Alex wasn't mad
Apple Intelligence summaries look equally sure of themselves whether they're summarizing a delivery or a feeling, and feelings are where they miss. I redesigned the lock screen around one rule: summarize facts, quote feelings.
- Role
- Product Designer (solo)
- Timeline
- 6 weeks
- Type
- Concept project
- Tools
- Figma, ProtoPie, Maze
Jimmy's Tuesday, slide 4 of 31
Jimmy keeps their phone face-up on the conference table. Not to be rude. Just in case.
Jimmy reads texts the way some people read tea leaves: a period instead of an exclamation point can ruin an afternoon. Their partner, Alex, texts in all lowercase, sends four short messages instead of one long one, and ends most thoughts with 🙃.
It's 2:15 on a Tuesday. Quarterly planning, slide 4 of 31. Alex sends four texts in a row:
Jimmy spends slides 5 through 31 quietly drafting an apology for something they can't remember doing. By slide 19, Jimmy has the florist near Alex's office open in a browser tab.
At 3:00, Jimmy finally opens the thread and reads “not mad.” And “love you.”
Jimmy is a composite persona, built from experiences that came up again and again in research.
The summary kept the facts and lost the feeling.
Every summary looks equally sure of itself
iPhone users who get summarized notifications can't tell at a glance how much a summary can be trusted. Every summary looks equally confident, whether it's a delivery update or an emotional message from someone they love. When a summary misses the tone, the lock screen gives no cue to check the original, so people either act on a wrong understanding or turn the feature off.
Summarization was treated as a formatting problem: make it shorter, mark it in italics. For personal messages, it's a judgment problem.
A package update and “not mad, just wiped” go through the same process.
Same glyph, same italics, same weight, whether it's a delivery time or a feeling.
The strongest warning appears once. Jimmy agreed to it months ago.
Expanding the stack replaces the summary, so there's no side-by-side.
So the fix couldn't be another label. It had to be a decision about when to summarize at all.
Summarize facts. Quote feelings.
When a message carries emotion, the lock screen shows the sender's real words instead of a paraphrase. Facts still get summarized. Italics still mean the AI is talking; quotation marks now mean the person is.
Jimmy sends a heart, puts the phone face-down for the first time in months, and pays attention to slide 5. Slide 5 is about Q3 vendor consolidation. Some problems design can't solve.
Jimmy got an afternoon back. For Apple, the stakes are bigger.
Summaries are how people meet Apple Intelligence
People see them dozens of times a day. When summaries feel trustworthy, the rest of Apple Intelligence (Writing Tools, Siri) gets the benefit of the doubt.
People who switch summaries off rarely go back to Settings. Keeping them opted in keeps the daily habit.
Getting attribution right for news protects Apple's relationships with the organizations behind the headlines.
| Type | Metric |
|---|---|
| Primary | Rate of turning summaries off, per app and overall. Lower is better. |
| Primary | Rate of opening the original when the AI is unsure. Higher means the signal works. |
| Secondary | Reports of inaccurate summaries per 1,000 summaries. |
| Guardrail | Time spent reading notifications shouldn't rise meaningfully. |
Proposed rollout test: today's design vs. quote mode, measuring the turn-off rate.
Jimmy's metric: apology texts sent for disagreements that never happened. Target: zero.
None of this starts from zero. Apple has been working on it since launch.
Apple already solved who wrote it
Apple has been improving this feature thoughtfully since launch. The current design already marks every summary: a glyph, italic type, a label and a warning at setup.
- Oct 2024iOS 18.1Summaries collapse notification stacks into one AI-written line.
- Dec 2024The BBC caseA summary combined several BBC News alerts into one inaccurate sentence about a real criminal case.
- Jan 2025iOS 18.3News summaries paused; all summaries set in italics; a beta notice at setup.
- Sept 2025iOS 26News returns with a setup warning, a “Summarized by Apple Intelligence” label and a way to report concerns.
- TodayiOS 27The lock screen design is largely unchanged.
Disclosure is solved. People can tell a summary was written by AI.
