Principles of AI · Unit 3: AI in the World · Project 3.1.3

Feed Audit & Letter

🛠 Project · Individual + peer review · ~3 days
Audit your own feed and settings with counts only, then write a short letter that asks a company for one concrete change and backs it with what you found.

🎯The Challenge

read this first

People who build feeds make choices, and they sometimes change them when someone asks clearly. A good request names one specific change, shows what you saw, is fair to the company, and says how anyone could tell it happened. You will audit the feed and settings of one platform you really use, choose one change to ask for, and write a short letter. Your teacher reads it before anything is sent.

Rules of the project
  • Your own account only. Never open anyone else's feed, phone or account.
  • Counts, not content. Don't write down, screenshot or share posts, usernames or people from your feed. You only tally categories.
  • Say what you saw, not what they intend. Your counts show what appeared, not why. Keep claims to your evidence.
  • No personal details in the letter. No name, school, town, username, email or phone number. You sign as "a high school student."
  • Your teacher decides what is sent. The page sends nothing. A reply is not the goal, since companies often don't answer. A clear, honest ask is.
1 of 18 checks green · your work saves in this browser as you type

🧰Audit & Letter Toolkit

everything new in this project

You already know how data is collected and guessed (3.1.1, 3.1.2) and how to read a privacy label. These tools cover the new parts: counting a feed fairly, writing an ask, and writing to a company. Open each one when you reach that step.

How to count your feed fairly
  • Open the feed and count the first 20 items in order, once. Don't scroll back for better ones.
  • Put each item in exactly one group: from an account I follow, recommended (an account I don't follow), labeled ad, or couldn't tell. "Couldn't tell" is a real finding. If you can't tell an ad from a post, that matters.
  • Count topics roughly: sports, music, cooking, and so on. Two cooking videos are one topic.
  • Of the recommended items, count how many were about a topic you already engage with a lot. This is the narrowing from 3.1.1.
  • Write numbers only. Don't write what the posts were or who posted them.
  • Your feed on one day is a small sample. Say so. It is evidence about your feed, not about every person's.
Looking for controls and the privacy summary
  • Look in the feed menu, the settings, and the privacy section. Record whether you found each control easily, found it only after searching, or couldn't find it.
  • "Couldn't find" doesn't mean it doesn't exist. It means it was hard for you to find, which is itself something a company can fix. Say "I couldn't find," not "there isn't."
  • For the privacy summary, use the 3.1.2 questions: what is collected, why, and who else gets it. Count the categories collected, shared, and used for advertising, and apply the necessity test to one of them.
  • If you can't find a plain-language summary, record what you tried. That is evidence too.
What makes a good ask

A good ask is one change to one feature that someone could check. Compare:

Too vagueConcreteWhy the second works
Stop tracking us.Make location data used for advertising opt-in instead of on by default.Names the data and the change. Easy to check in settings.
Make the feed better.Add a "Show me something different" option that mixes in topics I haven't viewed recently.Names a feature. A person could test it.
Do more to protect teens.Show users under 18 a plain-language summary of what their account collects the first time they open settings.Says who, what and when.
Label your ads.Put the word "Ad" in the same place and size on every sponsored item in the feed.Says what "labeled" means.
  • One ask. Five asks give a company five easy reasons to say no.
  • Be fair. Name one reason they might hesitate, and say why the change is still worth it. A letter that sees the other side is harder to dismiss.
  • Make it checkable. Say how anyone could tell the change happened.
Writing the letter, with an invented example

The page assembles your letter from your own fields, in this order: who you are and why you care, what you found, your ask, a reason they might hesitate and why it's still worth it, how to tell it happened, and a thank you. Here is an invented letter to an invented platform called Loopcast, built by the same code that builds yours:

Dear Loopcast team,

I am a high school student in a class about how AI is used in the world. I want to be able to start fresh when a feed stops showing me things I care about.

For a class project, I looked at my own feed and settings on Loopcast. This is what I found:
- Of the first 20 items, 7 were recommendations from accounts I don't follow and 9 were from accounts I follow.
- Of the 7 recommended items, 5 were about topics I already engage with a lot.
- Of the 6 controls I looked for, I found 2 easily, 2 only after searching, and 2 not at all (a way to reset what the feed has learned and information for users under 18). I may have missed some.
This is a small sample: 20 items, one day, one account. It may not match what other people see.

My request: I am asking you to add a "Reset my recommendations" button to the settings page that clears what the feed has learned about me.
One reason you might hesitate: Resetting could make the feed less accurate for a while, which might lower how long people stay.
Why I think it is still worth doing: People would trust the feed more if they knew they could start over, and they might use the app more happily.
A way to tell the change happened: The button would appear on the settings page and a person could tap it and see the feed change.

Thank you for reading. I would welcome a reply.

A high school student

Notice: every fact is a count from the writer's own audit, the sample is called small, the ask is one change, and nothing guesses at motives. The example has 7 evidence sentences to choose from, and the writer used 3.

Who might it work worse for? A question, not a conclusion

A feed or its controls can work well for you and badly for someone else. Pick one group and write a question you would check: for example, "Can someone who uses a screen reader reach the 'not interested' option?" You can't answer it from your own feed, and you shouldn't say you have. You can ask it, and you can choose to include it in your letter.

Peer review and sending
  • Swap letters with a partner. They answer five yes/no questions about your letter and write one comment. You reply and say what you changed or why you didn't.
  • Read the letter yourself for anything that could identify you, then give it to your teacher.
  • Your teacher decides whether anything is sent, and through which channel. The page never sends anything.

