simple but effective control strategy to improve cultural alignment: cultural prompting, I.e. instructing the model to respond like someone from your region. It works best for the latest versions.

Researchers concerned about cultural bias in large language models asked ChatGPT questions about values used in international surveys. All versions of the model answered like people in English-speaking and Protestant European countries. In PNAS Nexus: ow.ly/4Oy350Trj4S

We are hiring at @CornellInfoSci ! Open rank: academicjobsonline.org/ajo/jobs/28405 Cornell is a fantastic place to work, please consider applying and sharing this opportunity.

Happy to share that @CornellInfoSci is hiring teaching faculty in all areas! Teaching Cornell students has been a wonderful experience. Please share this opportunity widely: academicjobsonline.org/ajo/jobs/28402

Returning from perpetual Twitter hiatus to spread the word: @CornellInfoSci is hiring! Tenure-track hires at all levels in Ithaca, focusing on qual/interpretive methods, CSS/quant approaches, study/design of human/AI interactions. Review starts 11/10: academicjobsonline.org/ajo/jobs/28405

Our SI on Advancing Equity and Inclusion in educational practices with AI-powered educational decision support systems (AI-EDSS) is out! Check all the valuable contributions! bera-journals.onlinelibrary.wiley.com/doi/10.1111/bj… @BJETEds @whynotyet @IoanaJJ @alywise @NiaDowell

What does the Life Cycle of Large Language Models look like? Check out our new BJET article mapping how LLM-based applications are developed and what biases can arise in each step. bera-journals.onlinelibrary.wiley.com/doi/10.1111/bj… arxiv.org/pdf/2407.11203 @JinsookLee1021 @cab938 @renzheyu @YannHicke

📣Cornell Postdoc Position: Join our interdisciplinary, NSF-funded project on deterring objectionable speech online with @nataliebazarova @drewmargolin @difrad Details here: academicjobsonline.org/ajo/jobs/27892 Thank you for sharing

Proceedings of the 2024 ACM Learning at Scale Conference are now available in the ACM Digital Library: dl.acm.org/doi/proceeding… #las24ed

📢 Super excited to share that I'll be joining @Columbia @ColumbiaDBMI @columbiahci as an Assistant Professor in Jan'25! Pls RT: I'm hiring PhDs & Postdocs in Sp/Fa'25 in #DBMI & #CS. My lab focuses on #HCI and #AI for health. Check out orsonxu.com for more info!

What cultural values do GPT-4o, 4, 3.5, 3 express? Using World Values Survey questions, we find GPT consistently aligns with English-speaking countries/Protestant Europe. We show that Cultural Prompting improves alignment. arxiv.org/abs/2311.14096 @yan_ytyt @OlgaOvi @BakerEDMLab

The #las24ed program is available online: learningatscale.hosting.acm.org/las2024/255-2/ And if you haven't registered yet, make sure to take advantage of regular rates by July 1! @OlgaOvi @A_MartinezMones @zpardos @SusanRSinger @rebquintana @openHPI @marimar_ps @choux @cdemmansepp @BakerEDMLab

Today is the last day to get #EarlyBird rates on registration for ACM Learning at Scale 2024 and @EDMConf2024. Co-registration options available. Join us in Atlanta! web.cvent.com/event/e0dcaf2a… #las24ed #edm2024