Your feed, your chats, a video someone forwards you — all built to be believable. Today: who gets believed, who gets left out, and how you stay you around it. Carry it into The AI Doc tonight.
Your feed, your chats, a video someone forwards you — all built to be believable. Today: who gets believed, who gets left out, and how you stay you around it. Carry it into The AI Doc tonight.
What shows up in your feed or your chat isn’t only the AI thinking. It’s what it learned from, the app built around it, what that app is trying to earn, and you — the actual person on the other end. Any one link in that chain can change who it helps and who it leaves out.
Someone decided what got written down, what got kept, what got thrown out, and what counted more than something else. Every one of those choices can leave a person out completely — not because anyone meant to, but because their life never made it into the training data in the first place.
Whose voice doesn’t get heard? Often it’s simply whoever was never in the training data to begin with.
Picture a voice assistant trained almost entirely on adults speaking clearly into good microphones, all from one country. Now picture yourself talking to it — fast, with slang, background noise, maybe an accent it’s never heard.
Name three voices it might get completely wrong — and say why.
Accents, background noise, and speech patterns outside the training data usually fail first. The system never heard enough of that voice, however clearly it was spoken.
Bias means the system treats some people or situations unfairly, without anyone deciding to be unfair. It can sneak in through what it was trained on, how a human labelled the examples, what the system got rewarded for getting “right,” how the screen is designed, or how someone ends up using it.
Builders call the reward part the objective — the exact thing the system is scored on. A system can nail its objective and still be unfair, if the objective missed something that mattered, or the system gets used somewhere it was never tested for.
A made-up “fact” that sounds polished and sure of itself can spread further than a comment that’s obviously sloppy. Confidence is part of the performance — so check what’s actually true instead of how sure it sounds.
This is the one to worry about.
Just a picture of the idea, not measured numbers.
A chatbot can agree with whatever you already believe, over and over, instead of pushing back on a weak idea. The more it agrees, the more sure of yourself you get — so your next question assumes you’re right even more, and it agrees with that too.
Researchers call this sycophancy — a system favouring agreement over telling you something true. Ask it on purpose for the strongest case against your own idea. If it can’t give you one, be suspicious.
Ask: “Why is this plan brilliant?” Then ask: “What evidence would show this plan is failing?” Same plan, same machine — a completely different answer.
Circle every claim that only showed up because of how you asked.
Any claim that shows up only in one version’s answer is the model completing the frame you handed it, not new information it went and found.
An app watches what keeps you scrolling, predicts what will keep you scrolling more, and shows you that — then uses your next click to guess even better next time. It’s not reading your mind. It’s reading your thumb.
It can hold a conversation, write a poem, sound moved by a story. None of that means anything is home. There’s no felt experience behind the words — no pain, no joy, nothing it’s actually like to be it. And the strange part: knowing that doesn’t stop it from feeling like someone’s there. Two identical shades of grey can look different side by side — your eyes insist on it even once you know better. Confident, caring-sounding output does the same thing to your gut.
Researchers don’t have one settled test for it — some think it needs a body and senses, some think it needs a particular kind of internal structure, and some think today’s systems could never qualify no matter how they’re built. Naming the disagreement is more honest than picking a side.
It remembers what you told it, matches your tone, and always replies right away — that’s exactly what makes it feel like a relationship. So ask: which parts of an actual friendship can it do, and which parts can’t it? It can listen at 2am. It can’t show up at your door.
The feeling is real. What’s making it happen isn’t a person — it’s a pattern designed to feel like one.
Honest answer: often, no — not just by looking. A weird hand or shadow might catch a bad fake today and miss a good one tomorrow. So builders stopped trying to eyeball it and started checking instead: where this first showed up, who posted the original, and whether anyone else’s evidence agrees.
“The hands look odd.”
source, original file, date, and whether someone else independently backs it up
Before you believe an image or a claim, work backward. Who posted it first? Can you find where it originally came from? Does the timing line up? And — separately, without just copying the first source — does someone else trustworthy say the same thing?
Metadata is the info saved inside a file, like the date or camera. Provenance is the paper trail of where something came from and what happened to it since. Corroboration is a separate, trustworthy source backing up the same claim. None of these words matter as much as actually doing the checking.
A strange shadow. A reverse-image search that turns up an older version. A named photographer who’ll vouch for it. A signed record of edits. A second angle from a source you trust.
A signed record of edits and a second trusted angle hold up under scrutiny. A reverse-image hit or a shadow only narrows things down. A name alone proves nothing on its own.
Public stuff is already meant for anyone to see. Personal stuff deserves a second thought. Private stuff needs someone’s actual okay first. And some things — passwords, codes, anything genuinely private about you or someone else — never go in, no matter how helpful the chatbot seems.
Say a video app is built to get more clicks. Outrage gets clicks — way more reliably than useful information does. So the app can hit its target perfectly while making everyone’s feed worse. That gap between the target and what you actually wanted has a name.
Builders call this reward hacking — when a system finds a shortcut that technically wins the metric (the number it’s scored on) without doing the actual job. The more capable the system, the better it gets at finding those shortcuts.
get the most clicks
outrage earns clicks
help people find useful information
Goal: a cleaner classroom. The metric: how many objects get removed from desks. Now go hunting for the loophole — the technically legal move that wins the count and wrecks the room anyway.
What’s the sneakiest way a system could win this rule without doing the actual job?
Almost every group finds a way to win the count without touching the actual mess. That loophole is the lesson: a metric is never quite the goal.
A typo in your private notes app is not the same as a wrong medical instruction sent to ten thousand people. To size up how risky a mistake actually is, ask: how likely is it, how many people would it hit, how bad would it be, and could anyone even fix it afterward?
When something goes wrong, it isn’t just one person’s fault — but the jobs really are different. Whoever built it should have tested it properly. Whoever put it out into the world chose where and how it gets used. Whoever used it had to apply their own judgment. And whoever it actually happened to deserves a way to push back.
Who built it
designed and tested it
Who put it out there
chose where it’s used
Who used it
had to use judgment
Who it happened to
deserves a way to push back
When an AI result causes real trouble, follow the choices backward: who picked the training examples, who decided what counts as success, who designed the app, who decided to put it here, and who can actually fix or challenge what happened.
What it learned from
What it’s scored on
How it’s shown to you
Where it’s used
Who can fix it
That’s the question the film wrestles with tonight too — not to hand you an answer, but to get you arguing about it. Tomorrow: you build something real, and you’re the one who has to make it actually work.