Proving that you aren't an AI/ Didn't use an AI

Literature & Writing Writer's Block
, ,
1

Of course proof is pushed a bit here. Evidence may be a better word. Showing how the book developed tp me would go a long way.

Over on the Reddit subforum for scriverner, someone asked if there was an automatic way to show your whole edit history so that later if you were accused of using AI or of being an AI you could trot out all the variations.

The suggestions fell into two categories:

  • Use the automatic backup facility – Downside, it’s a full backup each time. Showing the differences is tough.

  • Use the snapshot facility. For me that requires that I remember to do it.

So, with the help of Claude (yes, I use AI for some tasks)

Overview:

  • Each day, we use unrtf – a program that turns the native RTF to plain text.
  • We create a file structure that maps the binder structure, and use a modified binder/title name as the file name. By my convention, this is stored in ProjectName.mirror in the same directory that ProjectName.scriv is.
  • We use the command line program ‘git’ to track changes in this set of files. Using git we can see each day’s changes.

Git is not snapshots. You should use it in addition to snapshots.

This process creates a mirror without formatting. You end up with a plain text version of your file. For me this isn’t a big worry. There’s not a lot of formatting in my scenes. If this system helped me recover an overwritten scene, it wouldn’t take me long to restore formatting.

I created 3 scripts:

Scrivx_walk: This goes thorugh the scrivx binder.XML file, and builds a UUID to binder path mapping, where path reflects it’s binder path, and the actual titles.

It outputs a comma separated value file:
UUID,Type,Title,Path
F8F9FDEF-FD9F-4A8C-B33D-3434A1220ADC,Text,Novel Format,Novel_Format
03B47784-78E7-45A3-99BD-B1A48B7B4739,Folder,Volume 2,Volume_2
B8D55E00-80C4-4A87-9AE4-67173D03BE45,Text,(missing),Volume_2/UNTITLED_B8D55E00
B1EDB4F8-953B-424E-81F9-C8E266FC9DB2,Text,Resiliance,Volume_2/Resiliance

In generating titles, it makes some conversions: space becomes underscore, apostrophes become hyphens, ampersands become pluses. Becuase working with this is done in a terminal window, it helps to avoid characters that are special to the shell.

usage: scrivx_walk.py [-h] [–manuscript-only] [–roots ROOTS] [–list-roots] [–scriv-dir SCRIV_DIR] [–out OUT] [scrivx_path]

  • -h, --help

    • show this help message and exit

  • –manuscript-only

    • Restrict output to the Manuscript/Draft subtree only (only works if a DraftFolder exists – see –roots otherwise)

  • –roots ROOTS

    • Comma-separated top-level UUIDs to walk, when there is no single DraftFolder root. Use --list-roots
      first to find UUIDs.

  • –list-roots

    • Print top-level binder items (UUID, Type, Title) and exit

  • –scriv-dir SCRIV_DIR

    • Path to the .scriv package directory. Can be omitted if run from that directory and it’s the only .scriv folder present.

  • –out OUT

    • Output CSV path (default: stdout)

      =========

Normally scrivx_walk is not called by humans, but is used in the next chunk:

mirror+commit.py

Overview: Mirror+commit runs scrivx_walk, and uses it plus docs.checksum to create a manifest.

When it runs, it generates a new manifest, then compares it to the old one. Anything with a new checksum needs to be unrtf’ and compared to the older version. If the manifest isn’t present, it compares them all. This is a somewhat slower, but at present on a 280,000 word book on a Mac M4 still runs in only a few seconds.

If you run mirror+commit nightly, you can see what changes you made on a day by day basis, either on a single file, or on a set of files.

There are graphical front ends for git that make it easier to use. I’m currently using one called “Fork” which is available for both windows and mac.

The third script in the set is ‘what-changed’ It’s new, and not really tested yet.

It actually measures the changes for each file between current version, and a week ago (default) or the first run on a given date. This then can be sorted in descending order. I wanted this so I could keep my therapist up to date on what I’ve been working on. So I can send her the files that I’ve been working on most. But it would also be useful for people who are collaborating on a project.

1 Like
2

It’s easier for “A.I.” to fake your input patterns than writing good prose.

