The Research

AI does not always tell
complete stories.

Four independent institutions — Cornell, UC Berkeley, aiEDU, and DAIR — documented the same problem from different angles. Civics Remix was built to address exactly what the research identifies.

4
Independent
Research Sources
107
Countries in
Cornell/UPenn Study
1,000+
Skills Analyzed by
aiEDU & Burning Glass

From Isabella, Founder

"I noticed that AI Ethics was not always neutral — that depending on how you asked the question, AI could give you an incomplete or slanted answer. That last part is the key: how you ask the question. I wanted to understand why prompting matters, learn how to use it to surface more complete answers, and then teach other young people to do the same — so they could think more critically, question the systems around them, and actually become the changemakers who improve society. That is why Civics Remix exists."

Isabella, Founder — Civics Remix / RockStars4Impact  ·  Holton-Arms School, Class of 2027

Four independent sources.
One shared conclusion.

The problem is real, the skills are necessary, and the approach works. Here is what the research found — and why it points directly to what Civics Remix does.

4
Independent Sources

Cornell University / University of Pennsylvania

Cultural bias and cultural alignment of large language models

Tested five major AI models against data from 107 countries. AI default responses can systematically favor Western, English-speaking cultural values — telling incomplete stories about everyone else. Targeted prompting measurably reduces that bias for the majority of countries tested.

When AI is not carefully prompted, it defaults to a narrow cultural lens. Teaching students to recognize and correct that default is a civic skill — and the core of what Civics Remix does in every session.
Tao et al. — PNAS Nexus, Vol. 3, Issue 9, 2024 ↗

aiEDU & Burning Glass Institute

Which Skills Matter Now? A data-driven framework for K–12 in the age of AI

Analyzed 1,000 workforce skills and 140 high school learning objectives. The report identifies ethical reasoning and ethical judgment as among the highest-priority skills for the AI era — essential not only for the workforce but for a thriving democracy. They must be taught deliberately. AI cannot build them on its own.

Ethical reasoning and ethical judgment are essential not only for economic opportunity but for a thriving democracy in the age of AI. Civics Remix is built on exactly those skills — and delivers them in every session.
Andreason, Sigelman et al. — Burning Glass Institute / aiEDU ↗

UC Berkeley / Google DeepMind

Rediet Abebe — AI equity research rooted in community and lived experience

Abebe's framework: the same AI tools that can tell slanted stories can, if carefully prompted, be redirected toward more complete and honest answers. The difference is in how the question is framed — and who is doing the asking. Communities most affected by AI bias are best positioned to identify and correct it.

Young people from the communities most likely to be underrepresented in AI outputs are not just the beneficiaries of this work. They are the most qualified to do it. That is the civic premise of Civics Remix.
Black in AI — blackinai.github.io ↗

Distributed AI Research Institute (DAIR)

Timnit Gebru — Independent AI research centered on communities most affected by AI bias

Born in Addis Ababa, Gebru founded DAIR after being ousted from Google for publishing research on AI bias. Her core argument: when AI systems are built without the voices of affected communities, they are more likely to tell those communities' stories incompletely — until those communities demand better.

Demanding more complete answers is a civic act. Teaching young people to demand better at scale is what Civics Remix exists to do.
DAIR Institute — dair-institute.org ↗

What Students Use to Investigate AI Bias in Real Time

The AI Bias Detector is not a product. It is what students do — in every session — using real, publicly available tools. These four tools make bias visible, measurable, and discussable. Students query them, compare outputs, and document what they find using the Four-Filter Protocol.

How Civics Remix uses these tools: Facilitators guide students to run the same query across different communities or phrasings, compare the results, and apply the Four-Filter Protocol to name the bias, identify whose story is missing, and document the civic consequence. No account or download required for Gender Shades, AI Incident Database, or Hugging Face Model Cards. The Perspective API demo is viewable without an account. Live querying may require a free API key — contact us at RockStars4Impact@gmail.com and we will help you set it up.
🔬

MIT Media Lab

Gender Shades

Timnit Gebru and Joy Buolamwini's documented investigation of facial recognition AI bias across skin tones and gender. Real data. Real disparities. The research that helped launch the AI ethics field.

Students examine accuracy rates across demographic groups and ask: whose face did the AI learn from? Who was left out of the training data — and what are the consequences?
gendershades.org ↗
📋

Partnership on AI

AI Incident Database

A running public database of documented cases where AI systems caused real harm — in hiring, criminal justice, healthcare, education, and more. Every entry is a real civic consequence of AI bias.

Students search by category — criminal justice, education, healthcare — and find specific incidents where AI bias produced a civic consequence. The harm is not hypothetical. It is documented.
incidentdatabase.ai ↗
⚖️

Google Jigsaw

Perspective API

A live AI tool that scores text for toxicity. Students input the same sentence written about different communities and compare the scores in real time. Bias becomes visible and measurable in seconds.

Students write one sentence about a community they know, then rewrite it substituting a different community. They compare toxicity scores and ask: what does the AI assume? Whose language is treated as more threatening?
perspectiveapi.com ↗
📄

Hugging Face

Model Cards

Every major AI model published on Hugging Face includes a Model Card — a document written by the builders themselves that names known limitations, biases, and gaps in the training data.

Students read what AI builders themselves document as missing or skewed. They ask: if the people who built this acknowledge it is incomplete, what does that mean for the decisions being made with it?
huggingface.co/models ↗

The Civics Remix Method

Every tool is evaluated through the Four-Filter Protocol.

Students do not just notice bias. They name it, locate its civic consequence, and document their finding. This is the Four-Filter Protocol — applied to every AI output in every session.

Filter 01

AI Ethics

“What bias does this answer carry?”

Filter 02

Humanities

“Whose story is absent from this answer?”

Filter 03

Trust

“Would a civic institution rely on this?”

Filter 04

Civic Consequence

“What decision gets made wrong because of this gap?”

Civics Remix was built to address exactly what the research identifies.

Students learn to recognize when AI may be missing part of the story, identify the bias behind it, and solve for something more complete through the lens of AI ethics. They do not just learn about the problem. They become the people equipped to address it. That is what a changemaker looks like.

Critical thinking Prompt engineering AI ethics and bias Harkness discussion Classical and cultural texts Peer facilitation Research Civic action

From the Classroom

“The use of Civics Remix sparked excitement and engaging conversation between fifth graders as they explored the use of AI. Students were intrigued by the discrepancies between AI generated responses and primary sources, reinforcing the need to consider multiple sources without reliance on technology.”

Ms. Felicia Baskin, 5th Grade Teacher, Naomi L. Brooks Elementary School

Sources

Tao et al. — Cultural bias and cultural alignment of large language models. PNAS Nexus, Vol. 3, Issue 9, 2024  ·  academic.oup.com ↗

Andreason, Sigelman et al. — Which Skills Matter Now? aiEDU / Burning Glass Institute, February 2026  ·  burningglassinstitute.org ↗

Rediet Abebe — Black in AI  ·  blackinai.github.io ↗

Timnit Gebru — Distributed AI Research Institute (DAIR), 2021  ·  dair-institute.org ↗