How AI Agents Shift Opinions Like Real People — The Mechanics of Starling

The previous post introduced the possibility — that dozens of AI agents can show you the outcome before you decide (read part 1). The natural next question is "how is this actually possible?"

This post walks through the mechanics of AI opinion simulation step by step — the flow from input to output, and why this structure is fundamentally different from automating a survey.

1. Input — What You Provide

Two things are needed to start a simulation.

• Topic and context — what you want to see reactions to
• Population composition — whose reactions you want to see

For example, the input for the Galaxy S25 simulation looked like this:

■ Topic
- Lineup: S25 ($799), S25+ ($999), S25 Ultra ($1,299+)
- S24 → S25 spec changes (battery unchanged, chipset +37%)
- Competitor specs: iPhone 16 Pro Max, OnePlus 13, Xiaomi 15 Ultra

■ Population
- Year 1 (S24 users): early adopters ~15%
- Year 2 (S23 users): considering replacement ~30%
- Year 3 (S22 users): replacement window ~30%
- Year 4+: performance degradation ~25%

■ Simulation question
"How will consumers react to the Galaxy S25?"

The key here is that information that gives away the answer is never included. Pre-order numbers, sales data, reviews — these poison the model. The model must reason from facts alone, not regurgitate the answer fed to it. That is what makes the simulation verifiable.

2. Building an AI Persona

Each agent is a virtual person. Starling generates N agents from the input population spec. Each agent carries:

• Demographics — age, gender, occupation, region, income
• Personality — extraversion, conscientiousness, openness (based on the Big Five model)
• Background — values, interests, prior experience cues

This information forms the basis of "how this person would react." Two agents with the same demographics react differently because of personality and background. That is what produces a distribution of personas rather than an averaged single voice.

Dozens of agents are dozens of different people. Some are eager about new things; others are skeptical. Some are price-sensitive; others are brand-loyal. This diversity is what produces opinion.

3. Multi-Round Interaction in a Social-Media Environment

This is the crucial piece. Starling agents don''t simply "answer a question." They interact with each other across time in an environment that resembles a real social media platform.

In each round, agents:

• Post their opinions
• Comment on others'' posts
• Follow people they like
• Like posts to express agreement
• See new posts surfaced by the recommendation system

At the end of each round, agents integrate what they''ve seen and update their opinions. They get pulled by a strong individual argument; they get reinforced when others in their group agree.

Running 15 rounds is like spending several days on social media. During that time, opinions form, camps split, and influencers emerge. The very process of opinion formation is being simulated.

4. Opinion Change — Dynamics, Not Just Aggregation

Surveys aggregate one-time scores. Starling tracks opinion flow over time.

Tracked outputs:

• Sentiment trajectory — round-by-round positive/negative ratios
• Influence distribution — whose posts spread the most
• Echo chamber formation — which groups isolate and reinforce
• Inflection points — rounds where opinion shifted sharply, and the trigger
• Viral pathways — the network graph of how a post propagated

This isn''t just "what people said" — it''s the record of "how the opinion was made." The same decision splits differently depending on which influencer in which camp said what in which round. You need the whole process visible to trust the result.

5. Output — What You Get

When all rounds complete, the following results are produced.

1. Sentiment distribution — final positive/negative/neutral ratios
2. Purchase intent / approval — T2B (definitely/probably will) vs B2B (definitely/probably won''t)
3. Top criticism / interest topics — recurring keywords and their intensity
4. Influencer ranking — top-influence agents and their arguments
5. Opinion-flow visualization — round-by-round change graphs
6. Action recommendations — next-step suggestions based on the result

All of this comes out of a single simulation. What would otherwise take several separate surveys arrives at once.

6. Validation — Does It Actually Work?

The theory sounds plausible, but how did it perform in practice? Two cases in more detail than part 1.

Galaxy S25 (January 2025)

• Conditions — 25 agents, 15 rounds, 3 repetitions
• Inputs — specs, price, competitors, age-by-replacement-cycle
• Excluded — pre-order numbers, sales, reviews

Results (3-run average):

• Sentiment — positive 51% / negative 38% / neutral 11%
• Industry benchmark — positive 50–60% / negative 35–45% (smartphone category internal baseline)
• Within range
• Top criticism topics (5) — battery, charging, S Pen, RAM, camera
• Actual post-launch reviewer/community criticism — same

The meaningful part is that the pre-launch simulation pinned down the exact pressure points that later emerged in real reviews. Knowing what will be criticized in advance is the most actionable output for marketing — it lets you decide what to message and which weaknesses to address up front.

Thailand 2026 Election (February 2026)

• Conditions — Parliamentary dissolution and election occurred after the model''s training data cutoff (November 2024). The model did not know the answer.
• Inputs — border-conflict timeline, party platforms
• Excluded — poll numbers, election results

Results:

• Starling''s 1st: Bhumjaithai (matches actual 1st)
• Starling''s 2nd: People''s Party (Prachachon) (matches actual 2nd)
• Polls had the 1st and 2nd reversed; the simulation got it right

This isn''t coincidence, for a simple reason. Polls measure "where opinion is now." Simulation models "where opinion will go." For decisions that depend on the future, you need the second tool.

7. Limits — Honestly

AI opinion simulation isn''t a universal tool. What it does poorly:

• Sensory reactions — taste, smell, texture; these require real experience and cannot be simulated
• Legally binding research — medical devices, financial products, etc. require validated human panels
• Model''s own bias — biases baked into the LLM''s training data can show up in the result
• Rare or emerging cultures — accuracy drops for regions or generations the training data underrepresents

The recommendation, therefore, is not "AI simulation alone" but "AI simulation as the first-pass screen, supplemented by other methods where the decision cost is high." The simulation establishes the direction; further validation handles the high-stakes details.

If you want to try it yourself, sign up for the free tier — credits are granted immediately on signup.