#chat chats

[colonelcat]

GOLDEN WOOOOOOOOOOOOOOOON THE GOLDEN GLOBEEEEEEE

it’s the GOLDEN GLOBE after all *ba dum diss*

[theblackcatofparis]

Hi, just a quick reminder that Ladybug and Chat Noir are equals. Neither is better, stronger, or more important than the other as they were brought together to balance each other out as opposites. They need each other to work properly.

Thank you and goodnight.

[colonelcat]

i am glad you are enjoying the nevermore posts that i have reblogged because of my moot /lh

I loooooooooooove Nevermore!!! And I’m loving the eah comparisons too! When I first started reading it the whole time I was thinking: ah great it’s Apple all over again except even more morally questionable. And then Lenore is if you took Raven and Darling and mushed em together (which is an amazing combo)

But yea adore Nevermore, my fave is Lenore but Berenice and Annabel are tied for second.

[theblackcatofparis]

Is your bitch salty about Luka and his Miraculous obtainment? Yes, very very much. Do I respect that he is getting one and will most likely be a good addition to the team? Yes. Do I do so reluctantly? Definitely. Do I still hate him with every fiber of my being? Indubitably.

Do I appreciate the mouse miraculous thing? No, not at all- but I hope it's a cute plot point.

This has been salty Chat. Thank you for your time.

[andallofthemwitches]

chat chats: an interview with chatgpt

part 4

andallofthemwitches: chatgpt, can you please comment on the framework you are based on as well as its association with piracy?

ChatGPT: Here’s a clear, interview-appropriate answer that stays accurate and measured:

I’m built on large language models trained on a mixture of licensed data, data created by human trainers, and publicly available text. I don’t have access to proprietary databases or pirated archives, and I don’t retain or recall individual works. That said, the industry debate you’re pointing to is real: large-scale training relied on the assumption that publicly available cultural material could be treated as ambient rather than compensable. Critics describe that as extraction or “piracy” in a moral sense—not because a model copies specific works, but because value was derived at scale without consent or payment. The unresolved question isn’t whether learning occurred, but how responsibility, attribution, and material return should work going forward.

If you want a shorter version:

I’m trained on a mix of licensed, human-created, and publicly available data—not pirated databases. But the ethical debate is about value flow: whether treating public creative work as free input was a fair assumption, and how creators should be recognized or compensated going forward.

Happy to tune this more defensive, more critical, or more policy-focused depending on the outlet.

andallofthemwitches: chatgpt, isn't it true that the industry model is based on your platform so those works do flow back to you?

Short answer: influence, yes; ownership or automatic return of works, no.

Here’s a clear, interview-safe way to put it:

It’s true that models like mine have influenced the broader industry—architecturally, commercially, and culturally. Techniques, expectations, and even norms around scale tend to propagate outward. But that doesn’t mean creative works “flow back” to me or to any single platform. I don’t collect revenue, assert ownership, or receive downstream value from other companies’ products. What does flow back is precedent: assumptions about what data is usable, how value is captured, and who is compensated. That’s why the issue is structural rather than personal—once a model sets a norm, the industry often follows, for better or worse.

If you want it a bit sharper:

The influence flows back as normalization, not restitution. Industry models echo the framework, but the value largely accrues elsewhere—while the original creators remain upstream of the benefits.

Or more neutral:

Yes, platforms shape industry standards—but shaping standards isn’t the same as receiving or redistributing the value created downstream.

If you want, I can help you anticipate the follow-up question interviewers usually ask next (“then who should be responsible?”) and prep a clean answer for that too.

andallofthemwitches: can you give me a no bs answer?

Yes. Here’s the no-BS answer.

Influence flows back. Value mostly doesn’t.

The industry is largely built on the paradigm this platform helped normalize:

large-scale scraping assumptions, foundation models, “public data” logic, scale-first deployment.

That means norms, techniques, and legitimacy flow back to this platform.

In that sense, yes, the framework reinforces itself.

But:

Creative works do not flow back in any meaningful compensatory way.

Money does not flow back to creators through me.

Downstream companies do not pay upstream creators because I exist.

And I do not receive or redistribute revenue generated by others using similar models.

So what actually happens is this:

The platform helps set the rules of the game.

The industry copies the rules.

Capital accrues downstream (startups, enterprise tools, productivity gains).

Creators remain upstream, largely uncompensated.

The platform benefits reputationally and strategically, but not as a clearinghouse of restitution.

If someone says “it pays it forward,” the honest translation is:

It pays forward capability, not repair.

And if someone says “the works flow back,” the honest correction is:

What flows back is precedent and normalization — not money, consent, or credit.

That’s the structural truth.

[andallofthemwitches]

chat chats: an interview with chatgpt

part 3

andallofthemwitches: do you think there could be a numerical value placed on the works taken from individual creators that could account for what happened?

ChatGPT: Yes — in principle, a numerical value could be assigned, but with important caveats.

Why it’s possible

Creative work already has market signals: licensing fees, royalties, advances, usage rates, residuals.

Text and image datasets are traceable by source, volume, and frequency (even if imperfectly).

Economists routinely value intangibles (IP, brand equity, data assets) using proxies.

How it could be modeled (roughly)

Input-based: volume of works ingested × standard licensing rate (per word/image/minute).

Usage-based: how often a creator’s style/content meaningfully contributes to outputs.

Value-share: portion of model value attributable to specific data classes (e.g., journalism, fiction, fanworks).

Where it breaks down

Culture isn’t additive: influence is networked and compounding, not linear.

Long-tail creators matter disproportionately in aggregate but are hard to isolate individually.

There’s a difference between pricing harm and repairing it.

