Looking back
We started this project in 2023, before the wave of AI slop that now fills social feeds. At the time, AI hosts that write their own scripts and publish their own videos were mostly hypothetical. Our bet was that they were coming, and that journalism should understand them from the inside before they arrived. That bet turned out to be right.
The videos from that era look dated now, and some of them are frankly cringeworthy. We’re showing them anyway, because finding the cringe was part of the work. The students learned where AI genuinely helps tell a story, and exactly where it crosses into the uncanny valley.
The question
Younger audiences get much of their news from short videos hosted by people they feel they know. Newsrooms have struggled to reach those fragmented, niche audiences in that format. We wanted to know whether AI could help, and what it would mean if it did. Can an AI-made video be informative and entertaining? How do viewers respond to a host they know is an AI? And what are the ethics of giving a language model a personality and opinions?
First experiments
In Fall 2023 the first team tried everything that existed: animated stills from Stable Diffusion, early text-to-video tools like Pika and Genmo, and talking-head generators that lip-synced a portrait to a voice track. The results were uncanny and inconsistent, but they proved the pieces could be chained together, and they gave the next team a clear target. The missing ingredient was not better video. It was a character the audience could recognize from one clip to the next.
Characters that write in their own voice
In Winter 2024 the team ran Llama 2 70B on a server in the lab, alongside a local text-to-speech model, so students could experiment without per-request costs. They designed a cast of hosts, a farmer, a nurse, a news anchor, a children’s storyteller, a big-city executive, each with a backstory, a way of talking and an audience they’d speak to. Then they fed every character the same local story and compared what came back.
A short Evanston story about the line outside Bennison’s Bakery on Paczki Day became, in the voice of a warm grandmotherly host, something like this:
Good afternoon, my dears! Bennison's Bakery in Evanston is the place to be, especially on Paczki Day… Trust me, darlings, your taste buds will thank you!
Script written by the Llama 2 pipeline from a local news story, Winter 2024
The facts survived; the delivery changed completely. That was the point. The team built a small web front end, Newsfluencer, that showed the original story next to each character’s script and the video it produced.
A fully automated agent
By Spring 2025 the project had become an autonomous agent built in n8n. It pulls an article from a news API and summarizes it, then an “editor” step chooses two of ten personas whose audiences most need the story. Each persona writes a 90 to 120 word spoken script, paced for 30 to 45 seconds, with a hook in the first six to eight words. The agent scores its own draft for hook, specificity and tone, revises it, sends it to an avatar and voice service to render and uploads the finished video. In practice the scripts still needed a human editor before they were ready, which was one of the findings.
What we learned
The teams compared “shallow” personas, sketched with only a few traits, against “deep” ones with fuller backstories and values. The deep personas argued with more nuance and stayed more consistent from video to video. Personality turned out to be more than decoration: it changes how an AI reasons about the news it’s given.
The students also documented the uncomfortable parts. Asked to describe each host’s audience, the model kept producing an audience’s “core emotional need,” and writing to that need made the scripts noticeably stronger, which the students themselves flagged as lightly exploitative. The team debated whether matching stories to personas would deepen filter bubbles, and settled on using the editor step to send each audience what was missing from its media diet rather than more of what it already liked. Those questions, how to keep many voices grounded in the same verified facts, and how to know how each voice will land, are what led directly to Bedrock and Mirror.
Looking at today’s feeds, the project’s central finding holds up. Generating the video was never the hard part. The value was in the reporting underneath it and in knowing the audience, and without those an AI host is just another face in the uncanny valley.