NeurIPS 2026 · Creative AI Track

Whispers:

An Interactive AI Thriller Series

Bernie Su, Vida Adeli, Soroush Mehraban, Cole Clifford

Pickford · {bernie,vida,soroush,cole}@pickford.ai

Abstract

Abstract

Detective Marcus Kent.
Detective Marcus Kent

Whispers is an animated, episodic interactive AI thriller series in which the audience is not a spectator but a participant in the story itself. Each episode follows Detective Marcus Kent, a man who hears whispers that help him solve crimes. These whispers are, in fact, the live audience.

Dialog, scenes, and story beats are written and rendered in near real time by AI operating within a strict narrative framework, so that the audience can shape a murder investigation as it unfolds. The work is an experiment in distributed narrative agency, where authorship is continuously negotiated between a crowd, a character, and the humans who built and voice the world. Whispers is currently being shown in theaters across the US. It was created by three-time Emmy winner Bernie Su, and is a finalist for the 2026 Emmy for Outstanding Innovation in Emerging Media.

A crowdthe live audience
A characterDetective Marcus Kent
The humanswho built and voice the world
1

The work

Whispers is an interactive animated episodic AI thriller series, available on pickford.ai. Each episode runs approximately 45 minutes and unfolds as a murder investigation led by Detective Marcus Kent, a man with an extraordinary gift: he hears whispers, and they help him solve crimes. But the whispers Kent hears are provided by the audience.

Detective Kent on a hotel staircase as an audience whisper appears on screen: “Jessica was using Nathan to get close to Autumn!”
From an episode An audience whisper appears on screen as Kent hears it.
~45minutes per episode
Murderinvestigation, led by Detective Marcus Kent
Near real timedialog, scenes, and story beats written and rendered by AI

Though in concept the story may resemble a traditional detective series, Whispers is far from it. The dialog, scenes, and story beats are all written and rendered in near real time by AI within a strict narrative framework as each episode plays. Because of this method, the audience can change what is being made as the episode airs, far beyond what a classic branching narrative is capable of.

Classic branching narrative

Paths written in advance.

Whispers

Written and rendered as each episode plays; the audience can change what is being made.

Despite being an AI-powered animated series, every voice in the show belongs to a working human professional, cast and contracted under SAG-AFTRA’s newly established AI performance frameworks. Their credits span traditional projects across film and television. The future of storytelling does not have to come at the expense of the people who tell stories.

Detective Kent at the hotel lobby crime scene.
A character in a hotel corridor.
Detective Kent in the precinct bullpen.

Every voice: a working human professional, cast under SAG-AFTRA’s AI performance frameworks.

2

Agency

Interactive storytelling has existed for decades, but it has largely equated agency with choice: a poll appears, the story branches, the show waits for a click and then resumes. Whispers pursues a different, more unsettling kind of agency. One that is continuous, collective, and bounded. We see this as a small study of how agency actually behaves when it is shared.

Agency as choice

story a poll appears the show waits for a click story branches · resumes poll wait resumes

Whispers

the story never waits for the audience
No menu, no vote screen, no pause. Influence arrives through a continuous stream of chat.

Continuous

Most importantly, the audience influences the story without ever leaving it and the story never waits for the audience. There is no menu, no vote screen, no pause. Viewers exert real narrative force through a continuous stream of chat, so participation is seamless rather than transactional. It is an implicit guidance rather than explicit action.

Collective

Additionally, no single viewer steers Whispers. The system continuously ingests the audience’s communication and reads its emotional patterns in aggregate; those signals feed Kent’s instinctive reading of a scene. When the crowd collectively leans toward a suspect, Kent leans that way, too. Not because a vote said so, but because the weight of the whispers moved him. Agency emerges from the many, and no individual can fully claim it.

Bounded

Crucially, audience agency is not absolute. The audience can whisper anything, but those whispers do not reach Kent unfiltered; they pass through character guardrails and the constraints of the investigation. Kent’s ethics, instincts, and investigative focus are encoded as hard constraints, so he will not leave the case, commit felonies, or betray who he is, no matter what the crowd wants. The audience shapes his direction, but cannot reshape his character. Agency in Whispers is therefore perpetually negotiated between the crowd’s desire, the detective’s integrity, and the human authors who set the boundaries of the world.

3

The role of AI and machine learning

AI and machine learning are not a production shortcut in Whispers; they are the mechanism that makes distributed real-time agency possible. Three components carry that load.

Story Kent, the case Canonical memory World-building details Audience input Pickford Platform Audio generation Scene Visualizer Audience whispers loop back in
System overview Audience input flows into the Pickford platform alongside the story, world-building details, and canonical memory; scenes are generated, rendered, and played to the audience, whose whispers loop back in.

01 Dynamically written scenes.

Instead of pre-writing entire scenes, Whispers takes a very different approach. Every dialog scene in the series is generated in near real time by a large language model operating within the strict narrative procedure of a murder investigation. This allows the audience’s whispers to be injected into the story as quickly as possible.