Reliability is the next layer. People still can't tell whether a specific summary captured what the sender meant.
Apple made it clear who wrote the summary. The next question was whether to trust it.
How many Jimmys are out there?
A week of screenshots of summaries that stood out, good and bad.
People who still use summaries, people who turned them off, and heavy texters.
Lock screens shown for one second each: who wrote this, and how would it feel?
How Pixel, Galaxy AI, WhatsApp, Gmail and Slack handle AI summaries.
“I read the summary, panicked a little, and then the actual text was like ‘lol all good.’”
“I turned them off for my mom after one made her sound furious about a casserole.”
“Honestly I didn't know the slanted text meant anything. I thought it was just the font.”
Illustrative quotes: composites written to reflect themes from the interviews, not verbatim.
Remember the 12 of 19? At one second, italics read as a font, not a signal. The disclosure was there. It just wasn't registering when it mattered.
So the job wasn't to label harder. It was to know when to hold back.
Four rules for an AI that knows when to step back
The best summary is sometimes no summary.
Labeling is a start. Signal reliability too.
The sender's own words stay close at hand.
Put the signal where the decision happens.
Four rules, six ways to try them.
Six ways to keep Alex sounding like Alex
When the AI isn't confident, skip the summary and show the stack the familiar way.
Kept as the fallbackPick the most meaningful line and show it word for word, in quotation marks.
Core of the solutionMark low-confidence summaries with a softer glyph or a “tone unclear” tag.
Folded into quotingPress and hold to see the originals under the summary, side by side.
Supporting interaction“Don't summarize Alex”, offered right when someone opens the originals.
The correction loopFor news, never combine separate headlines into one sentence.
A rule for news| Idea | Why not |
|---|---|
| Confidence percentage (“73% sure Alex is upset”) | Nobody wants a probability attached to their relationship. Jimmy would refresh it hourly. |
| Ask before summarizing each time | That's a notification about a notification. |
| Mood emoji next to summaries | Asks the AI to guess feelings more, the opposite of the goal. |
| Longer summaries | It's a lock screen, not a letter. |
Four of the six made the cut, and they work better together.
One decision, three ways to show a message
The rule: italics mean the AI is talking, quotation marks mean the person is talking. Two conventions everyone already knows from reading.
Back to Tuesday, 2:15. Same four texts from Alex, run through the new flow:
Jimmy reads “not mad, just wiped from work” on slide 4. The florist never gets a visit.
Always shown as written: verification codes, bank and fraud alerts, medical results. “Your account is probably fine” helps no one.
Press and hold any summary to see the original messages right underneath, with the summary pinned on top.
People think about notifications in terms of people. The control appears right when someone opens the originals after a summary.
| Case | Example | Handling |
|---|---|---|
| Sarcasm | “great, love that for me” | Quote. The original words at least keep the joke intact. |
| Emoji only | “🙃” | Quote, as sent. Interpreting 🙃 is between Alex and Jimmy. |
| Group chats | 12 messages from 4 people | Summarize the topic, quote one person. |
| Negation | “not mad” | Always quote. This is exactly where Jimmy's afternoon went sideways. |
| Hard conversations | “can we talk about us tonight” | Quote, never paraphrase. |
| News | 3 headlines | One per line, no combining. |
A rule built on italics and quotation marks has to work for people who can't see either.
Italics can't be heard. Words can.
“Summary by Apple Intelligence: dinner moved to 7:30…”
“Alex said: not mad, just wiped from work.”
Quote mode uses quotation marks plus a name, which hold up with Bold Text, large Dynamic Type and Reduce Transparency.
Peeking also works from the expanded view, for people who can't long-press.
One rule, one lock screen. Here's where it could go next.
What I'd carry forward
Even less room, even more need for trust.
Bigger screens, same rule: facts summarized, feelings quoted.
Let Alex choose to never be paraphrased.
Sometimes the best AI design decision is knowing when to step back.
Jimmy's phone is face-down in meetings now. Alex still sends 🙃. Jimmy still isn't sure what it means, but at least now it's Alex saying it.
