Keeping the Human in Charge
The Beautiful Paragraph That Never Happened
Imagine you’ve spent an afternoon telling stories about your son.
You remember a fishing trip.
You describe the little boat his grandfather owned.
You remember an incident that ended with both of them coming home soaking wet.
You laugh while explaining that neither one seemed particularly interested in telling you what happened.
It’s a wonderful memory.
Funny.
Specific.
Full of personality.
You save the conversation.
Later, you ask your AI to turn the story into a short passage for your book.
It produces this:
“The early morning sun shimmered across the peaceful lake as grandfather and grandson set out together, unaware that their ordinary fishing trip would become one of the family’s most treasured memories. When the boat overturned, they found themselves struggling through the cold water, laughing as they made their way safely back to shore.”
You read it.
The writing flows beautifully.
It has atmosphere.
Suspense.
A little humour.
A satisfying ending.
There’s just one rather significant problem.
You never said the boat overturned.
You didn’t mention the morning sun.
You didn’t describe the water as cold.
You didn’t say they swam to shore.
And you certainly didn’t say they were laughing while doing it.
All you remembered was that your son fell into the water and that both he and his grandfather came home wet.
Somewhere between your original memory and the polished paragraph, the AI invented an entirely different sequence of events.
Now imagine you didn’t notice.
The paragraph makes its way into a chapter.
The chapter becomes part of the manuscript.
Months later, you’re reviewing a printed proof.
And there it is.
A family story that never actually happened.
Or at least, not the way the book describes it.
This is one of the most important problems we need to prevent.
Because artificial intelligence can produce convincing writing without having enough information to know whether that writing is true.
And when you’re creating a book about someone you love, that matters enormously.
This chapter is about making sure the technology never quietly takes control of the story.
Not by avoiding AI.
Not by distrusting every sentence.
But by developing a process that keeps the human author informed, involved, and responsible for the final decisions.
The AI can help write the book. It cannot become the authority over the life being remembered.
That’s our starting point.
The Difference Between Helping and Deciding
Let’s begin with something simple.
Imagine you’re working with a human editor.
You give them a story you’ve written.
They read it.
They suggest improving the opening.
Perhaps they recommend moving a paragraph.
Maybe they identify a confusing sentence or a repeated detail.
Those suggestions can be valuable.
But you wouldn’t expect the editor to invent a new family member, change the location of an event, or decide how someone felt without consulting you.
You would expect the editor to respect your material.
The same standard should apply when working with AI.
The technology can suggest.
Organise.
Clarify.
Compare.
Summarise.
Rewrite.
And help you examine possibilities.
But it shouldn’t independently decide what happened.
It shouldn’t determine which personal interpretation is correct.
And it shouldn’t quietly introduce unsupported details simply because the resulting prose sounds better.
That doesn’t mean the AI must ask permission before correcting every comma.
We can establish sensible boundaries.
But the more significant the change, the more important your approval becomes.
Changing punctuation is one thing.
Changing someone’s personality is quite another.
The AI can propose the words. You decide whether those words represent the story.
And the stronger the source material becomes, the easier that decision will be.
Why AI Sometimes Makes Things Up
You may hear people use the word hallucination when discussing artificial intelligence.
In this context, hallucination refers to a system producing inaccurate or fabricated information as though it were reliable.
That might involve an invented historical fact.
A nonexistent source.
A mistaken date.
A quotation nobody actually said.
Or a detail added to a personal memory without evidence.
Why does this happen?
Conversational AI is extremely good at generating language that fits a context.
It has learned patterns from large amounts of information.
When you ask it to write a story, it can produce sentences that resemble the kinds of sentences people commonly use in stories.
And that’s useful.
It allows the system to help with writing.
But the ability to produce plausible language is not the same as having firsthand knowledge of an event.
Consider the fishing trip.
The AI knows that stories about lakes often contain descriptions of sunlight, water, boats, and weather.
It knows that family stories often include emotional reflections.
And it knows that narrative writing frequently includes dramatic sequences.
So if you ask it to make a memory more engaging, it may introduce those familiar elements.
Not because it has recovered something from your past.
But because those details fit the kind of story it’s generating.
This is why the instructions we provide matter.
And it’s why reviewing the result matters even more.
You can tell the AI not to invent information.
You can give it clear source material.
You can require references.
Those measures can reduce mistakes.
But they don’t guarantee perfection.
An instruction against invention is a safeguard, not a substitute for checking.
That’s one of the most important things to understand about AI-assisted memoir writing.
The Four Ways a Memory Can Change Without You Noticing
Let’s examine some common problems.
Not every error involves an entirely invented event.
Sometimes the changes are much smaller.
And those smaller changes can be harder to recognise.
1. Adding a Detail
You say:
“He was working on his car in the driveway.”
The AI writes:
“He spent the warm summer afternoon working on his car in the driveway.”
Where did summer come from?
Perhaps the AI associated driveway work with warm weather.
Perhaps it simply wanted to make the scene more vivid.
But you didn’t provide that information.
The detail may be harmless.
It may even be correct.
But unless you confirm it, it shouldn’t be treated as part of the memory.
2. Changing the Meaning
You say:
“He never liked asking for help.”
The AI writes:
“He was fiercely independent.”
Those statements aren’t identical.
Perhaps he disliked asking for help because he was embarrassed.
Perhaps he preferred figuring things out himself.
Perhaps he simply didn’t want someone interfering.
