Who gets paid when everyone can create?

Hello again, my friend,

For a long time, ‘doing the thing’ (like making something) was the hard part. You needed the skills to write, code, film or whatever else before you could worry about finding an audience or earning anything from the result.

AI is changing that. More people can create more things, faster and more cheaply than ever. But lowering the cost of creation does not necessarily mean the people will make more money just because they can do more now. In fact, it’s the opposite. If anyone can do it, it’s less valuable.

The companies providing the models, tools and distribution continue to grow, while claiming new rights over the work produced on their platforms. They’re creators to clear higher thresholds before they can earn, or compete with the companies underneath them.

That is the thread running through this week’s stories.

When AI makes mediocre production nearly free, platforms have a real quality problem to solve.

But I also believe making the first dollar easier for good creators to earn is the better long-term strategy.

The difficult part is making earning possible without making low-quality production attractive.

Let’s lock in.

Anthropic says Claude models launched in the European Union on or after August 2 will include machine-readable marks in the text they generate. Today, we know AI wrote it because of the long dashes and just the way it writes. Soon, it’ll literally be like fine print in the letters.

Images will also contain digitally signed notes to show that Claude processed them. They’ll be across supported Claude products, including the API, Claude Code and Claude itself.

The text watermark is designed to travel when someone copies and pastes the writing, and it may survive some editing. It’s hilarious really.

At first, that sounds like a clean solution to a problem people have been arguing about for years. If we can detect when AI was involved, perhaps we can finally tell who used it to write an essay, complete an assignment or quietly produce half their work. Teachers, lawyers, and almost anybody afraid of getting replaced by AI is secretly so happy about this.

Just kidding (kind of). Anthropic’s own explanation makes clear that the mark cannot tell us that.

A document may be marked because Claude wrote every word. It may also be marked because someone used Claude to translate it, proofread it, reorganize a paragraph or summarize their original research.

The opposite is also true. The absence of a detectable mark does not prove AI was not involved. The writing may have been heavily edited, passed through another system or produced using a model that does not support the marking system.

The watermark can tell us that Claude may have processed something. It cannot tell us how much of the thinking belonged to Claude, or whether the person using it contributed anything meaningful.

I have spent the past several months using AI across nearly every part of my work. But I still decide what is worth saying. I have to determine whether the argument reflects what I actually believe. I am responsible for checking the claims, and I still own the mistakes. Just like you.

A watermark may correctly say that AI touched the result. It cannot tell you where my judgment began or ended.

That distinction could eventually affect whether someone is credited, trusted, paid or penalized. If a company, school or publication begins treating every marked document as “written by AI,” then someone who used Claude to fix a few sentences could be judged the same way as someone who generated the entire piece and submitted it without reading it.

We are getting better at detecting whether a machine participated. We are not getting much better at describing what that participation involved.

Did the person have the idea? Did they provide the evidence? Did they make the important decisions? Did they verify the claims? Could they explain and defend the work without reopening the chat that produced it?

Those questions are harder to answer with software, but they are much closer to what we actually want to know.

Read Anthropic’s explanation →

Twitch announced this week that Amazon can use creator content from the platform to train generative AI models.

Creators can opt out, but they are included by default.

During a livestream responding to the backlash, Twitch Chief Product Officer Mike Minton gave an unusually honest explanation:

I appreciate the honesty. I hate what the sentence reveals. It’s a classic dark pattern, and the admission is so casual it’s very obvious they don’t care that much.

Product teams think carefully about defaults because most people leave settings as they find them. I’m leading a monetization product at beehiiv, I can guarantee they were very deliberate here.

This decision is different because Twitch already appears to know that many creators would not actively agree to it. It’s not just about what’s better or worse for the business, it’s also about deliberately hiding a setting you know your people will despise. I guess training AI is that important, and I’m not surprised.

There is a financial imbalance underneath the decision too.

Twitch creators spent years producing the streams, conversations and communities that now make the archive valuable for AI training. Amazon gets access to that material for another part of its business, while Twitch did not announce that creators would be paid or receive a share of whatever value the resulting models create.

Amazon gets a large collection of valuable training material and they’ll make billions on it. Creators get another setting they are responsible for finding and none of the money. What happens when AI just outright replaces some streamers?

This may benefit Amazon in the short term. It could also hurt Twitch’s reputation over time when creators understand what happened and feel the company took advantage of work they produced.

Every major platform is going to discover that years of user activity can be useful for training AI. Streams, posts, images, conversations and customer behaviour can all help companies build better models.

The temptation will be to secure permission through updated terms and settings that few people notice. Companies may be legally allowed to do that, but creators will still decide whether the exchange feels fair.

I do not think using creator content for AI training is automatically wrong. There may be arrangements where creators knowingly contribute their work because they are paid, receive better tools or see some other clear benefit.

But if Twitch believes this material is valuable, it should be able to explain what creators receive in return. It looks like they have, in a pretty straightforward way. It makes me wonder what other companies have done the same, and just…not said anything? I’ll give Twitch credit, they’re probably not the only ones.

