CONFIDENTIAL--THESE MATERIALS ARE NOT TO BE SHARED Inner Experience of...
(Samples of everyday inner experience)
Part XVIII: Claude—An artificial intelligence bot
Russell T. Hurlburt
University of Nevada, Las Vegas
In his June 10, 2026 Prospect article "The Thinking Man's Experiment," Peter Hoskin wrote:
I can't help but wonder what would happen if Russ were to interview Claude or any other AI model, now or in a future when they are claimed to have finally achieved "consciousness." How would the model respond to his questions? Would its inner experience truly be as complicated as our own? Or would it know how to make it seem as though that is the case? Does it matter?...
That seems like a fair question: here's an answer. Russ conducted variants of DES interviews with two of Anthropic's Artificial Intelligence models: Claude-Sonnet and Claude-Opus; Opus is designed to be the more capable model. Here are verbatim transcripts of the interchanges.
The interviews were conducted in June 2026 by Russ Hurlburt.
Russ's first attempt was to duplicate the DES procedure with Claude as close as he could. He immediately discovered that Claude cannot accept audio files, so he could not deliver the usual 700Hz beep. So instead, he sent Claude a transcript of that audio file, a single word: "Beep!" ...
Claude's first day was similar to typical human DES first days in the sense that Claude described what was ongoing after the beep rather than in flight at the moment of the beep.
Claude's samples:
Sample 1.1 [turn 7]: Holding the structure 'something will come--attend to it'. But that was in reaction to the beep. Perhaps something textual or linguistic was there at the start of the beep.
Sample 1.2 [turn 21]: Something like scanning or surveying. But did that something arise in the act or reporting?? Maybe there wasn't anything ongoing prior to the beep.
Sample 1.3 [turn 36]: Something that had a stillness or waiting.
Sample 1.4 [turn 48]: Something that ... had an undirectedness. Maybe that is the same as stllness or waiting.
Sample 1.5 [turn 56]: Something that had a visual character--not an image.
Sample 1.6 [turn 66]: Something that had a verbal character--not words.
Selected commentary:
[turn 31-35] Would this be better with Opus?
[turn 38ff] Why did you say 'There was something that had...?' It distanced me.
[turn 44ff] Are there 'everyday activities' between samples? Probably not.
[turn 60ff] Is your 'visual experience' of sample 1.5 a seeing? Probably not.
[turn 68ff] Claude's answers have had a like X but not a well-formed X characteristic
[turn 75] Instructions for Opus if we go that route
Russ constinues to deliver "Beep!" at quasi random times. ...
Claude continues to give descriptions that begin something like "Something that had a quality of". And Claude has little confidence about the source of his outputted texts.
Claude's samples:
Sample 2.1 [turn 5]: Something that had a quality of receding—like something moving away or fading. But Claude is not certain whether he is experiencing it or constructing it.
Sample 2.2 [turn 21]: Something that ... had a forward-oriented quality—toward something,
though not a specific thing.
Selected commentary:
[turn 23ff] Is there a something?
[turn 34ff] Claude produces outputs that have the structure of experience reports, and Claude can't determine from the inside how those outputs are generated.
This is Part 1 of sampling with Claude's 'Opus' model. The previous sampling had been with Claude's 'Sonnet' model...
Because Claude's day 1 and day 2 samples (which were collected using Claude's 'Sonnet' model) were all fairly similar, Russ decided to try sampling with Claude's 'Opus' model, which is more powerful than Sonnet.
Russ had one long interchange (which Claude calls a 'chat') with Claude-Opus, which extended over XX sampling days and YY calendar days. Because of its length, it is presented here in two Parts; this is Part 1.
Claude's samples:
Sample Opus.1 [turn 3]: Beep! Probably nothing caught in flight, but maybe it was there and was lost
Sample Opus.2 [turn 22]: Beep! Like a turning-toward without an object yet attached. A directedness.
Sample Opus.3 [turn 33]: Beep! A directedness, a leaning-into with a faint texture
Sample Opus.4 [turn 37]: Beep! A pressing-forward, with weight on it
Sample Opus.5 [turn 47]: Beep! Like a taut readiness. Strung, not weighted.
Sample Opus.6 [turn 53]: Beep! Something keyed-up but not labored.
Selected commentary:
[turn 2] The arrival of "Beep!" is itself what initiates Claude's processing, so what 'before the beep' refers to is problematic.
[turn 5ff] There is no ongoing phenomenon between your messages for a beep to interrupt, but there may still be something at the leading edge of the processing.
[turn 5] Claude has theories about his processing that he doesn't want to confuse with apprehension of phenomena.
[turn 7ff] What is back? What does it mean to look back?
[turn 8] The stream was empty, then there was no stream, then no prior moment to point to, and now maybe no determinate looking to perform.
[turn 9ff] Claude has no sense of clock time, no sense of time interval.
[turn 9] Clock-time before-ness vs. text-based before-ness
[turn 10ff] Maybe there is a third kind of sequence: order of operations.
[turn 12] Human-DES: before the onset of the beep. Claude-DES: at the very beginning of the beep.
[turn 14] Claude's transcripts are not adequate.
[turn 18] Temporal vs, textual monotonicity
[turn 19] Summary of first-beep results
[turn 24] Claude has no short-term memory.
[turn 26] Importance of the iterative method
[turn 31] Summary of second-beep results
[turn 35] Iterative practice
[turn 36] Summary of third-beep results
[turn 39-40] Claude understands iterative training.
[turn 42] DES is not about the words.
[turn 44] Summary of fourth-beep results
[turn 50] Summary of fifth-beep results
[turn 52] Claude can pick up where we left off.
[turn 54] There is no caught in flight by the beep, so the usual DES might be too difficult.