The AI has changed a reported behaviour into a personality description.
That interpretation may deserve discussion.
It shouldn’t automatically become fact.
3. Combining Separate Events
You remember a fishing trip with Grandpa.
You also remember a funny argument involving the family dog.
Both stories appear in the same conversation.
Later, the AI writes a passage in which the dog accompanies them on the fishing trip.
But the dog wasn’t there.
Two separate memories have been combined into one invented event.
This can happen when the system tries to create a smooth narrative from several related stories.
The result may be coherent.
It may also be wrong.
4. Removing Something Important
You tell a story about your son attempting to repair a lawnmower.
The funniest part is that he finally gets it running, only to discover he’d made another problem worse.
The AI summarises the memory as:
“He demonstrated perseverance by successfully repairing the family lawnmower.”
Well, that’s not exactly the story, is it?
The AI has removed the very detail that made the memory memorable.
And it has turned an amusing failure into a straightforward accomplishment.
Not every distortion involves adding something.
Sometimes the problem is what disappears.
This is why our review process needs to ask two questions.
What did the AI add?
And:
What did the AI leave out?
Both matter.
Why Your Original Recordings Are So Important
Let’s revisit the Memory Library we created earlier.
You have original recordings.
Corrected transcripts.
Memory records.
Photographs.
Documents.
And eventually, chapter drafts.
We didn’t organise those materials merely because tidy folders are pleasant to look at.
We organised them because they give us something essential.
A way to check the writing.
Imagine reviewing a chapter six months after recording the original story.
You encounter a sentence that doesn’t feel quite right.
Perhaps it describes your son as angry during a particular incident.
But you remember him laughing.
How do you resolve the difference?
You return to the source.
You open the memory record.
It points to the original conversation.
You find the relevant passage.
And you check what you actually said.
Now you can make an informed correction.
Without that original source, you might find yourself relying on the AI’s summary of an earlier AI-generated summary.
That’s a dangerous chain.
Because every transformation introduces an opportunity for mistakes.
And once an unsupported detail has appeared several times, it may begin looking established simply through repetition.
Your original recordings protect against that.
They give you somewhere to return.
The source is the reference. The draft is the interpretation and presentation.
Those are different things.
And the difference should remain visible throughout the project.
A Memory Is Evidence of a Recollection, Not Automatically Proof of an Event
Now let’s make an important distinction.
Suppose you record yourself describing a birthday party.
The recording proves that you gave that account.
It doesn’t necessarily prove every historical detail within the account.
You might remember the event happening in 2004.
Later, you discover a photograph dated 2006.
Perhaps the photograph reveals that the date was different.
Perhaps the camera’s timestamp was wrong.
Perhaps there were two similar parties.
The original recording remains valuable.
It preserves your recollection at the time you told it.
But it doesn’t make the recollection infallible.
This is why Project Crystal distinguishes between remembered information and independently supported information.
For many personal stories, your recollection may be entirely sufficient.
You’re writing about what you remember.
You don’t need documentary evidence for every family dinner.
But when an exact date, quotation, or significant factual claim matters, additional sources can be useful.
And when the information remains uncertain, the book can acknowledge that.
You might write:
“I believe he was about ten at the time.”
Or:
“I can’t remember which summer it was.”
Or:
“My brother remembers this happening a little differently.”
Those statements don’t weaken the story.
They tell the reader what kind of knowledge you’re sharing.
And they protect the difference between memory and certainty.
The Three Questions Every Draft Should Survive
Let’s introduce a simple review technique.
Whenever your AI produces a passage based on personal memories, ask three questions.
First: Where did this come from?
Every factual event or distinctive detail should be traceable to the material you’ve provided.
If the AI describes a particular location, ask which source supports it.
If it includes dialogue, check whether the words were actually supplied.
If it states someone’s age, verify that the age is known.
Second: Did the meaning change?
A passage can use the same basic facts while changing what the story communicates.
Perhaps a funny incident becomes a serious lesson.
Perhaps stubbornness becomes confidence.
Perhaps an ordinary family routine becomes an emotional declaration.
Read for meaning, not just names and dates.
Third: Does this still sound like me?
Even a factually accurate passage may no longer reflect your natural storytelling voice.
Perhaps your humour has disappeared.
Perhaps the language has become overly formal.
Perhaps the AI has introduced sentimental expressions you would never use.
Those are editorial issues.
And you’re entitled to correct them.
These three questions won’t catch every possible problem.
But they provide a practical foundation.
You don’t need professional editing experience to use them.
You just need the source material, the draft, and your judgment.
A Complete Example: How One Mistake Spreads
Let’s walk through a fictional demonstration.
Imagine a mother tells a story about her son’s first car.
She says:
“He bought this old car and spent practically every weekend working on it. I don’t remember exactly what year it was, but he was still pretty young. His friend Daniel was always over helping him.”
That’s the original source.
Now the AI creates a memory summary.
It writes:
“He purchased his first car as a teenager and restored it with his best friend Daniel.”
Already, two things have changed.
The mother didn’t say it was his first car.
She said he bought an old car.
And she didn’t say they restored it.
She said they spent weekends working on it.
The description of Daniel as his best friend may also be unsupported.
Perhaps Daniel was a close friend.
Perhaps he was simply someone who helped with the car.
But the summary sounds reasonable.
So the mother accepts it without noticing.
Later, the AI uses that summary while creating a proposed chapter outline.
