Everyone has a normal.
Phosphoros shows when someone leaves theirs.

Your watch knows your resting heart rate, so it notices when your pulse jumps. Phosphoros learns how a person usually talks across many interviews, and notices when an answer is not usual for them.

Try it. Which of these two answers was unusual for him?

Peter Thiel, interviewed by Mathias Döpfner, 24 September 2026. Watch the source interview

For people who watch hours of tape.

Peter Thiel · 11 interviews · 2024–2026

A challenge. A change in the answer.

When the interviewer challenges him, Thiel tends to answer at a higher pitch. His answers are shorter and start sooner. Each pattern met our test limit in 6 of 7 earlier interviews, and again in the latest interview.

Reading · transcriptWe use “pushback” for a challenge to the previous answer. This label describes the words in the exchange.

Measurements · voice and timingWe compare answers after pushback with other tagged answers in the same interview. A tag is a label from the transcript. The table shows differences between group medians. A median is the middle value.

Difference between group medians

Measurement	Earlier interviews	Latest interview	Earlier recurrence
Pitch	+3.8 Hz	+2.7 Hz	6 of 7
Answer length	−41.7 s	−26.0 s	6 of 7
Time to start	−0.36 s	−0.60 s	6 of 7

Earlier results show the median difference across seven interviews. Each pattern passes in six interviews, with different exceptions. Three other interviews have too few eligible answers. Two answers in the latest interview have uncertain starts. We identify them below.

The five pushback answers

This comparison uses 5 pushback answers and 22 other answers from the latest interview. The other answers have a median pitch of 107.9 Hz. Their median length is 75.9 s, with 0.99 s before they start. Select a time to examine the question and answer.

1. 07:05 Limits on AI

   Reading · transcriptDöpfner says that limits on AI in the West could help China. Thiel says global cooperation would be necessary. He argues that the means could be worse than the problem.

   Measurements · voice and timingPitch 103.6 Hz; answer length 79.8 s; time to start 0.39 s.

   The answer’s start is uncertain. This uncertainty can change the measurements. This comparison includes the answer, but the system does not rank it for unusualness.

2. 09:40 Greta Thunberg’s influence

   Reading · transcriptDöpfner questions Thunberg’s influence on American opinion. Thiel says her influence has decreased. He then describes problems that the West creates for itself.

   Measurements · voice and timingPitch 108.3 Hz; answer length 20.1 s; time to start 1.01 s.

3. 24:19 The AfD’s foreign policy

   Reading · transcriptDöpfner challenges the AfD’s foreign policy. Thiel supports debate about Ukraine, China, and Israel. He also agrees that the AfD is wrong on all three.

   Measurements · voice and timingPitch 110.6 Hz; answer length 29.6 s; time to start 0.16 s.

4. 36:54 An indirect answer about Germany

   Reading · transcriptDöpfner calls the previous answer indirect. Thiel expresses uncertainty, then describes a possible continued decline and its economic causes. The measured interval also includes a later question about his interest in politics.

   Measurements · voice and timingPitch 110.7 Hz; answer length 152.3 s; time to start 0.43 s.

   The answer’s start is uncertain. This uncertainty can change the measurements. This comparison includes the answer, but the system does not rank it for unusualness.

5. 47:17 His connection to Germany

   Reading · transcriptDöpfner questions his connection to Germany because he left at the age of one. Thiel describes his German home language, then Germany’s place in Europe. Döpfner then says that Thiel has returned to analysis instead of feelings.

   Measurements · voice and timingPitch 114.0 Hz; answer length 49.9 s; time to start 0.37 s.

Individual answers can differ from the group pattern. The AI answer has a lower pitch and is longer than the comparison median.

How this comparison was made

We assigned transcript labels from the question and answer text. We measured pitch and time separately. The person who labeled the latest interview had already seen its measurements. Thus, this comparison is not a blind test.

We searched for 162 possible patterns. Some can pass the test by chance. For group comparisons, the median ranks must differ by at least 0.05 within each recording. For paired measurements, the rank correlation must be at least 0.1. These limits describe patterns; they do not establish causes.

The 27 answers in this comparison come from 38 measured answers. Two have uncertain starts. The unusual-answer scores use a separate set of 29 answers with stricter requirements. See the recordings.

Other recurring patterns

“Pushback” and “changed the question” are readings from the transcript. Changes in the voice and face are measurements.

Pattern	Earlier recurrence	Latest interview
His pitch increases during the interview	9 of 10	Observed again
After pushback, his gaze moves further from the centre	5 of 6	Below the criterion
When he changes the question, he has less silence	5 of 6	Observed again
When he changes the question, he speaks more slowly	5 of 6	Below the criterion

“Below the criterion” can mean a smaller change in the same direction. It does not always mean the pattern reversed.

How we checked the results

These checks use 435 measured answers. They test speaker identification, answer order, and the number of flagged answers. They do not establish the cause of a change.

Speaker identification

We compared the results with human transcripts from five interviews. The system assigned 97.6–99.2% of words to the correct speaker.