Thus when Detective Kent speaks to suspects, partners, and witnesses, almost none of it is pre-written. It is constructed, in the moment, from the investigation process, the parameters of the scene, and the weight of the audience’s influence. The dialog is never the same across any two viewings.

the investigation process the parameters of the scene the weight of the audience’s influence
Large language model within the strict narrative procedure of a murder investigation
Kent’s dialog with suspects, partners, and witnesses, constructed in the moment
Viewing A ≠ Viewing B: the dialog is never the same across any two viewings.

02 Collective emotional guidance.

Most interactive experiences ask the audience to register input. A poll appears. You make a choice. The story shifts. But in whispers it is near seamless for the viewers. The show doesn’t stop, and it doesn’t ask. The story just flows.

Through the whispers, we continuously ingest the audience’s communication with the system, understanding their emotional patterns at all times. These signals feed directly into Kent’s instinctual reading of a scene. When the audience collectively leans toward a suspect, Kent leans that way too. Not because a vote said so, but because the collective whispers have gotten him to lean in that direction.

Thus, when it’s time to choose a suspect for interrogation, it isn’t a series of clicks or prompts at a specific moment. It’s the analysis of all the whispers up until that point that allows Kent to make the most informed choice.

Kent’s lean, from all whispers so far

Autumn
Nathan
Cho

Time to choose a suspect for interrogation: —

Illustration Whisper text taken from a theater screening; the weighting shown is a simplified illustration, not the production model.

03 The narrative integrity engine.

Even though the audience can whisper anything, no matter how emotional, obscure, or eccentric; those whispers don’t reach Detective Kent unfiltered. They pass through his defined character and the setting of the murder investigation.

Kent’s ethics, instincts, and investigative focus are encoded as hard constraints on the story universe model. He’s a righteous detective on the hunt for justice, and whispers that suggest he veers from that path in any way won’t be considered by him. He won’t leave the investigation, he won’t commit felonies, and he won’t indulge in personal desires. The audience shapes his direction, but they cannot reshape his character. This ensures that the story maintains its integrity and that the characters remain consistent to the story world.

The audience can whisper anything

  • autumn is hiding something, press her!
  • Forget the case, Kent. Quit the force and open a taco truck.
  • Steal the murder weapon and sell it online
  • Tell her you think Cho did it
  • Ask the suspect out on a date. Right now. In the interrogation room.
  • Rob the hotel safe while everyone’s distracted
  • Drop everything and fly to Vegas, the case can wait
  • Check the fingerprints on the glass

Hard constraints

  • He won’t leave the investigation
  • He won’t commit felonies
  • He won’t indulge in personal desires

Kent’s ethics, instincts, and investigative focus

Reaches Kent

    Won’t be considered by him

      Illustration Example whispers; the sorting shown illustrates the hard constraints described in the paper.
      4

      Artistic goals

      The series is not simply an interactive crime procedural; it is a meditation on the gap between what we believe in private and what we create in public. The goal is a form of storytelling where AI does not replace the creative act but enables one that was previously impossible. Not AI as a shortcut, but as a dance with a human creator to allow a previously impossible experience.

      The combination of custom designed systems gives Whispers the familiarity of a thrilling scripted detective series while still inviting the audience to be consequential and shape the story. More concisely, we allow the audience to scream at the TV and sometimes the TV might scream back.

      Photo A Whispers audience at a theater screening.

      Author biographies

      Bernie Su

      Bernie Su is Head of Creative at Pickford. A three-time Primetime Emmy and Peabody Award winning creator and showrunner, he has spent over a decade pioneering interactive storytelling, earning the first Emmy awarded to a YouTube series and the first Emmy and Peabody awarded to a Twitch original. Relevant prior work includes The Lizzie Bennet Diaries and Artificial.

      Vida Adeli

      Vida Adeli is a Research Scientist at Pickford. She holds a PhD in Computer Science from the University of Toronto, where her research focused on generative AI, computer vision, 3D human modeling, and 3D character animation, video and scene generation for interactive storytelling. Relevant prior work includes PickStyle, a context-preserving video style-transfer framework, Puppeteer, an object- and posture-aware co-speech gesture generation system, and GAITGen, a generative model for synthesizing realistic 3D human gait.

      Soroush Mehraban

      Soroush Mehraban is a Research Scientist at Pickford. He is completing a PhD in Biomedical Engineering at the University of Toronto, where his research focuses on generative AI, computer vision, 3D human modeling, human motion analysis, and video generation for interactive storytelling. Relevant prior work includes PickStyle, a context-preserving video style-transfer framework; Puppeteer, an object- and posture-aware co-speech gesture generation system; MotionAGFormer, a transformer-based framework for 3D human pose estimation; and FastHMR, a fast and accurate monocular human mesh recovery framework.

      Cole Clifford

      Cole Clifford is a co-founder at Pickford. He has spent 10 years putting state-of-the-art ML research into production and generally exploring how to use neural networks in interesting ways. Prior work ranges from astronomy in space2vec, to unique embedding methods in NWT: Towards natural audio-to-video generation with representation learning.

      Citation

      BibTeX

      @inproceedings{su2026whispers,
        title     = {Whispers: An Interactive AI Thriller Series},
        author    = {Su, Bernie and Adeli, Vida and Mehraban, Soroush and Clifford, Cole},
        booktitle = {NeurIPS 2026 Creative AI Track},
        year      = {2026},
        url       = {https://openreview.net/forum?id=9QPGiyCW7j}
      }