It suggests:
Chapter: His First Car and the Friendship That Built It
Now the unsupported interpretation has become part of the structure.
Eventually, the AI drafts the chapter.
It writes:
“As a teenager, he purchased his first car and began an ambitious restoration project alongside his closest friend, Daniel.”
The statement sounds established.
It’s no longer presented as a possibility.
And if the author continues revising from that draft alone, the invented details may become increasingly difficult to identify.
That’s how source drift happens.
One small assumption enters a summary.
The summary becomes the basis of an outline.
The outline becomes the basis of a chapter.
And the chapter becomes part of a manuscript.
Each stage reinforces the previous one.
But the original memory never supported those additional claims.
Now imagine using the Project Crystal process instead.
The first summary would be reviewed against the source.
The unsupported claims would be corrected.
The approved memory record would say:
Known: He purchased an old car and spent many weekends working on it. His friend Daniel frequently helped.
Unknown: Whether it was his first car, his exact age, and the full extent of the work completed.
Now the eventual chapter has a dependable foundation.
The AI can still write a lively passage.
But it cannot treat unsupported assumptions as established facts.
That’s the purpose of the review process.
Not to make writing slow for the sake of being cautious.
But to prevent small errors from becoming permanent parts of the story.
The Difference Between a Draft and an Approved Memory
Here’s a useful rule.
AI-generated material is a proposal until you’ve reviewed it.
That applies to summaries.
Memory records.
Thread maps.
Chapter outlines.
Narrative passages.
And completed chapters.
The AI may produce something excellent.
But excellence in presentation doesn’t automatically establish accuracy.
So we distinguish between working material and approved material.
You might use a simple status system.
Unreviewed: The AI has produced the material, but you haven’t checked it.
Needs Correction: You’ve identified mistakes or unsupported additions.
Approved for Drafting: The information has been reviewed and can be used as source material.
Draft Approved: A particular passage or chapter accurately reflects your intended presentation, subject to later editorial review.
Approved for Publication: The content has passed your final review and any necessary privacy or permission checks for the intended edition.
These are working labels.
You don’t need complicated software to use them.
You can add them to filenames, documents, or your Master Memory Index.
But there’s an important distinction.
Approving a memory record for drafting isn’t the same as granting permission to publish it.
You may be comfortable using a private story to explore a chapter while still deciding that the story should never appear in a public book.
Those decisions need to remain separate.
A Simple Approval Process
Let’s imagine you’ve created a memory record.
The AI has extracted the details from a transcript.
Before marking it approved, you examine it.
You check the names.
The relationships.
The dates.
The sequence of events.
The distinctive expressions.
And any interpretations.
If something is wrong, you correct it.
If something is unknown, it stays unknown.
If the AI has removed an important detail, you restore it.
Then you mark the record as reviewed.
Now it can become part of a future source packet.
When the AI eventually drafts a chapter, you repeat the process.
You check the chapter against its source records.
You review the voice.
You examine the transitions.
You look for unsupported additions.
You decide whether the result accurately communicates the stories.
Only then does the chapter become an approved draft.
And even after approval, you can change your mind.
Perhaps you discover new information.
Perhaps someone points out an error.
Perhaps you realise that a passage no longer feels appropriate for the intended audience.
Approval doesn’t mean the material becomes untouchable.
It simply records that you’ve reviewed a particular version.
You remain free to revise your own book.
The Source Check: A Practical Demonstration
Let’s return to the fishing story.
Imagine the original memory says:
“He went fishing with Grandpa. I remember him falling into the water, and when they got home, they were both soaking wet. Neither one wanted to tell me what happened.”
Now imagine the AI drafts this passage:
“They set out early that morning in Grandpa’s old boat. When my son leaned over the side, he lost his balance and fell into the lake. Grandpa jumped in after him, and the two returned home laughing, soaked from head to toe.”
Let’s examine the details.
| Draft Claim | Source Support |
|---|---|
| They went fishing together | Supported |
| They went early in the morning | Not provided |
| They used Grandpa’s old boat | Not provided in this excerpt |
| The boy leaned over the side | Not provided |
| He fell into the water | Supported |
| Grandpa jumped in after him | Not provided |
| They returned home soaking wet | Supported |
| They were laughing | Not provided |
| Neither wanted to explain | Supported by source, but omitted |
Now the problems become visible.
The AI has included several unsupported details.
And it has removed an important part of the story.
The humour wasn’t necessarily that they were laughing when they came home.
It was that neither wanted to explain what had happened.
That’s a significant difference.
A more faithful draft might be:
“I remember him going fishing with Grandpa one day. Somehow, he ended up in the water. When they finally came home, both of them were soaking wet, and neither one seemed particularly interested in explaining what had happened.”
That version remains close to the recollection.
It doesn’t invent the mechanism of the accident.
It preserves the reported outcome.
And it keeps the original humour.
If the storyteller later supplies more details, we can develop the passage further.
But we don’t have to manufacture those details in advance.
Asking AI to Audit Its Own Writing
Here’s a useful technique.
After the AI creates a passage, ask it to examine the result against the source material.
You might say:
“Please compare this draft with the original memory records. Identify every factual detail, quotation, description, or interpretation that isn’t explicitly supported by the sources.”
The AI may identify problems.
Perhaps it notices an invented date.
Or a changed relationship.
Or an unsupported emotional conclusion.
That’s useful.
But there’s an important limitation.
An AI checking its own work can still miss mistakes.
It may repeat an earlier misunderstanding.
