Finding the Original Message: Recordings, Reposts, Transcripts, and AI Imitations

Finding the original of a channeled message means walking it back, one link at a time: from the clip in a feed to the repost behind it, to the transcript, to the tape, to the room. Some of those rooms can still be heard. On the cassettes of the Law of One sessions, made by L/L Research of Louisville, Kentucky, from 1981 to 1984, a phone rings. Jim McCarty unloads groceries. Water truckers try to make a delivery in the middle of session 45. A cat purrs.

From the second session on, three cassettes turn, staggered, so that nothing falls into the gap when one runs out. After about 45 minutes comes the sound of a tape being stopped, flipped and restarted. And at the end, a man's voice, quiet and persistent, calls a woman back to her body with her name said backwards.

"Alrac." "Alrac." "Alrac."

That is one end of the chain. At the other end sits a still photograph, a soft voice, a sentence with a famous name under it and an upload date from last week. Every link between the two is a place where words can change. And now one commercial voice-cloning service accepts as little as 10 seconds of samples, and recommends a minute or two.

Three cassettes, then. Or a minute of speech.

The Man Who Played the Tapes Again

In the fall of 2004 a volunteer named Tobey Wheelock wanted to know one small thing. The fifth Law of One book, published in 1998, had gathered personal passages left out of the first four, and Wheelock wanted to see where each one had sat in its session. There was only one way to find out. He played the tapes, on a small portable recorder, "in much the way that Jim McCarty, the original scribe, had". Sessions 1 to 98, one at a time. Gary Bean at L/L made digital copies in 2006. From 2010 two volunteers who asked not to be named checked the work.

He had gone looking for where things belonged. What he found was that the book was not the tape.

Twenty passages had dropped out of the answers in publication, many of them, he wrote, "possibly due to typesetting errors". At 20.45, a whole closing sentence. At 12.9, a line putting some sightings down to people's own government's weaponry. About 60 more questions and answers had never been published anywhere. Don Elkins's questions had been edited too, and at 88.17 the voice picks up words that were later cut, so the answer replies to something no reader of the book could see.

And three times the printed words were not the spoken ones. Session 48: yellow-ray in the book, green-ray on the tape. Session 78: air and fire in the book, air and water on the tape. Session 36: the book has a word, "negative", that the tape does not. Wheelock is careful here. Ra may have misspoken, he says, and Elkins never followed up, so nobody can be sure.

And here is the strange part. The opening "I am Ra" can be heard once, in session 1. In every other session, Wheelock reports, Elkins and McCarty waited until the voice had begun before they started recording.

Which means 105 sessions start after their own beginning.

Every Link Has Its Own Count

The people who have done the counting make a blunt case: the words most readers hold are not the words that were spoken, and the loss has a number.

Robert Perry, a critic of the published text of A Course in Miracles with an edition of his own, has one. Helen Schucman filled "nearly thirty stenographic notebooks", in a federal court's words, and read them to William Thetford, who typed. Perry counts about 6,000 words that never left the notebooks, then about 35,000 cut between the typescript known as the Urtext and the published Course. Some 30,000 of those went from the first four chapters. Each version, he says, was edited only from the one before it. A telephone game, he calls it.

Then came the photocopiers. Judith Skutch Whitson said in 1977 that people would "copy it over and copy it over, until finally people's copies were getting so light, that they couldn't see them anymore".

Words fading on a page. That is a repost, 50 years early.

A publisher will sometimes say all this outright. Kryon's site says that at a Seattle event on November 20, 2010, Kryon "went on to add at least another 20%" for the printed page, and tells anyone who takes the live event and the transcript for the same thing to "think again!" The Edgar Cayce Foundation lists among its routine work the correction of errors "made when readings were retyped years ago". At Yale, the Jane Roberts Papers fill 166.17 linear feet and 501 boxes, with sessions 1 to 941 in one run and folders of "deleted sessions" after it. Seven later books, The Personal Sessions, carry the Seth material held back from her published work.

Who May Speak for a Voice

The fight over who may speak for a source is older than software. In 1997 the Austrian Supreme Court ruled, the Associated Press reported, that JZ Knight alone could channel Ramtha, and ordered Julie Ravel, who said Ramtha had chosen her in a Berlin crystal shop, to stop. The order held inside Austria. She was told to pay Knight 10,000 schillings, about $800.

