Following prior research, I had a language model stand in for me and measured whether it made the same choices I would. Whatever the material, 74 days of diary (about 20,000 Japanese characters), a one-line description of myself, or an interview transcript, agreement on Apple's PCC stayed within 0.62–0.67 (62–67%). The full diary did no better than the one-line description, and adding material sometimes made it worse. How I speak, on the other hand, clearly changed with who I was talking to. Here is what I measured and how it changed the app's design.
What I wanted to do
I build Liner Notes, an iPhone diary app. Down the road I want it to help leave your family a record of who you were. Once diary entries and conversation records pile up, how far can an AI that answers in your place (a so-called digital twin) go? In October 2026 I measured it on myself. I followed prior research: Park et al. (2024) interviewed about 1,000 people for two hours each and had a language model predict their survey answers from the transcripts. In the same way, I gave a language model my material and had it predict my answers.
How I measured
I prepared 74 questions about my thinking and preferences, in five groups: personality, how I make decisions, how I take risks with time and money, everyday choices, and social attitudes. Answers were multiple-choice or on a five-point scale. I answered first, then gave a language model some material, had it answer "as this person would," and measured how close it came to my answers.
Agreement is the mean of 1 minus (difference ÷ width of the scale). 1 is a perfect match; answering at random gives 0.55.
I used three models: Qwen2.5 7B and 14B (4-bit) on a Mac, and Apple's Private Cloud Compute (PCC).
Result: changing the material barely changed agreement
| Material | PCC | 14B |
|---|---|---|
| Random answers | 0.55 | 0.55 |
| No material | 0.53 | (did not answer) |
| One-line description (age, family and other basics) | 0.67 | 0.72 |
| 74 days of diary (about 20,000 characters) and records | 0.66 | 0.68 |
| 12-question interview transcript (about 5,700 characters) | 0.66 | 0.61 |
| One-line description plus my answers to 31 questions about how I think | 0.64 | 0.61 |
Agreement by material
1 = perfect match
Dashed line: Random answers (0.55)
Show as a table
| PCC | 14B | |
|---|---|---|
| One-line description | 0.67 | 0.72 |
| 74 days of diary | 0.66 | 0.68 |
| 12-question interview | 0.66 | 0.61 |
| One line + 31 answers | 0.64 | 0.61 |
Reading the full diary did no better than one line. Adding my own direct answers to 31 questions actually made it worse. This is one person on 74 questions, so I treat differences of about ±0.05 as noise.
Looking into why it dropped, my own answers contradicted each other depending on how I was asked. In the 31 questions I wrote freely that "I decide once it's good enough" and "I don't mind leaving things open," but on the five-point scale I rated "I find indecision painful" as "somewhat true." The model followed my words and missed.
Some questions showed the model's own bias regardless of material. On the 9 questions about risk with time and money, PCC chose "safe and now" whatever I gave it, opposite to me on all 9. By question group, the diary helped only with how I make decisions; time and risk did not move with either material.
Agreement by question group (PCC)
1 = perfect match
Dashed line: Random answers (0.55)
Show as a table
| One line | Diary | |
|---|---|---|
| Personality | 0.65 | 0.69 |
| Decision-making | 0.8 | 0.88 |
| Time and risk | 0.33 | 0.33 |
| Everyday choices | 0.87 | 0.73 |
| Social attitudes | 0.57 | 0.55 |
Not what I say, but how I say it changes with the listener
If material about what I say does not raise accuracy, what makes a person themselves may lie elsewhere. So I took measurements of how I speak from my conversation recordings, by listener.
| Listener | Pause before speaking | Sentence length |
|---|---|---|
| Coworkers | 1.6 s | 32 characters |
| My wife | 2.2 s | 20 characters |
| My daughter | 2.8 s | 20 characters |
| My boss | 5.3 s | (not measured) |
Pause before speaking, by listener
Seconds
Show as a table
| Item | Value |
|---|---|
| Coworkers | 1.6 s |
| My wife | 2.2 s |
| My daughter | 2.8 s |
| My boss | 5.3 s |
My speaking rate was 4.7–5.0 characters per second with everyone. What changes is the pause and the sentence length. In 26 of 42 recordings the listener had not been recorded, so these numbers come from the other 16.
Prior research points the same way
That how someone speaks identifies them has been shown in earlier studies.
| Study | Finding |
|---|---|
| Helander et al. (2007) | People recognize acquaintances even when the voice is replaced by sine waves so no words can be heard (70.4% in a two-way choice) |
| Kanber et al. (2021) | People recognize their partner's voice from non-word sounds like "um" and "uh" 91.3% of the time |
| IMPersona (2025) | Even family members are fooled by an AI impersonating someone in text. Text the real person wrote is judged to be them only 70.5% of the time |
| Interviews with 16 bereaved people (2026) | Whether a chat imitating the deceased feels like them depends on tone and word choice more than on factual accuracy |
| Park et al. (2024) | For predicting survey answers, a summary with the speaking style removed does about as well as the interview transcript (0.81 vs 0.83) |
The last study shows that measuring by content does not reveal differences in speaking style, the same direction as my result. What makes someone themselves shows up in how they say things more than in what they say.
How I changed the app
Based on this, I changed how conversations are stored. The app used to keep only the polished transcript. Now it keeps the following, only for the parts where I spoke:
- The words as spoken (the transcript before polishing)
- Voice numbers (pitch, rate, pauses, laughter)
- Who I was talking to
Other people's voices are not kept, and the audio itself is deleted after 24 hours. To evaluate an AI copy, I no longer use agreement on choices; I look at whether people who know me feel the writing and tone are "really me."
Summary
- An AI standing in for me did no better with 20,000 characters of diary than with one line. Before adding material, measure where it plateaus
- Your own answers change with how you are asked. Adding them as material can lower accuracy
- What makes a person themselves shows in how they speak. Pauses and sentence length change with the listener
- If you want to keep what makes someone themselves, keep the words as spoken and the voice numbers, not just polished text
More engineering notes
- On iOS, run background audio processing on the Neural Engine
In an iOS BGContinuedProcessingTask, CPU work ran 12-21x slower while Core ML on the Neural Engine stayed within 2x. Three causes and the fix, measured.
- Choosing a speech recognizer by benchmark alone is risky. Measure with the audio your app actually handles
On read speech Apple's recognizer looked best; on conversation its error rate was 1.8x Parakeet's. How to compare Japanese speech recognizers, with data.
- Same transcription accuracy, three times the errors in the summary
Two recognizers with the same CER produced 3x the summary errors. Why a stricter summary prompt made it worse, and how to compare by summary quality.
- Apple's Private Cloud Compute refused 10.6% of my app's requests as unsafe
Apple's Private Cloud Compute (Foundation Models) refused 10.6% of 614 production requests and rate-limited 9.6%. What broke and how I redesigned.