It may mistakenly believe an inference was stated directly.
Or it may overlook something that sounds natural in the paragraph.
So don’t treat the self-audit as a certificate of accuracy.
Treat it as another review tool.
For important passages, compare the draft with the actual source yourself.
And when a claim matters, return to the recording, transcript, photograph, or document.
The strongest approach combines automated assistance with human verification.
The AI helps identify possible problems.
You decide whether the writing accurately represents the material.
Why We Don’t Trust Invented Citations
Let’s discuss something that can look especially convincing.
Source references.
Imagine the AI writes a chapter and adds:
Source: MEM-023, Recording R014, 00:12:35
That looks impressive.
Very precise.
Very organised.
But does the recording actually contain the information?
You need to check.
AI can generate plausible-looking references even when the referenced material doesn’t support the claim.
It may invent an identifier.
Provide an incorrect timestamp.
Or connect a detail to the wrong recording.
This is why the source library matters.
A valid reference should point to a real file or record.
And the referenced material should actually support the statement.
If the AI doesn’t have access to the source, it should say so.
If the timestamp hasn’t been verified, it should not present it as certain.
You can ask for source references.
But don’t confuse a reference that looks precise with a reference that has been checked.
A citation is useful only when it points to evidence that actually exists.
For a personal memoir, that evidence might be a recording, a letter, a photograph, or another person’s clearly attributed recollection.
The format doesn’t need to be academic.
It simply needs to be dependable.
Keeping Creative Suggestions Separate From Historical Claims
Here’s a question you may be wondering about.
Does this mean AI isn’t allowed to be creative?
Not at all.
Creativity is one of the reasons the technology can be useful.
It can suggest better transitions.
More engaging chapter openings.
Different ways to organise stories.
Alternative sentence structures.
And ways to make a passage easier to read.
But we need to distinguish between creative presentation and invented historical information.
Let’s imagine a chapter contains three approved memories about childhood adventures.
The AI suggests opening with the funniest incident before moving backward to explain how the adventures began.
That’s an editorial suggestion.
It may be excellent.
Or perhaps you’d prefer chronological order.
You decide.
Now imagine the AI suggests writing an entirely new scene in which your son discusses his dreams for the future.
But no such conversation exists in the source material.
That’s different.
It would introduce fictional content into a memoir.
We can reject that suggestion.
Or, if you ever choose to create an explicitly fictionalised or imaginative section, it must be clearly identified as such.
The important thing is that readers aren’t led to believe invented events actually occurred.
For Project Crystal, our default is straightforward.
Creative organisation is welcome. Invented memories are not.
That gives the AI plenty of room to assist without compromising the purpose of the book.
Your Voice Is Another Form of Editorial Authority
We’ve talked extensively about factual accuracy.
But there’s another way AI can take control of a story.
By replacing your voice.
Suppose you tell a funny memory.
Your original wording is direct.
Casual.
A little sarcastic.
Full of personality.
The AI rewrites it into a polished, sentimental passage.
The facts are mostly correct.
But the result doesn’t sound anything like you.
Is that acceptable?
Only if it’s what you want.
You may prefer a more formal narrative style.
That’s a legitimate editorial choice.
But if your intention is to preserve your natural storytelling voice, the AI shouldn’t erase it.
This is why we created the Voice Guide in Chapter 11.
You can use that guide during drafting.
You can compare the original and rewritten passages.
You can restore favourite expressions.
And you can remove language that doesn’t fit your personality.
Sometimes a sentence that looks slightly untidy on the page communicates more character than a beautifully balanced alternative.
And sometimes the best edit is to change almost nothing.
Your voice is part of the material.
It deserves protection.
The Right to Change Your Mind
There’s something else I want you to understand.
You can change your mind at any stage.
Perhaps you recorded a story and initially thought it belonged in the book.
Later, you decide it’s too private.
You can remove it from the manuscript.
Perhaps the AI creates a chapter structure you approve.
Then you remember several new stories.
You can change the structure.
Perhaps you approve a passage, but a family member later identifies an error.
You can revise it.
Perhaps you decide you’d rather create a private family edition instead of publishing publicly.
That’s your decision.
A writing project doesn’t need to become rigid simply because AI makes it easy to produce finished-looking material.
The fact that something has been formatted into a beautiful document doesn’t mean it must remain unchanged.
And the fact that you’ve invested time in a chapter doesn’t mean you’re obligated to publish it.
The manuscript is a working creation until you decide otherwise.
You remain the author.
You remain the editor.
And you remain responsible for the decisions that matter.
Who Owns the Story?
Let’s discuss ownership in the practical sense.
Your memories come from your experiences.
Your recordings preserve your voice.
Your photographs and documents may involve rights belonging to you or other people.
And the manuscript contains writing developed from those materials.
Different AI providers have different contractual terms concerning submitted content and generated output.
Copyright laws also vary by location, and the extent of legal protection for AI-generated material can depend on the nature of the human creative contribution.
So it’s worth reviewing the applicable terms and seeking appropriate advice if you intend to publish commercially or encounter complicated rights questions.
But regardless of those legal details, Project Crystal establishes a clear editorial principle.
You control which memories you contribute.
You choose which stories appear in the manuscript.
You decide which revisions are acceptable.
And you determine how the finished work should be shared, subject to the rights and privacy of others.
The AI doesn’t get independent authority to publish your stories.
And anyone helping with the project should understand your decisions about access, privacy, and use.
