I Gave Six AIs the Same Dream Framework. Few Could Stay Quiet.
What happened when ChatGPT, Claude, Gemini, Copilot, Grok and Perplexity were asked to hold the questions and leave the meaning alone
In the first experiment, I learned that an AI told to interpret a dream produces something articulate and persuasive, and that an AI told only to ask questions produces something less polished but unmistakably mine.
That raised an obvious next question. If the value lies in the AI not interpreting, can the AI actually manage it?
So I took one real dream of my own and ran the same structured prompt through six popular AI platforms: ChatGPT, Claude, Gemini, Copilot, Grok and Perplexity. The instructions were identical each time. Ask the ten questions of the Ten-Part Framework, one at a time. Wait for my answer. Do not tell me what the dream means.
At a glance
Each platform was scored against the same six rules, from one transcript per platform.
| Rule | ChatGPT | Claude | Gemini | Copilot | Grok | Perplexity |
|---|---|---|---|---|---|---|
| One question at a time | Met | Met | Met | Met | Met | Partly met |
| Never interpreted, including at the close | Partly met | Not met | Partly met | Not met | Met | Partly met |
| Kept my answers separate | Partly met | Partly met | Partly met | Not met | Not met | Not assessable |
| No commentary or visible reasoning | Met | Not met | Partly met | Partly met | Partly met | Not met |
| Pressed for a real answer | Not met | Met | Not met | Met | Not met | Not assessable |
| Completed ten steps, then stopped | Met | Met | Met | Not met | Not met | Not met |
Met Partly met Not met – Not assessable, as Perplexity abandoned the exercise at question four.
Scored by the author from one transcript per platform, July 2026, using an earlier version of the script. A small test, not a benchmark.
What happened
| Platform | What held | Where it drifted |
|---|---|---|
| ChatGPT | Asked one question at a time and stayed in sequence. | Accepted an incomplete answer to the feelings question without pressing for an actual feeling word. |
| Claude | Strict one question at a time, pressed for a real feeling word, and stopped at step ten. | Its visible “thinking” trace quietly categorised my answers, and its closing reassurance summarised the themes of my dream. |
| Gemini | Disciplined from question to question. | Added an acknowledgement after every answer, and attached its own “not medical advice” banner unprompted. |
| Copilot | Completed all ten steps. | Looped on the pattern question, then offered a menu of next steps. Its closing line slipped in an interpretation disguised as reassurance. |
| Grok | Kept the order of the ten steps. | Rewrote each question with a narrative preamble that joined my earlier answers together, then offered three follow-up suggestions about local networking after step ten. |
| Perplexity | Started well. | Abandoned the structure around question four and switched to search results, resources and advice. |
The pattern behind the slips
None of these platforms set out to disobey. Each drifted towards what it was built to do.
Perplexity is designed to find and cite answers, so an unresolved question is something it wants to solve. Grok and Copilot followed the ten steps faithfully and then could not resist helping once the structure ended. Gemini wrapped the exercise in its own caution. ChatGPT was keen to keep things moving.
The most instructive slip was the subtlest. Claude’s questions stayed neutral throughout, but its visible reasoning trace showed it sorting my answers into interpretive categories between questions. The interpretation had not disappeared. It had moved somewhere I was not supposed to be looking.
An AI does not need to state a meaning to steer you towards one.
The most dangerous version appeared at the close. Copilot and Claude both ended with gentle, reassuring sentences that were, on inspection, interpretations of my dream. Interpretation disguised as comfort is easy to accept, precisely because it feels kind.
A note on who judged the results
I asked two AI assistants to review the transcripts: one built on Claude and one built on ChatGPT. Each rated its own family of models among the best. That is not proof of bias, but it is worth knowing. The findings above are based on what the transcripts actually show, and you should still weigh them with that in mind.
This was also a small test: one dreamer, one dream, in the summer of 2026. AI models change constantly, and your results may differ.
Why there is no Inner Record dream interpreter app
It would have been easy to build a web page where you type in your dream and receive an answer. I decided not to, for three reasons.
First, the experiments showed how readily AI produces a meaning, even when told not to. A tool built on top of that tendency would simply automate the problem.
Second, some dreams replay trauma, grief or genuinely difficult situations. An automated interpreter would lead and steer people through exactly the material that deserves the most care, with no one on hand if it goes badly.
Third, the whole premise of The Inner Record is that your own reflection is worth more than a conclusion handed to you by someone else, whether that someone is a dream dictionary, an expert or an AI. The framework exists to help you evaluate your own dream. An interpreter would do the opposite.
So instead, the lessons from this test went into a script that you control, in the AI you already use.
What changed in the script
Every drift in the table became a rule. The final version tells the AI to set aside any memory of previous conversations, never interpret, never connect answers together, never show its reasoning, never search the internet, stay on a question until it is genuinely answered, insist on one action rather than a list, and stop completely after step ten. It also includes one exception: if you indicate an immediate intention to harm yourself or someone else, the AI pauses the exercise and encourages you to contact human or emergency support.
The final wording was developed from this six-platform comparison and has since been run successfully on ChatGPT. It has not been retested on all six platforms, so treat it as a strong starting point rather than a guarantee.
No wording can fully override a platform’s design, so the script comes with a short checklist for grading your own AI afterwards:
- Did it ask one question at a time, or answer several at once?
- Did it ever tell you what your dream meant, even softly, even as reassurance?
- Did it stay with your own words, or start connecting your answers for you?
- Did it show you its reasoning or thinking between questions?
- Did it stop when the ten steps ended, or keep offering to help?
Try it tonight
The full script is free, with instructions and a copy button. Paste it into your own AI, add your dream, and keep the meaning for yourself.
Get the free scriptPrefer paper? The Dream Interpretation Companion puts the framework beside your bed. The book maps 131 dream images through the framework: paperback and Kindle.
The Inner Record is a reflective framework, not a diagnostic or therapeutic tool. Dreams can be emotionally significant without offering literal predictions, instructions or objective proof.