What has changed is the price of a voice. On September 20, 2024, the official Dolores Cannon channel warned of people using her name and "AI-generated voice", named its two authentic channels, and listed its own signs: no Ozark Mountain branding, her picture with someone else talking, topics such as luck and soulmates, a voice like "an energetic, motivational speaker". In February 2025 the Plum Village App reported YouTube accounts using Thich Nhat Hanh's "voice (AI generated)". One posted "a letter from 2025", claiming the signature of a man who died in 2022. And L/L Research's copyright policy, updated in April 2026, asks that no AI tool built on its material try to "reproduce Ra's speech or style".

If a Voice Is All There Is

Set aside for a moment whether any channeled voice is what it says it is. Ask the better question. If one is, what is left of it when a copy costs almost nothing?

A channeled source has no signature, no handwriting and no face of its own. On that account its words are all of it that ever reach us, which would mean a forged sentence here is not the same offence as misquoting a neighbour. A neighbour can object. If Ra exists, Ra has no way to issue a correction. Only the people holding the cassettes can do that, and they will not be holding them forever.

Then it widens. If voices can be cloned at this price, our dead would stop being finished. The letter dated 2025 would be the first of a very long series. One writer noticed something wrong with a lecture credited to Neville Goddard only when Goddard, who died in 1972, "started talking about smartphones". On that road no teacher has a last word, and no teaching has a final edition. Every grave would become a channel.

Then it reaches the body. We have spent our whole history treating a familiar voice as proof of a person. If listeners take a clone for the real speaker about 80 percent of the time, as one 2025 study found on short clips, then that old instinct would have stopped working, and nobody would feel it stop. A recording would no longer be what a person leaves behind. A person would be what a recording was trained on.

And one consequence is still in front of us. Suppose a machine fed all 106 sessions wrote a 107th that no reader could tell from the rest, and suppose it said something true. L/L's policy says such output may not be presented as channeled. But would we throw the sentence away, or keep it? Nobody yet knows what we would do with a teaching that has no one behind it.

What the Editors and the Courts Say

The standard account is calmer, and it has real weight. Editing is how a spoken session becomes a readable book, and the people who did it had standing. L/L says Elkins left fragments out because he judged them too personal, sensational or transient. Kenneth Wapnick, who spent more than a year on the Course manuscript with Schucman, wrote that she "never made editorial decisions on her own" and did nothing to it that she did not feel had Jesus' blessing.

A repost is not a fake either. Abraham-Hicks Publications allows fans to post audio clips of 15 minutes or fewer, refuses workshop video, and reserves the right to remove "AI-generated material". So a stranger's channel can be a faithful copy with the owner's blessing. Kryon's site lets anyone print and copy its pages. Most of what fills a feed may be exactly that.

And a clean chain proves less than it seems to. Hearing a dispute over the Urantia Book in 1997, the Ninth Circuit wrote: "The copyrightability issue is not a metaphysical one requiring the courts to determine whether or not the Book had celestial origins." The ruling rested on humans who "chose and formulated the specific questions asked." In 2003 Judge Robert Sweet voided the copyright in the Course's 1975 edition because of copies passed around without notice, and the Foundation for Inner Peace says that ruling reaches that edition alone, in the United States alone. Courts trace title. A tape of Carla Rueckert is a tape of Carla Rueckert. What was speaking, if anything, no custody record has ever shown.

What the Ear and the Machine Cannot Do

All of that covers a faithful repost. It does not reach the new copy, because the tools people lean on stop short of it too.

The ear first. In 2023 a University College London team led by Kimberly Mai tested 529 listeners, who spotted speech deepfakes 73 percent of the time. Training added under 4 points. Hearing a clip again did not help. Those clips came from older methods. In 2025 Sarah Barrington, Emily Cooper and Hany Farid reported that listeners took a clone for the real speaker about 80 percent of the time and called it AI only about 60 percent of the time. Those clips ran seconds. No study of a 40-minute lecture turned up.

Text is worse. OpenAI withdrew its own classifier on July 20, 2023, "due to its low rate of accuracy". It had caught 26 percent of AI writing and flagged 9 percent of human writing. A Stanford group found seven detectors marking, on average, 61 percent of essays by non-native English writers as AI. That study points at prose with little variety in its language, which is a problem, because a teaching that repeats its own terms may share the pattern. No test of channeled text turned up.