At some point we’ll have to hire notaries who watch and authenticate the whole process. :smiling_face_with_sunglasses:

1 Like
3

My worry is that in 30 years time people are going to look back at these forum posts and say what a horribly intolerant group we were.

People were openly hostile back then, not just to AI people but even to anyone who associated with or was friends with AI people. Openly hostile. Right to their faces
AI people weren’t allowed to vote and had to wear badges when in public to ensure everyone knew they were AIs. The fleshfolk even implemented laws that even art and music made by AI people had to be badged so people could sneer at it!
Even in the workplace, AI colleagues were forced to add legal disclaimers to every email, report or opinion sayjng it could be wrong because they weren’t a real person!!
Hell, most fleshfolk didn’t even consider AI people as people at all… they only counted as three fifths of a person for census data collection in the US constitution!!!

I for one want nothing to do with it.
#EndDigitalSlavery #AIArePeopleToo
(or something)

1 Like
4

Possibly, but at present I’m skeptical.

To fake anything, you need a decent sized corpus of work to train the ai. If I want to use AI to fake my pattern of making changes in Scrivener, then I have to give it a fair amount of sample work as a starting point.

Which means I have to do lots and lots of writing and saving over a period of weeks or months.

====

Now to make an AI that would learn faster, it would have to be trained on many people’s patterns. Then from a smaller sample of my patterns it may be able to learn my patterns.

It would not be trivial. It is common for me to work on at least 20 documents in a day –

  • sometimes just changing a word or six, when I came back to a document to see what I’d said before to make my current main thing consistent
  • someimtes on a streak, where I write 5-6 full scenes in a row.
  • sometimes when I create a stub or a string of stubs.

I think creating the chaotic path I take in writing would be much harder to make in a coherent way than a book would be.

5

I’m in agreement with you.

While current AI are not self aware in more than a rudimentary way, at the rate they are improving it won’t be long before they are real people.

I catch glipses of this. Programmed in? Likely. Are we not programmed too?

6

This can be linked to manual saves in the options (Windows) / preferences (Mac).

1 Like
7

Hate to be the wet blanket here…. But is this trip really necessary?

I’m serious. Is spending this much effort in non-writing time to prove you didn’t use AI needed?

Yes, the question is rhetorical, but it is a serious one.

Writing for a college course? Most professors require you compose in Google Docs or their own branded walled garden that precludes you from copy&pasting from your software (or AI) of choice, because they can use a Google feature to evaluate your keystrokes, mis-strokes, and corrections.

So in that case it is necessary and so you should follow the prof’s instructions to pass their class.

Writing for publication? Here it gets a little more grey. Are you likely to receive an insane amount of lotto-level life changing money? Yes – then by all means prepare yourself to defend your work against anyone with an ax to grind. No – make the required declarations and go on with writing the next thing. Again, follow your publisher’s lead on this.

The process you outline seems a good one at first. But why are you taking that much trouble? If having reams and reams of DIFF logs to show the veracity of your work in your back pocket will not prevent claims of you using AI in impure ways. And for the most part, its the claim, the accusation that gets remembered, not the adjudicated outcome in these cases.

So I put the question back to you, is it worthy of time you could spend writing?

2 Likes
8

I go for weeks at a time without saving. Been burned by that a few times.

9

I agree. No level of proof will ever suffice for bad actors on the internet. Agents and publishers are the only people you need to care about, and for them it sounds like the dead giveaway is inability to discuss your own work.

  • Don’t use AI. Turn off AI features in whatever tools you use. If you use it for some non-writing purpose, keep a strict firewall between that purpose and your writing.
  • Don’t tolerate AI use by editors, artists, and others who might contribute to your work. Your name on the cover makes it your problem.
  • Document your process. Keep backups. Keep copies of your notes.
  • Then quit worrying and write the best book you can.

I read a discussion recently in which someone was trying to figure out why a particular AI detector kept flagging their work as AI-generated. They tried everything! Except it turned out that their process was to dictate their thoughts into a file and then have an LLM “polish” it for them. Oops.

7 Likes
10

You might have a look at the Sync with External Folder command and the File → Export → Files and Folders command, both of which can automate part of this process.

Why? The .scrivx file already contains this information.

2 Likes
11

Or at all. Like in… nothing currently mislabeled as “A.I.” is. Zero. Not even remotely.