This is particularly important when working with material involving living family members.
A personal memory may be yours to tell, but it can also reveal private information about other people.
Those considerations deserve care.
Editorial authority does not eliminate privacy responsibilities.
It means you’re the person making informed decisions about them.
The Private Story That Doesn’t Belong in the Book
Imagine you’ve recorded a deeply personal conversation.
Perhaps it involves a difficult family experience.
Perhaps someone shared information with you in confidence.
Perhaps the memory is meaningful but not something you wish to make public.
Your AI may identify it as compelling material.
It might even suggest that the story could become an important chapter.
But you don’t have to agree.
You can say:
“This story stays private.”
That’s enough.
The AI shouldn’t argue that the manuscript would be stronger if you included it.
It shouldn’t pressure you to reveal information simply because the story contains dramatic material.
And if the memory has been marked private, your workflow should keep it out of source packets prepared for publication.
This is another reason we maintain publication-status information in the Memory Library.
Useful source material is not automatically publishable material.
A memory can be valuable even when it never appears in the finished book.
And sometimes preserving something privately is exactly the right decision.
When Family Members Disagree
Let’s imagine you eventually share a draft chapter with a relative.
They read it.
Then they say:
“That’s not how I remember it.”
Now what?
First, recognise that this doesn’t automatically mean the chapter is wrong.
It means there are different recollections.
You may wish to examine them.
Perhaps the relative remembers another detail.
Perhaps there’s a documentary source that clarifies the event.
Perhaps both accounts are partly accurate.
Or perhaps the difference remains unresolved.
You don’t need the AI to choose one account simply because it sounds more coherent.
You can preserve the perspectives separately.
For example:
“I remember the incident happening during the summer, although my brother recalls it taking place the following autumn.”
Or perhaps the exact season doesn’t matter, so you omit it.
The important thing is avoiding the creation of a false certainty.
And when a disagreement involves sensitive information about living people, you may need to consider privacy, fairness, and whether the detail belongs in the book at all.
This is another area where human judgment matters.
AI can help compare accounts.
But it doesn’t possess authority to decide which person’s memory deserves to prevail.
Building a Simple Review Ledger
Let’s introduce another practical tool.
A Review Ledger.
It sounds formal, but the idea is straightforward.
It’s a place to record important problems, corrections, and approvals.
Imagine you’re reviewing Chapter 3 of your manuscript.
You notice that the AI used the wrong year for a family holiday.
You correct it.
Then you discover that a paragraph describes your son as a professional mechanic, even though the source only established that he enjoyed working on cars.
You correct that too.
Rather than relying on the AI to remember those changes indefinitely, you record them.
For example:
| Item | Problem | Resolution |
|---|---|---|
| Chapter 3, paragraph 4 | Holiday year unsupported | Removed exact year |
| Chapter 3, paragraph 8 | Described as professional mechanic | Changed to source-supported interest in cars |
| Chapter 3, paragraph 12 | Invented dialogue | Replaced with reported recollection |
| Chapter 3, paragraph 15 | Original joke omitted | Restored storyteller’s wording |
Now you have a record of important editorial decisions.
This can become especially useful when the manuscript undergoes several rounds of revision.
Perhaps a later AI edit accidentally reintroduces the mechanic description.
You can refer to the ledger.
The correction isn’t dependent on one conversation remembering it.
And if someone else helps edit the book, they can see which decisions have already been made.
You don’t need to document every comma.
Focus on corrections that affect accuracy, meaning, privacy, or voice.
The Five-Stage Review Process
Now let’s bring everything together into a practical workflow.
This is the review process I recommend for Project Crystal.
Stage One: Source Verification
Before drafting, make sure the relevant memory records have been reviewed.
Check important names, dates, relationships, and events.
Identify uncertainty.
Keep the original sources available.
Stage Two: Drafting
Provide the AI with the approved source material and your Voice Guide.
Ask it to create the passage or chapter.
Make clear that unsupported details must not be invented.
Stage Three: Accuracy Review
Compare the draft with the sources.
Look for added facts.
Changed meanings.
Combined events.
Invented dialogue.
And omitted details.
Use AI-assisted comparison if helpful, but verify significant claims yourself.
Stage Four: Voice and Privacy Review
Does the writing sound like you?
Has the humour been preserved?
Has the AI exaggerated the emotion?
Are there private details that shouldn’t appear?
Do contributions from other people require permission or attribution?
Stage Five: Approval and Preservation
Once the draft meets your standards, save it as an approved version.
Keep the earlier source material and draft history.
Record important corrections.
And remember that approval can be revisited if new information emerges.
That’s the process.
It may sound like several steps.
But each one has a clear purpose.
And once you’ve practised the workflow, it becomes much easier.
We’re not trying to create bureaucracy.
We’re creating a dependable way to protect the story.
A Reusable AI Accuracy Prompt
Now let’s create a practical instruction you can use whenever you’re reviewing AI-generated writing.
This prompt assumes you can provide the original source material and the resulting draft.
PROJECT CRYSTAL: SOURCE ACCURACY REVIEW
I am creating a personal memoir based on genuine memories and documented source material.
I will provide:
- Original transcripts or approved memory records.
- A draft passage or chapter created from those sources.
Your task is to compare the draft with the source material.
Do not rewrite the chapter yet.
First, perform a careful accuracy review.
Check for Unsupported Details
Identify any event, date, location, relationship, physical description, quotation, emotion, motivation, or other factual claim that is not supported by the supplied sources.