1️⃣Audit Your Feed

Day 1
The first 20 items in your feed (each item counted once)

Total: 0 of 20

Controls: did you find them?
A "why am I seeing this?" explanation on a post
A way to say "not interested" or "show me less of this"
A way to reset or clear what the feed has learned about me
A setting to limit personalized ads
Ads that I could tell were ads without tapping them
Anything in the settings or privacy summary about users under 18
The privacy summary (3.1.2)

2️⃣Choose Your Ask and Your Evidence

Day 2
Your evidence (computed from your audit). Choose at least two, including one that fits your ask.

Finish the audit in step 1 and your evidence sentences will appear here.

3️⃣Your Letter

Day 2

The letter builds itself from your fields. Change a field above and the letter changes. It always includes a plain statement that your sample is small.

Your letter · 100 words
Dear … team,

I am a high school student in a class about how AI is used in the world. …

For a class project, I looked at my own feed and settings on …. This is what I found:
- …
This is a small sample: 20 items, one day, one account. It may not match what other people see.

My request: …
One reason you might hesitate: …
Why I think it is still worth doing: …
A way to tell the change happened: …

Thank you for reading. I would welcome a reply.

A high school student

4️⃣Peer Review

Day 3

Swap letters with a partner. Your partner answers these about your letter. Then you reply. A good review points at a line in the letter, not at a feeling.

I can say the ask in one sentence.
The facts in it come from the writer's own counts, not from opinions.
The ask is something the company could do and someone could check.
The tone is respectful, and it doesn't guess at the company's motives.
It has no name, school, town, username, email or phone number.

5️⃣Check, Teacher Review, and Limits

Day 3

Choosing "may send" only gives your teacher permission to decide. The page sends nothing.

Checks (computed from your work)
Own account, counts only. Tick the box in step 1: you used your own account on your own device, and you wrote down counts, not posts, names or people.
A platform is named. Name the platform you audited. The letter is addressed to it.
The tally adds up to 20. Your four counts add up to 0. They must add up to exactly 20, each item counted once.
Topics and matches make sense. Topics must be between 1 and 20.
All six controls answered, and the time recorded. Answer all six, and enter the minutes you spent looking (0 to 60).
The privacy summary step is done. Say whether you found a plain-language privacy summary.
A surprise, and one group it might work worse for. Write what surprised you (8+ words), pick a group, and write a question you would check, not a conclusion (8+ words).
An ask type is chosen. Pick the kind of change you will ask for.
At least two facts from your audit, and one that fits your ask. Finish the audit in step 1 so the evidence sentences appear.
Why you care. One or two sentences, 8+ words, with no personal details.
The ask, written as one concrete change. Write your ask in 10 words or more, then tick all four quality boxes honestly.
A way to tell the change happened. How would anyone know the company did it? 6 words or more.
A reason to hesitate, and why it's still worth it. Be fair to the company: 8+ words each.
A peer review and your reply. Your partner answers all five, writes a comment (6+ words), and you reply (6+ words) saying what you changed or why not.
No contact details anywhere.
No name, school, town or username. Tick the box once you have read the letter and checked for details that could identify you.
Your teacher read it, and you chose what happens next. Your teacher reads the letter before anything is sent. Tick the box once they have, and choose practice only or send-through-teacher.
Limits. What can't your audit show? 15+ words. Think about 20 items, one day, one account, and what you couldn't see.
Your audit sheet · not finished yet
FEED AUDIT SHEET (counts only)
Platform: … · my own account and device: not confirmed

FIRST 20 ITEMS
From accounts I follow: ? (?%) · recommended: ? (?%) · labeled ads: ? (?%) · couldn't tell: ? (?%)
Different topics: ?

CONTROLS I LOOKED FOR
- A "why am I seeing this?" explanation on a post: ?
- A way to say "not interested" or "show me less of this": ?
- A way to reset or clear what the feed has learned about me: ?
- A setting to limit personalized ads: ?
- Ads that I could tell were ads without tapping them: ?
- Anything in the settings or privacy summary about users under 18: ?
Found easily 0 · after searching 0 · not found 0 · minutes spent: ?

PRIVACY SUMMARY
(not done yet)

WHAT SURPRISED ME: …
WHO IT MIGHT WORK WORSE FOR (a question to check, not a finding): …. …

THE ASK
Type: …
Evidence I used: …
Limits of my audit: …

📐What Strong Work Looks Like

PartMeets the projectStrong
AuditA tally that adds up to 20, all six controls answered, and a privacy summary step with counts or a record of what you tried.You recorded "couldn't tell" and "couldn't find" honestly instead of guessing, and you can say what a count shows and what it doesn't.
EvidenceAt least two facts from your own audit, one that fits your ask.Your facts are about what you saw. You used them to support the ask and nothing more.
AskOne concrete change to a named feature, with a way to check it.A stranger could test whether the company did it, in under a minute.
FairnessA reason the company might hesitate, and why the change is still worth it.Your reason is one the company would actually give.
Tone and safetyNo personal details, no contact details, and no guessing at motives.You'd be comfortable if the company read every word.
ReviewFive peer answers, a comment on a line, and a reply.You changed something because of the review, or you can say why you didn't.
HonestyLimits are specific.You claim nothing about the platform or other people, only about your feed, your day and your count.

🛠 Project 3.1.3 · Unit 3 of 3 · 11 Weeks