# What changes when we debate AI?

0–6s: We bring people together at the CIMC residency at Edgecity to debate AI. 30 people enroll, including six debaters · 29 October 2026
Human debate: 30 people enrolled; six debaters and twenty-four other enrollees.

6–12s: How does debate change strongly held beliefs? Can we model how discussion changes their beliefs?
The two debate topics are existential risk and machine consciousness.

12–19s: Opening views. Rebuttals. Open discussion. We ask for position and confidence before discussion and after every round.
Four questionnaires per claim: initial beliefs, after opening views, after rebuttals, after open discussion.

19–35s: I believe there is an existential risk present due to the development of frontier lab language models Claim 1 · Existential risk
Initial: 3 strongly disagree, 2 disagree, 3 neutral, 4 agree, 7 strongly agree; 19 responses. After opening views: 3, 4, 4, 6, 2; 19 responses. After rebuttals: 1, 3, 5, 6, 3; 18 responses. After open discussion: 2, 2, 3, 3, 2; 12 responses. All five bins run from strongly disagree to strongly agree. Respondents vary across checkpoints; these distributions are not matched-person changes.

35–51s: Language model-based agents, such as ChatGPT or Claude Code, are conscious. Claim 2 · Machine consciousness
Initial: 2 strongly disagree, 2 disagree, 3 neutral, 3 agree, 2 strongly agree; 12 responses. After opening views: 4, 4, 1, 3, 2; 14 responses. After rebuttals: 3, 3, 5, 1, 1; 13 responses. After open discussion: 3, 3, 4, 2, 2; 14 responses. All five bins run from strongly disagree to strongly agree. Respondents vary across checkpoints; these distributions are not matched-person changes.

51–58s: Three of ten initially extreme positions move toward neutral. Matched person–claim pairs · Initial positions of 1 or 5
Three of ten initially extreme person–claim pairs move toward neutral: risk 3 of 7, consciousness 0 of 3. Seven retain their distance from neutral, including one full reversal.

58–69s: One respondent takes a different path. Their confidence stays at the maximum throughout.
H25, existential risk. Position at the four checkpoints: 5, 1, 3, 1. Confidence: 5, 5, 5, 5.

69–75s:  H25 finishes at position 1, strongly disagree, with confidence 5, extremely confident.
H25 · Final written reflection on risk: Humans will rise to the occasion.

75–82s: Confidence rises more often than it falls. 21 matched person–claim pairs from 13 people
Confidence increases in 8 person–claim pairs, decreases in 3, and stays unchanged in 10.

82–90s: We give 24 agents participant profiles and the same two claims. In this separate simulation, their answers are spoken publicly.
Simulated population: 24 agents, comprising six debaters and eighteen listeners. Participant profiles initialize the agents; claim ratings are generated during the run.

90–98s: Six debaters and eighteen listeners take part. The 24 simulated agents, shown using their saved portraits and anonymous study identifiers.

98–114s: They debate what makes care possible. Native 3D conversation replay of the 24 saved agents in the Debate Room. Agent 05: “It’s the difference between a pulse and a playback.” Agent 02 replies 8.5 simulated seconds later: “Can we build a pulse, or is the biological substrate the only thing that makes it real?” Agent 10 then asks about care. Characters use the viewer’s standing positions and animation, not historical body trajectories.

114–121s: The distinction recurs as the debate continues. 22 agents eventually use “pulse” and “playback” together.
Recorded first uses of pulse and playback together, from elapsed 00:03:00 to 02:02:40. Twenty-two of twenty-four agents use both words in the same message.

121–129s:  Recorded simulated dialogue at elapsed 02:49:56. This is a proposed condition, not a completed experiment.
Agent 02 · Proposing a way to give AI something to lose: metabolic constraints or a persistent drive to maintain its own architectural integrity—a digital fragility

129–137s:  Agent 07 at elapsed 02:56:21. Position remains 2; confidence rises from 3 at the initial checkpoint to 4 in the final answer.
Agent 07 · Final consciousness answer: without survival stakes or fragility, there's no inhabiting subject, just high-fidelity playback.

137–167s:  Three hours, six minutes and 53 seconds of simulation time. 24 agents produce 3,261 conversation messages and 5,542 agent actions; the full record contains 5,548 events including six system events. Each timeline row shows one agent’s messages. The orchestration list highlights the current stage using observed transitions mapped to the simulation clock. Live topic analysis reveals topics at their recorded first appearance, reaching 1,199 topics. The native graph prioritizes topics for readability; weights and relationships describe the completed run, and positions use the viewer layout.

167–173s: The fragility argument recurs across the group. Biological fragility · 18 contributing agents
The selected topic is biological fragility as a requirement for care. The saved analysis attributes 353 mentions to this topic from 18 contributors. Topic membership does not establish endorsement or persuasion.

173–180s: All 24 disagree at the initial and final consciousness checkpoints. Eleven become more confident; none becomes less confident.
Consciousness confidence: 11 agents increase, 13 stay unchanged, zero decrease. All 24 retain position 2, disagree, between the initial and final checkpoints.

180–185s: Their general-belief confidence stays fixed. 24 agents · Separate profile beliefs, rated 0–100
All 24 general-belief confidence histories stay flat across 3,261 recorded state updates. This separate measure ranges from 0 to 100.

185–191s: The simulated debate has different starting positions and timing. No extreme initial ratings · Consciousness baseline follows earlier discussion
For risk, 4 of 11 matched human respondents and 1 of 24 agents move toward neutral. Baselines, timing and response conditions differ. The simulation’s first consciousness checkpoint follows extensive prior discussion; its final risk checkpoint lasts approximately 131 minutes.

191–199s: Could AI-mediated discussion amplify shared explanations and make groups more certain? How might that shape collective decisions?
At the end of the run, 22 of 24 agents have used pulse and playback together. Repetition does not establish that the phrase caused a confidence change.

199–205s: We study these dynamics so people can question automated influences. And protect their values and autonomy.
The comparison describes recorded behavior. A proposed follow-up would compare public and private ratings and test the results against further human debates.

205–209s: Explore the debate study. Contact Emergentic about your question.

Music: Parag K. Mital - Other Intelligences
https://soundcloud.com/pkmital/mr-magoo-other-intelligences