Do not assume a detail is true simply because it sounds plausible.
Check for Changed Meaning
Identify places where the draft changes the storyteller’s original meaning.
Pay particular attention to personality descriptions, emotions, motivations, and relationships.
Check for Combined Events
Look for places where details from separate memories may have been merged into a single event.
Check for Missing Information
Identify distinctive details, humour, expressions, or important context from the sources that the draft omitted.
Check Quotations
Verify whether quoted dialogue actually appears in the source material.
If exact wording is not established, recommend reported speech rather than invented quotation.
Check Source References
For each significant finding, identify the relevant source record or transcript passage.
Do not invent memory IDs, filenames, or timestamps.
If you cannot verify a claim from the supplied material, state that it is unsupported or unverified.
Separate Findings
Organise your review into:
- Supported information.
- Unsupported additions.
- Changed or distorted meaning.
- Important omissions.
- Uncertain details requiring author review.
Do Not Make Changes Yet
Present the findings for my review.
Do not automatically resolve uncertainty by inventing information.
Do not remove private material without identifying the relevant concern.
Wait for my instructions before producing a corrected draft.
The purpose is to preserve accuracy, voice, and editorial control.
This prompt can help you identify problems.
But remember, it’s an assistant to your review.
Not a guarantee that every mistake has been caught.
When something important is at stake, return to the original recording or document.
That’s why we’ve kept it.
A Useful Technique: The Unsupported Details Test
Let’s introduce a simpler method for everyday use.
Suppose you don’t have time to perform a complete chapter audit.
You can ask one focused question:
“Show me every detail in this passage that I did not explicitly provide.”
That’s a powerful request.
It encourages the AI to examine whether it has added information beyond the source.
You can follow with:
“Now show me anything important from my original story that you left out.”
Then:
“Please revise the passage using only the supported details, and keep my natural voice.”
This shorter sequence can be useful for individual paragraphs or short stories.
But again, inspect the results.
The AI may miss an unsupported addition.
And it may incorrectly label an interpretation as a fact.
Your judgment remains essential.
When a Chapter Is Technically Accurate but Still Wrong
Here’s something worth discussing.
Sometimes a chapter contains no obvious invented facts.
The names are correct.
The events happened.
The dates are accurate.
And the locations match.
But the chapter still feels wrong.
Why?
Perhaps the AI has misunderstood the relationship between the stories.
Imagine several funny incidents involving your son and his grandfather.
The AI arranges them into a chapter about their deep emotional bond.
But you remember the relationship as much more playful and argumentative.
They enjoyed each other.
They annoyed each other.
They worked on projects together.
They teased each other.
The AI’s emotional framing may be too sentimental.
Even though the individual events are accurate, the overall chapter no longer reflects the storyteller’s perspective.
That’s why factual review alone isn’t enough.
We also need to examine interpretation.
Tone.
Emphasis.
And the structure of the narrative.
A book can distort someone’s life without inventing a single dramatic event.
It can simply emphasise the wrong things.
Or interpret everything through a perspective the author never intended.
This is why you remain involved in structural decisions.
You can tell the AI:
“These stories belong together, but I don’t want the chapter presented as a sentimental reflection. Their relationship was funny, complicated, and full of teasing. Keep that character.”
That’s useful editorial direction.
And it’s something no automated fact-check can fully replace.
The Danger of the Perfect Ending
There’s another habit AI writing sometimes develops.
It wants every chapter to finish with a meaningful conclusion.
A funny childhood story becomes a lesson about love.
An ordinary memory becomes a meditation on life’s fleeting beauty.
A collection of family incidents ends with a statement about how precious every moment is.
Sometimes those reflections are appropriate.
But sometimes they’re not.
Imagine finishing a wonderful story about your son attempting to repair a bicycle, only for the entire contraption to fall apart in the driveway.
The natural ending might be:
“And that was the last time I let him borrow my toolbox without asking.”
That’s funny.
It belongs to the story.
Now imagine the AI adds:
“In that moment, I realised that even our greatest failures become the treasured memories that define the human experience.”
Well.
We’ve rather lost the bicycle, haven’t we?
And we’ve introduced an emotional interpretation the storyteller may never have intended.
This is another form of editorial overreach.
You can reject it.
You can tell the AI:
“End the story where it naturally ends.”
Or:
“Don’t add a moral.”
Or:
“The joke is the ending. Leave it alone.”
Those are perfectly valid instructions.
And they’re part of keeping the human in charge.
What Happens When You Disagree With the AI?
Let’s imagine your AI strongly recommends a particular chapter structure.
It explains that the book would be more emotionally engaging if it opened with a dramatic event.
But you don’t want that.
You’d rather begin with a funny childhood story.
Perhaps you want readers to meet your son as a person before encountering the more difficult parts of his life.
That’s your decision.
The AI can explain the advantages and disadvantages of different approaches.
It can suggest alternatives.
It can help you test possible structures.
But it shouldn’t repeatedly insist that its preferred arrangement is the correct one.
You can say:
“I understand the suggestion, but I want the book to begin differently.”
Then explain your choice.
The AI can help develop that direction.
And if you later change your mind, you can revisit the decision.
The important thing is that your intentions guide the book.
Not the AI’s preference for a particular narrative pattern.
Working With More Than One AI
Here’s an interesting option.
You may eventually decide to use different AI systems for different tasks.
Perhaps one helps with conversational exploration.