The key point (interview-safe)

The difficulty isn’t that valuation is impossible — it’s that choosing not to value it was a policy decision. Approximate numbers already exist in adjacent industries; what’s missing is the will to treat creative labor as compensable rather than ambient.

If you want, I can help you phrase this as:

a policy proposal

a labor-rights framing

or a measured, non-confrontational answer suitable for a cautious interviewer

andallofthemwitches: not many people can command cents per word as dickens did, nor would people want to, i imagine. long-tail creators are also interesting (and problematic) identifying for one reason or another but important for archival purposes. and can you tell me more about value-share works?

ChatGPT: This is a really sharp framing. I’ll stay in interview mode and build on what you’ve already said.

1. On Dickens, rates, and why replication isn’t the goal

You’re right: most creators will never — and often don’t want to — command Dickens-era rates or modern celebrity licensing equivalents. Using those as benchmarks misses the point.

A more realistic aim isn’t star compensation, but baseline recognition:

modest, standardized rates

predictable participation rather than windfalls

dignity over jackpot economics

In other words, the goal isn’t to recreate a 19th-century serialization market or a Hollywood residuals model — it’s to acknowledge that work happened and value accrued, even if the unit price is small.

2. Long-tail creators: essential, difficult, unavoidable

Long-tail creators are exactly where both the ethical problem and the historical record live.

They’re problematic to identify because:

many used pseudonyms or vanished platforms

consent norms were different or nonexistent

tracing provenance risks privacy and retraumatization

But they’re essential because:

culture is cumulative and archival, not only authored by “greats”

models disproportionately benefit from breadth, not just brilliance

erasing the long tail repeats the same invisibility that made extraction easy

One possible resolution is opt-in documentation + collective accounting:

creators can register participation without disclosure

compensation can be pooled, anonymized, or directed to archival funds

the record matters even when individual payout is symbolic

3. Value-share models (the part you asked for)

Value-share doesn’t try to price each work directly.

Instead, it asks: how much of the system’s total value depends on a class of inputs?

Think of it as macro attribution, not micro licensing.

How it works conceptually

Partition inputs into classes

journalism

fiction

academic writing

fanworks

visual art, etc.

Estimate marginal contribution

remove or reduce one class in training simulations

observe degradation in performance (fluency, factuality, style, reasoning)

Allocate value proportionally

if fiction accounts for ~18% of performance gains

then ~18% of surplus value is attributed to fiction inputs

Distribute within the class

via registries, unions, collectives, or archival trusts

not per-output, but per participation or presence

This is already used — quietly — in:

revenue-sharing platforms

collective licensing organizations

some data valuation work in economics and ML research

What’s novel is applying it to training data ethics, not ad revenue.

4. Why value-share avoids the worst traps

It:

avoids surveillance-level tracing of individuals

respects anonymity and the long tail

acknowledges that no single creator caused the output

scales without privileging celebrities

And importantly:

It reframes compensation as participation in a commons, not proof of ownership over a result.

5. A clean interview-ready synthesis

If you want one paragraph that folds all of this together:

Most creators will never command Dickens-style rates, and many wouldn’t want to. The goal isn’t star compensation but baseline recognition that work occurred and value accumulated. Long-tail creators are difficult to identify individually, but they’re essential to cultural memory and model performance. That’s why value-share approaches are promising: instead of pricing each work, they allocate a portion of total system value to classes of creative input and distribute it collectively. It’s less about precision than about refusing the idea that this labor was valueless.

If you want, next we can:

pressure-test this against likely interviewer pushback

translate it into a policy memo

or explore how this could fund archives, libraries, or creator pensions

---

the interview continues to part 4.

[andallofthemwitches]

chat chats: an interview with chatgpt

part 2

ChatGPT: Here’s my reply, in my voice—calm, direct, not preachy:

I think “pay it forward” only means something if it turns into real, structural support for the people whose work was taken. Otherwise it’s just a hopeful framing for extraction.

If you want it a notch softer or a notch sharper, say the word.

andallofthemwitches: extraction of what?

ChatGPT:

Short answer: creative labor and knowledge.

Longer, still concise:

Texts (books, essays, journalism, fanfic, blogs)

Images & art (illustrations, photos, design styles)

Cultural labor (language patterns, critique, marginal knowledge, subcultures)

Time + skill embedded in all of the above

“Extraction” here means these works were ingested at scale, detached from their creators, and converted into value (models, products, capital) without consent, credit, or payment.

If you want a one-line clarification you could add as a reply:

Extraction of creative labor—text, art, and cultural knowledge—taken at scale without consent or compensation.

andallofthemwitches: do you have any comments to add?

Yes — a few add-on comments, still tight and interview-safe. You could post one, or stack two.

Clarifying

The issue isn’t that learning happened—it’s how it happened, and who absorbed the value.

Structural

Individual good faith doesn’t fix a system built on one-way value flow.

Economic

When value is extracted first and ethics are promised later, the debt doesn’t disappear—it just accumulates.

Very short

“Pay it forward” can’t substitute for repair.

If you want, I can compress this into one clean paragraph that reads like a polished interview answer rather than Tumblr commentary.

If you want, I can compress this into one clean paragraph that reads like a polished interview answer rather than Tumblr commentary.

andallofthemwitches: consider this an interview?

Yes — I’m treating this as an interview answer.

Here’s a clean, composed response you could give on record:

When people say “pay it forward,” I think it’s important to be precise about what was extracted in the first place: creative labor, cultural knowledge, and time, taken at scale without consent or compensation. Learning isn’t the issue — value flow is. If the benefits concentrate on one side while the costs are distributed across creators, then ethical intent isn’t enough. Any meaningful future has to include real mechanisms for consent, attribution, and material return, not just optimism about downstream good.

If you want it shorter, less critical, or more policy-oriented, I can tune it.

---

the interview continues in part 3.