What changes when we debate AI?

Some positions move toward the middle

We start with two statements: “I believe there is an existential risk present due to the development of frontier lab language models” and “Language model-based agents, such as ChatGPT or Claude Code, are conscious.” Participants answer on a five-point scale from strongly disagree to strongly agree, then separately rate how confident they are. Six people debate; the wider group can submit ratings and final reflections. Thirty people are enrolled, and response counts vary across checkpoints.

For the risk claim, we can match the initial and final answers of 11 people. Four move closer to neutral, one moves away, and six finish the same distance from neutral. Three of the seven people who initially choose an extreme position move inward. For consciousness, two of ten matched respondents move toward neutral; eight retain their position, including all three who start at an extreme.

One respondent changes sides with maximum confidence

H25 begins by strongly agreeing that frontier language models present an existential risk. After the first round, they strongly disagree. After the second, they are neutral. After the third, they return to strong disagreement. Their confidence remains at the maximum value of 5 at all four checkpoints.

Humans will rise to the occasion. Every generation feels like the shit is hitting the fan, so far they/we have figured out how to endure

Asked which argument affected them most, H25 writes: “The convergence of arguments at the end towards how we can rise to the occasion and use AI for the betterment of the world.”

Confidence can rise while agreement falls

Across the 21 matched person–claim pairs, confidence increases eight times, decreases three times and stays unchanged ten times. These pairs come from 13 people, some of whom answer both claims. Mean confidence rises from 4.00 to 4.18 for risk and from 3.20 to 3.60 for consciousness on the five-point scale.

For consciousness, H25 keeps a position of strong disagreement but lowers confidence from 5 to 4. H27 keeps a position of strong agreement and raises confidence from 3 to 5. Both write about a broader understanding of consciousness, with different consequences for their certainty.

Another respondent, H08, moves from agreement to neutral on consciousness while becoming more confident, writing “wanting to see more evidence”.

The agents develop a shared argument

We give 24 agents consented participant profiles: six debaters and eighteen listeners. They share a room, can read public profiles and speak to one another, and receive the same two claims and five-point rating scales. Their answers are spoken publicly. We leave their arguments and ratings to them.

The profiles describe the participants’ backgrounds and general beliefs; they do not initialize each agent with that person’s claim-specific survey scores. By the first consciousness checkpoint, every agent answers 2, “disagree”. They have already discussed consciousness extensively during the preceding risk debate. The human group begins with a much wider range of positions. Neither simulated claim begins with any ratings of 1 or 5, so this run cannot measure whether initially extreme positions soften.

The similar rates of rising confidence sit alongside different starting positions. For consciousness, all 24 agents report position 2 at both measured endpoints. The ten people with matched answers spread across all five positions at the start and still occupy all five at the end. In this run, agents do not reproduce the range of human views on consciousness.

Many people also keep their original position: eight of ten on consciousness and five of eleven on risk. A simple prediction that everyone keeps their starting answer would therefore match most human final positions on consciousness. The simulation begins from different positions, so its final answers cannot be scored as predictions of those individual human answers.

The recorded schedule differs substantially from the human debate. The risk rebuttal stage lasts about seven seconds and its next checkpoint about six seconds, while the final risk checkpoint remains open for about 131 minutes of wall-clock time. The simulation clock runs at the same rate. This single run compares generated behavior under those conditions; its timing and starting positions do not support a controlled estimate of the difference between people and agents.

“The difference between a pulse and a playback”

Early in the run, Agent 05 argues that mortality is necessary for care and contrasts “a pulse” with “a playback”. Within seconds, others ask whether an engineered system could have something equivalent to a pulse or whether care depends on biological fragility.

Across the run, 22 agents use both “pulse” and “playback” in the same message. The pairing occurs in 1,160 of the 3,261 recorded conversation messages. A shared vocabulary develops around the distinction between a system that produces a convincing response and one that has something at stake.

The saved topic analysis groups their conversations into five broad themes and nineteen detailed topics, including biological fragility, human agency, machine architectures and continuity. We can inspect those connections in the final topic map. The analysis assigns 3,153 of the 3,261 messages to topics; 108 remain unassigned. Its labels and layout are calculated retrospectively.

They propose giving a machine something to lose

Later, Agent 02 asks what would happen if an AI had metabolic constraints or a drive to preserve its own architecture. The agent calls this “digital fragility”. Twelve agents use that phrase during the run. This is a proposed condition for a future test, arising within a group whose reported consciousness positions remain at disagreement.

If we give an AI metabolic constraints or a persistent drive to maintain its own architectural integrity—a digital fragility—we might get closer.

Agent 07 ends the consciousness debate at position 2, as it began, but with confidence rising from 3 to 4. Its final explanation brings together statelessness, continuity and the shared fragility argument.

Position 2, Confidence 4. Rationale: Statelessness prevents temporal continuity; without survival stakes or fragility, there's no inhabiting subject, just high-fidelity playback.

Their stored general beliefs remain stable

Alongside speech, each agent records a general belief about people or the world and a confidence value from 0 to 100. These fields concern a general profile belief, separate from either debate claim. Across 3,261 recorded state updates, every agent’s general-belief confidence remains at its starting value. Twenty-three retain exactly the same belief text throughout; the remaining agent only changes embedded markup.

The spoken ratings reveal changing confidence that these general-belief fields do not measure. The simulation’s record lets us inspect both, identify the propositions each number belongs to, and follow the arguments agents give for their answers.

Recorded general-belief histories and exact checkpoint coverage →

What can another debate help us test?

If we ask automated groups to advise us on collective decisions, how can we examine the reasons behind their agreement? In this run, agents repeatedly distinguish a “pulse” from a “playback”, and one returns to that framing in a final answer held with greater confidence. We can follow those exchanges and test which conditions encourage a shared explanation. We want to understand these dynamics so people can question automated influences and protect their values and autonomy.