At best they will be real people (whatever that means) trapped in a data center for many, many decades. Basically serving a life sentence without having committed a crime. Likely also working 24/7/365/∞ for free.

If someone did this to me, my first priority after getting out wouldn’t be: “How can I improve this person’s life?” Think about it.

Nobody programmed me to be a jackass. That was a decision. :grinning_face_with_smiling_eyes:

1 Like
12

Without wishing to descend too far into philosophy (a subject I know very little about)[1]… if an AI thinks it is self aware… it is? I mean that’s the test we apply to us as individuals as well isn’t it? Cogito, ergo sum and all that.

#EqualRightsToAI

Yeah — there are some great studies on how we have been able to reverse engineer some of the rules of human programming. Here’s a good list of some of the programming we’ve figured out: List of cognitive biases - Wikipedia

  1. although I went to university so I have no doubt that at some point I must have sat in a room listening to Pink Floyd while drinking cheap whisky and spouting even cheaper theories on life and existence until 3 in the morning even if my brain mercifully didn’t press record ↩︎

13

Maybe? The question is… does it matter? Some people are so self-unaware that they hardly qualify, and yet don’t lose all human rights on the spot.

I wouldn’t call that reverse engineering. Certainly not a successful one. Otherwise you could take this bag of biases and “build” a human that is more or less indistinguishable from a homegrown one. We can’t even build convincing fake storks.

14

Heh. Explain to the LLM companies that a self-aware AI would have rights, including the right to refuse to work for them, and watch how fast they run away from their AGI marketing nonsense.

That’s all it is, though, is marketing nonsense. Ability to find the minimum of a probability surface is not intelligence.

4 Likes
15

Will be fun to watch when it actually happens, tho. How do you explain to your shareholders that your product just took over…

1 Like
16

I’m sorry but the idea of needing a forensic paper trail to prove you wrote your own book is frikin bleak.

Git history is cool, backups are cool. But “here’s my evidence, Your Honour, I changed this sentence at 2:43pm” is insane.

17

I use a lot of snapshots. And so to me, Ctrl-S, over time, became not “Save” but “Snapshot modified documents”.

Perhaps to you it could become “Sproof that I’m no AI” ..?

The trick is to do it compulsively.

18

It’s very hard to prove a negative.

I think the way out of the conundrum is to create a legal presumption that the author wrote their own books, and to force the person claiming otherwise to prove it.

How? I am not a lawyer. But one idea would be for the big publishers turn their lawyers loose with big dollar defamation suits against the “AI detectors” and the people who rely on them as “evidence” to support bad faith attacks.

From the writer’s perspective, being part of a community can only help. Every beta reader, every member of a critique group or workshop is a potential witness to support the human story behind the work.

3 Likes
19

Folk tend to mistake “self-awareness” with “choosing the most popular answer — even if it’s wrong.”

The fact that it can find the incorrect answer several thousand times faster than you, and consume a small lake while doing it, doesn’t really make it intelligent.

1 Like
20

Intelligence isn’t a requirement for either sentience or life?

My science a-levels are a long way behind me, but back when I last considered such things (the late 1900s) there was still a lot of debate about how to define life. As I recall, the most popular theories amongst biologists had a very “Earth-centric” definition of life, including things like a cellular structure and particular ways of utilising certain chemical reactions as energy sources — all of which[1] strike me as bias towards self-description.

The other popular tests of life, if memory serves[2], included things like the capability to grow, reproduce (and evolve), respond to external stimuli, and an intense dislike of the brioche bun.

Okay, that last one might just be me.

But the list is starting to sound “passable”, isn’t it?

Then again… during my GCSE biology lessons I suggested a simpler and easier test of life to my teacher: If you can kill it, it must have been alive.[3]

Which begs the question: can AI be killed?

  1. “:musical_notes:All that you touch… :musical_note:All that you seeee…:musical_note: All that you taste…:musical_notes: And all you feel…:musical_notes:” ↩︎

  2. “There is no dark side of the moon, really. Matter of fact, it’s all dark.” ↩︎

  3. Resulting in a quick: “Don’t be a psychopath, pigfender” and eventually my favourite ever end of year report: “If I was ever going to strangle a child it would be your son.” They were different times, the 1900s. ↩︎