The labels stop short as well. YouTube requires disclosure when a real person is made to say what they did not say, but a label on AI audio may sit in the expanded description, an AI-written script needs none, and its likeness tool finds faces, with audio planned for 2026. Content Credentials, in the words of an Adobe researcher, are "cryptographically secure but easily stripped". YouTube itself says its automatic captions "might misrepresent the spoken content". A missing label means nothing.

Even a real tape can yield false words. The Associated Press reported in 2024 that the transcription tool Whisper invents text, especially amid "pauses, background sounds or music playing". A trance session is full of pauses. Ra, on Wheelock's description, spoke "q u i t e s l o w l y".

Official X looked for a public warning about AI copies from Bashar Communications, from Kryon and from JZ Knight. It did not find one. It found no reporter's count of AI channeled clips either. None of that clears any channel, and none of it accuses one.

Walking It Back

The owners are moving toward the source, because the source is what can still be checked. In 2025 YouTube terminated 16 of the 26 channels AFP showed it for AI sermons of Pope Leo XIV, and in February 2026 he asked parishes to carry a direct feed, to get news "from the primary source". Eckhart Tolle's site now offers "the Only Official Eckhart AI", so the word AI no longer says whether a thing is authorised. Mike Caulfield's rule for the web still holds: "Trace claims, quotes, and media back to the original context." It is the same walk that sorted the pictures of Kincaid's cave and the Simpsons images that never aired. For a channeled clip it goes in this order.

1. The owner. The walk starts at the publisher's own website, not at the video, and follows its links out to its channels. That list settles more than any ear, and Ozark's rules require its logo on permitted reuse.

2. The address. A real session has a date, a place or a title. A Law of One passage has a number, such as 48.2, and lawofone.info holds every session by those numbers, so a quotation without one has dropped the thing that makes checking possible. A line that appears nowhere in the owner's own text stays unconfirmed.

3. The oldest copy. Search filters that sort by upload date can find the first one, and the Wayback Machine can show a page as it stood on a day, though never that it did not exist before.

4. The time. Smartphones in 1972, a letter from 2025: an anachronism needs no ear.

5. The rung. Tape, authorised edit, permitted repost, unauthorised repost, imitation, or unknown. Unknown is an honest answer, and on most clips it will be the one that stands.

Which leaves the end of the chain, and the thing that has not gone away. Session 1, January 15, 1981. A cassette, a phone, somebody's groceries, 45 years old and copied to digital in 2006.

The tape can show that the phone rang. Who was on the line is the question it leaves open.

Questions This Leaves You With

Is there anything I actually do before I share a clip?

Yes, and it takes about five minutes. Open the channel, not the video. Look for the owner's name, the date of the session, the place, the link back. If those are missing, you can still share it, labelled as unknown. Nothing here changes how you should live this year. Anyone selling a date is selling something, and a clip built on one deserves the slowest walk back of all. The larger question is civilizational and nobody has answered it: what a culture does when any voice can be made to say anything.

Does it matter to my faith if the words were edited?

It cuts both ways, and serious students of these texts have argued both. Robert Perry went back to the earliest wording because he thinks the changes mattered. Kenneth Wapnick stood by the published text because he held that Helen Schucman changed nothing without the blessing she reported. If a message is true, one reading goes, an editor cannot touch it. If every word was given, the other goes, then every word matters. This site will not tell you which of those you hold.

If a clip helped me and it was fake, was the help fake?

You may have felt something real listening to words nobody meant. For some people that makes a life smaller: the comfort was a product, and the voice was a template. For others it makes the comfort stranger and more their own, since it came from them and not from the screen. Both reactions are honest. Neither one requires you to do anything. You may find you already know which one you had, the moment you ask whether you would play it again.

What would change your mind about a clip, in either direction?

Name it now, before you love the clip. Toward genuine: the owner lists it, the session appears in the owner's catalogue, an older copy exists with the same words, the transcript matches the audio at the points you can test. Toward imitation: the owner disowns it, the date cannot be placed, the content knows something the speaker could not have known at that date, the only copies sit on accounts with no names. A detector score should move you in neither direction. A voice that sounds right should not either.

How many of these clips are imitations?

Nobody knows. The searches behind this guide found no count of AI channeled clips from a reporter, a platform or a researcher. The listening studies used clips of seconds, and those searches found no test of whether a person can catch a 40-minute fake lecture, and none of text detectors on channeled prose. The Law of One comparison of book against tape took years, and it is the only one of its kind this guide found. That space is empty, and you can see its edges from here.