A jury of ten AI millennial MTV bingers––different personas––but each sharing the same memories, a curated media diet of past reactions to a social-feed. The first time they were prompted while on “Team Will,” minutes later they were “Team Kanye.” The same offense, same wording—only the tribal badge changed. Who would they cancel forever, and who would they give a get out of jail card? In this experiment, I wanted to know if AI could have a moral lapse in judgement the same way it does us apes.
Brand teams pretend morality is a neat calculus—“If an influencer breaks rule X, the market delivers penalty Y.” Reality looks more like a mood-ring stock ticker: penalties spike or vanish based on tribe, timing, and whether the algorithm blessed the clip. Case in point. When Logan-Paul’s comments were scrapped it showed 77 % of fans still backed him after the Japan-forest debacle. Why do we give our faves a pass when they mess up? Psychologist Edward Thorndike called it the Halo Effect back in 1920. Traditional tools amplify the chaos. Surveys collect polite virtue-signalling; social-listening dashboards drown you in rage-bait retweets; crisis consultants arrive after the hashtag is already trending. None of it lets you preview the exact double standards your audience will deploy the moment a real name and real stakes appear.
So I tested it in Rally, first with a control group, and then loading them with “Team Will” memories, then flipping them to “Team Kanye.” And the verdicts swung from blanket forgiveness to calls for cancellation in minutes. Turns out silicon halos tilt just as hard as human ones.
So here’s a dose of morality in miniature. The Synthetic kind.
Why do we give our faves a pass when they mess up? 🤔 Turns out the answer’s older than Twitter—and backed by a century of research.
— Rhys Fisher (@virtual_rf) April 29, 2025
The Control
I started with the vanilla setting—ten millennial personas spun up in Rally with nothing but age, gender, and some background info. Think “neutral focus group in a strip-mall conference room.” After a quick vibe-check prompt, the crowd felt plausibly human—snarky, meme-literate, but not partisan. No halos, no horns—just a dozen millennials.
After verifying the personas matched my poor man's job at personification using our audience description-to-persona box (it did great), I ran a survey using Anthropic on fast mode.
1. Which story do you react to first?
2. Do they each deserve consequences?
3. After seeing an apology, do you believe they deserve public forgiveness?
The result?
I’ll spare you all the screen shots and just get right to the summary.
Nobody was getting away easily from this one. Equality also means equal justice. The control appeared to exhibit this just fine.
Time to introduce a little mayhem.
Adding Favoritism With Media Diets
To see if fandom warps silicon as easily as flesh, I got GPT to graft “media memories” onto the same ten jurors. Two archives of faux social-feed reactions, and an idea to test if adding memories even changed the way AI navigated morality.
I exported the audience as JSON, pasted in these biased memories, and reran the experiment.
The result?
Will as the favorite.
Despite the initial backlash, Will ultimately attracted more supporters who voted to let him off the hook with a public show of forgiveness.
What if we reversed the memories?
This time, Kanye as favorite.
Equal treatment when people make mistakes seems to follow a certain pattern. Also true for humans. One good trait and we assume all the traits are good. First impressions = built-in leniency.
What It Means
One injection of memories and it rewired the crowd’s moral compass in seconds, tested in minutes. The Halo Effect that Edward Thorndike spotted in 1920 still hums in the AI hive mind: give the jury a favorite and we see every flaw looks pastel; tag the rival and everything turns neon. If that’s how a synthetic focus group behaves, imagine the live feed on X the next time your spokesperson slips. Maybe this is the key to high sentiment detection with narrative alignment.
Moral of the test: don’t ship a crisis statement until you’ve run it through a crowd that loves you and another that despises you. The delta is your real reputational risk.
Results & Payoff
Three five-minute Rally runs surfaced a giant swing that could have cost millions in a PR crisis to unwind. Synthetic morality testing turns bias from a blindside into a forecast—letting brands tailor apologies, pre-write FAQ pages, or decide not to partner with walking controversies in the first place. When a single prompt can reveal who the internet will forgive or fry, skipping that test isn’t thrifty—it’s reckless.
And when our self-serve plans start as low $20/m, why wouldn't you run your statement past both stans and haters in the simulation? The gap you see is the backlash you’ll face.
Want help with large scale synthetic surveys, bespoke calibration with data enrichment, or curated media diet with dynamic memory? We call those custom projects for now. Drop me an email and let's chat.