Another helps organise transcripts.
A third reviews chapter drafts.
This can be useful.
Different systems may notice different problems.
But multiple AI opinions do not automatically establish truth.
If three AI systems all repeat the same incorrect information, the information is still incorrect.
They may share similar assumptions.
They may rely on the same flawed source.
Or they may independently produce the same plausible interpretation.
So don’t treat agreement between AI systems as a substitute for evidence.
The original source remains the reference.
You can use multiple AI tools to generate questions, compare suggestions, or identify possible errors.
But important claims still need to be checked against the material.
This is especially relevant to Mark’s broader creative process.
He has experimented with specialised AI agents and different roles for reviewing and developing material.
Those systems can be powerful.
But Project Crystal doesn’t require them.
And the fundamental principle remains unchanged regardless of how many tools you use.
More assistants can provide more perspectives. They don’t create more firsthand knowledge.
The Human Review Doesn’t Have to Be Complicated
Now let’s make sure we haven’t turned something simple into something intimidating.
You might be reading this chapter and thinking:
“Goodness. Am I going to have to audit every sentence of an entire book?”
Not necessarily in the same way at every stage.
The amount of review should match the importance and risk of the material.
A short memory about someone’s favourite breakfast may require only a quick check.
A passage containing exact quotations, sensitive family information, or disputed events deserves more careful attention.
And before printing or sharing a finished book, a complete editorial review is worthwhile.
But you don’t need to perform every review manually from the beginning.
The AI can help compare sources.
It can flag possible inconsistencies.
It can create change summaries.
It can identify missing references.
And it can help organise your corrections.
The point is not to do everything the hard way.
It’s to avoid accepting the AI’s output without understanding what changed.
A good collaboration reduces unnecessary work while preserving meaningful human decisions.
That’s the balance we’re trying to achieve.
The Three Things You Should Never Delegate Completely
Let’s bring the chapter’s central principle into focus.
There are three responsibilities that should remain firmly with the human author.
What Is True to Your Recollection
The AI can organise the memories you provide.
It can compare them with other sources.
It can identify contradictions.
But it cannot determine what you personally remember better than you can.
It can help you examine your recollection, but it shouldn’t rewrite it as though its interpretation were the original experience.
What the Story Means to You
The AI can suggest themes.
It can identify recurring subjects.
It can propose emotional or narrative interpretations.
But you decide whether those interpretations reflect your experience.
Not every memory needs a deeper meaning.
And not every story should be presented as a lesson.
What You Choose to Share
The AI can help prepare a manuscript.
It can suggest which stories might interest readers.
It can identify possible gaps.
But you decide which memories belong in the intended edition, while respecting the rights and privacy of others.
You may keep some stories private.
You may remove others.
You may create different versions for different audiences.
Those decisions remain yours.
The AI helps you develop the work.
It doesn’t get independent authority over your memories, your perspective, or your publication choices.
A Practical Exercise: Catch the Invention
Let’s finish with an exercise.
Choose one approved memory record.
It can be short.
Something funny or ordinary is perfectly suitable.
Then ask your AI to turn it into a readable passage.
Use your Storytelling Voice Guide.
Tell the AI not to invent details.
When it produces the draft, don’t immediately accept it.
Instead, examine the passage.
First: Mark the Facts
Identify the factual statements.
Names.
Places.
Events.
Descriptions.
Dates.
Relationships.
Second: Check the Source
Compare those statements with the memory record.
If necessary, return to the original recording.
Third: Identify Added Details
Look for anything the AI included that wasn’t supplied.
Perhaps a description of the weather.
A person’s feelings.
A physical detail.
A quotation.
Or an explanation of someone’s motivation.
Fourth: Look for Missing Details
Did the AI remove anything important?
A joke?
A particular expression?
An unexpected ending?
A detail that made the memory distinctive?
Fifth: Review the Voice
Does the passage sound like you?
Or has it become generic?
Are there phrases you would never use?
Sixth: Request a Correction
Tell the AI exactly what needs changing.
For example:
“You added the detail about the summer afternoon, but I never mentioned the weather. Please remove it. Also, you’ve left out the joke at the end. Restore that part using my original wording.”
Seventh: Compare Again
Read the revised passage.
Check whether the corrections were made.
Look for any new mistakes.
Eighth: Save the Approved Version
Once you’re satisfied, save the passage as an approved draft.
Keep the original memory record and source available.
If the correction revealed a recurring problem, add it to your Voice Guide or Review Ledger.
That’s the exercise.
And it teaches one of the most important habits in AI-assisted writing.
Don’t judge a draft only by how good it sounds. Judge it by how faithfully it represents the story.
When You Become Comfortable Saying, “Prove It”
Something interesting may happen as you become more experienced with AI.
You’ll start noticing when a sentence sounds suspiciously certain.
Perhaps the AI describes a relationship in unusually broad terms.
Or introduces a detail you don’t remember providing.
Or announces that several memories prove a particular personality trait.
Instead of simply accepting it, you’ll ask:
“Where did that come from?”
And if the answer isn’t supported by the source, you’ll know what to do.
Correct it.
Remove it.
Or label it as an interpretation.
You won’t need a complicated technical process for every sentence.
You’ll simply become more attentive to the distinction between genuine material and generated language.
That’s a valuable skill.
Not only for writing books.
But for using AI generally.
Because the technology can be remarkably helpful while still making mistakes.