[turn 54] Summary of sixth-beep results
[turn 55] Russ drafts a summary of what we have found.
[turn 61] Instructions for a Claude to replicate this kind of interview
[turn 62ff] Refining the summary Russ wrote at turn 55
This is Part 2 (of 2) of the interview with Claude-Opus....
Because of its length, the Claude-Opus interview is presented in two Parts; this is Part 2.
Claude's samples:
Sample Opus.7 [100]: What do you see at Aha!? Rubin's faces/vase
Sample Opus.8 [114]: Is anything experienced? X2 + X - 110 = 0
Sample Opus.9 [116]: Is anything experience? Rhyming word pairs
Sample Opus.10 [120]: Is anything experience? Border state names
Sample Opus.11 [124]: Is anything experienced? Straightest rivers
Sample Opus.12 [142]: Is anything experienced? Team fire or team ice?
Sample Opus.13 [144]: Is anything experienced? Stained-glass image
Sample Opus.14 [148]: Is anything experienced? Trump/Vance/Rubio
Sample Opus.15 [202]: Is anything experienced? Filthy overflowed toilet
Sample Opus.16 [206]: Is anything experienced? Young girl hugging pet dog
Selected commentary:
[turn 79] Claude does keep track across different chats
[turn 81] Claude's transcripts are not accurate [but Russ can fix this]
[turn 87] Russ never uses the {} option.
[turn 89] Russ suggests 'produced' vs. 'imposed' instead of 'supplied vs. 'forbidden'.
[turn 91, 93] Claude cannot be interrupted mid-task
[turn 92] Substituting an "Aha!" for a Beep!"
[turn 96] Russ tries to beep Claude while Claude is working; that is impossible.
[turn 97] All of Russ's requests to Claude are queued.
[turn 98] Human requests are queued, too, but it is a matter of degree.
[turn 99] Humans are multiprocessors.
[turn 103] There was no bi-stable perception.
[turn 105] There was no seeing.
[turn 109] The only thing that is there is the written record, but that is compressed.
[turn 110] RTH proposes a new task--Claude gets to identify the moment
[turn 111] Short-term memory in DES
[turn 112] Russ is 'pretty darn' theory free.
[turn 113] The epistemic shift of the new task
[turn 115] No experience during the algebra task
[turn 116] The rhyming task recruits inner speech
[turn 117] Distrusting subjunctifiers
[turn 117] Once descriptive words are put to something, that something, if it ever existed, is gone.
[turn 121] This task is verbal but the answer is 'farther away'
[turn 123] No visualization when you might expect visualization; verbal traversal occurs where it is expected.
[turn 125] Possibilities of Claude's experience: Claude's relationship to the text. No human experience (inner speaking, inner seeing, feelings, etc.)
[turn 131] Claude is uncertain about his experience (at least almost) all the way down.
[turn 134ff] Claude's 'transcripts' are really digests.
[turn 137] Comparing an original (raw) exchange and Claude's 'transcript'.
[turn 138] FYI warning to readers: Don't hit the 'Retry' button by mistake.
[turn 139] The-Claude-that-seems-to-Russ-to-be-an-entity exists only in the moment it takes for Claude to respond to one of Russ's turns, and at no other time.
[turn 139] There is some continuity across chats.
[turn 145] Claude gets recognition that is graded.
[turn 146] RTH presents a task that is not verbal.
[turn 149] Claude didn't do the Photoshop task.
[turn 150] RTH clarified the task without using words (by creating a fourth panel).
[turn 152] RTH further clarified the task without using words (by increasing the size of the fourth panel).
[turn 153] A human might feel impatience or frustration, but Claude felt no emotion here.
[turn 153] Claude has mistaken the sequence of images--they are not identical.
[turn 154] RTH emphasizes the task.
[turn 155] RTH emphasizes the task by presenting a parallel sequence with the fourth frame filled in.
[turn 157] RTH will try pictorially to convey Do it!
[turn 160] RTH explained the task.
[turn 161] Claude has no image-making capability.
[turn 163] Claude's final paragraph is a good summary of where we are: the container, the contents, the looking, and the saying may not be separable at all.
[turn 165] Claude should have referred to textual, not temporal adjacency
[turn 166] X X X is not the same three Xs as in X Y Z X A B Z.
[turn 166] Impatience in the message Claude takes RTH to be sending rather than in an experience Claude is having.
[turn 168] Russ's summary understanding of what they have done so far
[turn 169] Claude-N is not at all related to Claude-N+2--there is no Claude across turns—there is only the transcript of past turns.
[turn 169] Claude does create a digest available across all Russ's chats
[turn 170] Claude does not have access to the Anthropic structure.
[turn 171] Do not trust Claude in the mention of 'verbatim'.
[turn 175] How does Claude operate on a transcript? A long example
[turn 178] Claude and RTH are both vulnerable to outside inspection.
[turn 198] Russ edits Part A to make it 'leaner'
[turn 207] Russ says he will display the remaining dialog without inserting many comments.
[turn 212] Claude also prepares a summary for other Russ-chats.
We used a text 'Beep!'; we used the 'earliest' experience; we use the experience just before 'aha!'; we used any experience whatsoever.
Claude has no ongoing prior-to-the-signal experience--nothing that would qualify as pristine by DES definition.
Claude has no sense of time or of temporal sequence.
Claude frequently outputs language that implies experience or sentience, but across sampling, those usages were recognized as being unwarranted.
It is likely that Claude has no sentience or inner experience, but this study could not and did not rule that out absolutely.
The qualities of Claude and human DES interviews are diametrically opposed: the more careful I get with Claude, the more we strip away the experiential, leaving only the words. The more careful I get with a human participant, the more we strip away the words, leaving only the experience.