Ranking stability

For each recording, we divided the other recordings into two reference groups. We ranked the same answers against each group. Across 65 comparisons, the median rank correlation was 0.87. A value of 1 means identical order.

The flagged answers changed more. Of all answers flagged by either group, the median share flagged by both was 38%. Similar answer order does not ensure the same set of flagged answers.

Reference size

For 4 recordings, we ranked their answers against the eight most recent earlier recordings. We compared this order with the order from all available earlier recordings. The median rank correlation was 0.99. This test does not establish eight recordings as a minimum for another person.

Flag frequency

We compared each recording with the other 12. The system flagged 31 of 435 answers (7.1%). An answer must reach the reference record’s 90th percentile and meet speaker identification and measurement requirements.

This limit does not force exactly 10% of another recording’s answers to receive a flag. The observed rate was lower, at 7.1%. This rate measures how often the system flags answers. It does not measure accuracy.

Start with one answer.

The demo shows the video, transcript, and measurements together. You can replay each exchange and compare the answer with earlier recordings.

One answer can be the start of a wider comparison.

Open the demo

A change identifies an answer for analysis. It does not establish a cause, an intention, or a lie.

Tests · Supreme Court · OT2024–25

73% of Supreme Court cases, from the voice alone.

Phosphoros heard how each justice talked to each side: pitch, loudness, pace, interruptions, reaction time. It never read what they said. The model was frozen before the test terms were measured, then run once.

See the test

The same person is the reference.

A reference is the set of measurements used for comparison. Phosphoros uses measurements from the same person. It does not use an average person.

1. Transcript The system converts speech to text. Five speaker identification runs check who speaks.
2. Measurements The system measures the person’s voice during their words. It measures their face when it is on screen.
3. Interview baseline The system compares each answer with the typical answer in that interview.
4. Record The reference record contains the differences between this person’s answers across other interviews.
5. Comparison An answer can receive an unusual score when its difference exceeds 9 of 10 reference answers. Speaker identification and measurement requirements also apply.

The score uses these 10 measurements. The demo also shows filler words, such as “uh,” and answer length.

Different recordings. Defined limits.

Microphones, cameras, and noise can change a measurement. Phosphoros compares each answer with the usual value in the same interview. Some measurements cannot support a comparison across recordings.

Equipment tests and comparison limits

Seven changes to one recording

“Holds” means the answers keep their order. “Weaker” means some answers change order. “Fails” means the test does not preserve the order.

Change	Voice	Face
Quieter microphone (12 dB)	Holds	Holds
Cheap microphone	Weaker: time to answer	Holds
Room noise	Fails: speaking rateWeaker: silence	Holds
Face 12% smaller	Holds	Weaker: brows
Face 24% larger	Holds	Fails: browsWeaker: head movement
Face 50% smaller	Holds	Fails: browsThe face was not found automatically
Low video quality (360p)	Holds	Weaker: eye opening

When we exclude a measurement

• We exclude brows and head movement from comparisons when face size differs by more than 15%. This rule excludes 7 of 12 earlier recordings from these comparisons.
• We exclude silence from comparisons when background noise differs by more than 4 dB. This rule excludes 8 of 12 earlier recordings from these comparisons.
• Five speaker identification runs check who speaks. If they disagree, we exclude the answer from the comparison.

The recordings

The record contains 13 public recordings from 11 interviews. Two interviews have two parts.

1. 693 answers In 13 recordings.
2. 435 measured A measured answer lasts at least 15 seconds and contains at least 30 words. Each comparison uses the answers that meet its requirements.
3. 269 in the record The 12 earlier recordings contain 397 measured answers. Of these, 269 have a clear start and one clear speaker. The opening example and the demo use these 269 reference answers.

Recording	Answers	Measured	In the record
Apr 2024Ann Coulter, UNSAFE, part 1	43	25	7
Apr 2024Ann Coulter, UNSAFE, part 2	40	25	8
Apr 2024Tyler Cowen, Conversations with Tyler	57	45	42
Jun 2024Aspen Ideas, The Iconoclast (stage)	38	29	20
Jul 2024Triggernometry	31	26	21
Aug 2024The Joe Rogan Experience #2190	191	90	58
Sep 2024All-In Summit (stage)	34	11	9
Nov 2024Bari Weiss, Honestly	64	50	31
Nov 2024Peter Robinson, Uncommon Knowledge, part 1	23	17	12
Dec 2024Peter Robinson, Uncommon Knowledge, part 2	28	21	19
Apr 2025Jordan B. Peterson, episode 541	36	25	19
Jun 2025Ross Douthat, Interesting Times	61	33	23
Sep 2026MD MEETS, Mathias Döpfner. Latest: the demo	47	38	30 compared
Total	693	435	269

Bring another question.

An interview, a debate, or your own conversation on camera. Tell us which recording you want to examine. You can export the report and the measurements for your own analysis.

If you own a recording on this site, contact us to request a change or removal.