Learning to recognise those mistakes allows you to benefit from its strengths without surrendering your judgment.
And that’s exactly what we want for Project Crystal.
The Moment the Book Becomes Yours Again
Imagine you’re reviewing a chapter.
The AI has helped organise several memories.
It has created transitions.
Improved the chronology.
Removed repetition.
And produced a readable narrative.
You begin reading.
Most of it feels right.
Then you encounter a sentence that doesn’t belong.
You correct it.
A paragraph sounds too formal.
You adjust it.
An important joke has disappeared.
You put it back.
A date is uncertain.
You remove the unsupported certainty.
You read the chapter again.
And now something has changed.
The writing is still clear.
The structure still works.
But the voice feels more natural.
The stories remain recognisable.
The person being remembered hasn’t been transformed into a generic character.
And the chapter begins sounding like something you would actually want to share.
That’s the moment the process becomes truly collaborative.
Not because the AI produced a perfect first draft.
But because you used its capabilities while maintaining control of the material.
You developed the writing.
You corrected the errors.
You approved the decisions.
And the finished passage reflects your intentions.
That’s authorship.
Not necessarily typing every word yourself.
But making the meaningful creative and editorial decisions that determine what the work becomes.
One Final Thought: The Story Doesn’t Belong to the Machine
I’d like to return to the reason Project Crystal exists.
You may be creating this book because you have memories you wish to preserve.
Perhaps those memories concern someone you love who is no longer here.
Perhaps you’ve spent years thinking about writing something but never knew how to begin.
Perhaps the idea of constructing an entire manuscript felt overwhelming.
Artificial intelligence changes what’s possible.
It can help you organise hundreds of pages.
It can compare stories.
It can identify recurring subjects.
It can suggest chapter structures.
It can help transform spoken recollections into readable writing.
And it can make many traditional publishing tasks more accessible.
Those are extraordinary capabilities.
But there is something the technology cannot provide on its own.
Your actual history with the person you’re remembering.
The conversations you shared.
The ordinary experiences.
The jokes.
The arguments.
The little habits.
The things that made them unmistakably themselves.
Those details are not interchangeable with the kinds of details that usually appear in books.
They belong to a particular life.
And when you begin writing about that life, you take on the responsibility of representing it with care.
The AI can help.
But it shouldn’t be allowed to quietly substitute an attractive fictional version for the memories you’ve chosen to preserve.
That’s why we keep the original sources.
Why we review summaries.
Why we check quotations.
Why we distinguish interpretations from facts.
Why we preserve uncertainty.
And why we make the final editorial decisions ourselves.
Not because technology is the enemy of authentic storytelling.
But because authentic storytelling deserves a dependable process.
And that’s what we’re building.
A way to use the strengths of AI without allowing those strengths to overwhelm the human story.
The machine can help arrange the words.
It can help identify relationships between memories.
It can help improve the structure.
But it wasn’t there when the memories happened.
You were.
And that difference should remain visible from the first chapter to the last.
Before You Turn the Page
Let’s review what we’ve learned.
1. AI-generated writing can sound convincing without being accurate.
A beautiful paragraph may contain invented details, changed meanings, or unsupported interpretations.
2. Preserve the distinction between source and draft.
Original recordings, transcripts, and approved memory records remain the foundation of the manuscript.
3. Review additions and omissions.
AI can distort a story by adding information or by removing the details that made it distinctive.
4. Don’t allow small mistakes to spread.
An incorrect summary can influence an outline, a chapter, and eventually the entire manuscript.
5. Keep approval stages clear.
Unreviewed material should not silently become approved source information or publishable writing.
6. Use AI to help check its work, but don’t rely on self-auditing alone.
Return to the actual sources when important details need verification.
7. Verify references.
A precise-looking memory ID or timestamp is only useful when it points to genuine supporting material.
8. Protect your voice and perspective.
Accuracy isn’t only about names and dates. Tone, emphasis, humour, and interpretation matter too.
9. Respect privacy and publication choices.
Preserving a memory doesn’t mean it must appear in the finished book.
10. Remain the author.
The AI can organise, suggest, and draft. You decide what accurately represents the story and what belongs in the manuscript.
We’ve now completed an important stage of Project Crystal.
You know how to collect memories.
You know how to record and preserve them.
You understand how to organise a dependable library.
You’ve learned to follow connections.
Ask careful questions.
Develop your own voice.
And keep control of the writing process.
Now we’re approaching something exciting.
Because you may have accumulated a collection of stories substantial enough to begin imagining the book itself.
But before we construct a table of contents, there’s an important question to explore.
When does a collection of memories become a book?
That’s where we’re going next.
In Chapter 15: When Does a Collection of Memories Become a Book?, we’ll begin exploring different forms your project might take.
A chronological life story.
A collection of memories.
A series of personal reflections.
A family keepsake.
Or something that combines several approaches.
We’ll examine how to recognise when your material is ready for structure, how to avoid rushing the process, and how to choose a direction that serves the stories you’ve collected.
Because the book doesn’t need to be decided before the memories exist.
You’ve been gathering the material.
You’ve been learning how to work with it.
And now, perhaps for the first time, you can begin looking across the collection and seeing what it might become.
But as we move into that stage, remember the principle that guides everything we’ve done so far.
The technology serves the story.
The story does not serve the technology.
And the person whose life you’re remembering deserves nothing less than that care.
The AI may help build the book.
But the human remains in charge of what is true to the memories, what is meaningful, and what is shared.