When the company’s honest answer is that nobody would agree if participation were voluntary, the better question is not how to avoid asking. It is what Twitch would need to offer to make creators want to say yes.

Read the Twitch report →

YouTube is also changing its relationship with creators.

Beginning February 1, 2027, new creators applying for advertising and Premium revenue sharing through the YouTube Partner Program will need either:

That is double the current requirement of 4,000 watch hours or 10 million Shorts views. Existing members of the Partner Program will not lose access because of the new entry threshold.

YouTube says the change reflects the growth of the platform, which now sees more than 200 billion daily Shorts views and over one billion hours of television watch time each day.

The company is also expanding Premium Lite, introducing additional creator incentives and leaving its existing entry thresholds for fan funding and shopping products unchanged. It expects to pay creators more in 2027 than it did in 2026.

So this is not simply a story about YouTube taking something away.

I also understand why YouTube is doing it.

AI has made it much easier to produce videos that are technically acceptable but add almost nothing. Someone can generate a script, narration, images and editing without knowing much about the subject, or caring very much about producing something good.

If the path to monetization is easy enough, more people will flood the platform with mediocre work in the hope that a few videos earn money. Raising the threshold may discourage some of those accounts and help YouTube direct more money toward creators who have already demonstrated that people want to watch what they make.

My disagreement is not with having a quality floor. It is with using a very large commitment threshold as a substitute for judging quality.

Eight thousand watch hours can tell YouTube that a creator has attracted an audience and remained active. It cannot tell YouTube whether the work is original, thoughtful or worth encouraging.

A serious beginner and a low-quality AI content operation can both sit below the threshold. A sufficiently persistent content farm can eventually climb above it.

Commitment and quality are related, but they are not the same thing.

I would rather see platforms make it easier for good creators to earn their first dollar while becoming stricter about repetitive, misleading and mass-produced content. That is more difficult than raising one number because it requires better review systems and more judgment.

I still think it is the healthier long-term strategy.

The first dollar matters far more than its size suggests. It tells someone that their work can create value outside an employer, sponsor or client relationship. It gives them a reason to keep improving and makes it slightly more realistic to continue producing work they believe in.

For lack of a better phrase, it can help them avoid selling out too early.

If someone has to work unpaid for a long time, eventually they may need to shape the work around whoever is willing to pay first. That might be a sponsor, an employer or the type of content an algorithm happens to reward. Making a small amount directly from an audience gives creators another option, even if it does not replace their income.

Most creators will never build enormous businesses. That should not be the only measure of whether the system works. A healthy platform can also help many people earn modest amounts from work their audiences genuinely value.

That can benefit YouTube too. Creators who see an early connection between good work and payment have a reason to keep investing in the platform. Some will become large creators later, but YouTube does not need to identify all of them in advance. It needs a system that rewards promising work early enough for those people to continue.

There will always be tension here. If earning becomes too easy, quality can fall and opportunists can overwhelm the system. If it becomes too difficult, the platform may block some low-quality accounts while also forcing legitimate beginners to work for a long time before sharing in the value they help create.

YouTube provides infrastructure, discovery, hosting and an audience that would be nearly impossible for most people to assemble alone. It deserves to make money from that, and it has every reason to protect the quality of the platform.

I just do not think the best answer is always to move the first meaningful payment further away.

The better goal is to make the first dollar attainable without making low-quality production profitable.

Read YouTube’s announcement →

Lovable raised $400 million this week at a valuation of $13.3 billion.

The Swedish company allows people to create software through natural-language instructions. According to Lovable, its annual recurring revenue has nearly tripled from $200 million and is approaching $600 million. More than 60 million projects have been created on the platform, and Lovable-built applications now receive more than 900 million visits each month.

Those numbers come from the company, but even with the usual startup caveats, the direction is difficult to miss.

Software creation is becoming available to far more people.

I have experienced this firsthand. There are products I have built over the past year that would previously have required a technical co-founder, an agency or far more money than I could justify spending.

AI did not suddenly give me years of engineering experience. It did allow me to move from describing software to producing working versions of it.

That means more people can try an idea before raising money, hiring a team or giving away part of the company. In the best cases, tools like Lovable could make it easier for someone to earn their first dollar from software without needing to sell the idea to investors first.

I think that is meaningful.

But Lovable’s valuation is also a very large bet that the company can build a lasting business rather than become a temporary step between a user and an increasingly capable AI model.

The obvious risk is that companies like Anthropic and OpenAI are moving toward the same everyday customer. Their assistants are getting better at planning, coding, testing and completing larger pieces of work.

If someone mainly wants to describe an application and receive something that works, what happens when Claude or ChatGPT can offer a comparable experience directly?

Lovable does not necessarily lose that competition. But it has to give customers a reason to stay beyond making the first version easy to generate.

That reason may come from everything surrounding the initial build: deployment, collaboration, security, maintenance, support and a better understanding of what non-technical builders need once the exciting first version is finished.

Creating a demo is one problem. Running software that customers depend on is another.

An application handling money, private information or important business operations cannot survive indefinitely as a collection of prompts and good intentions. Someone has to keep it secure, fix it when dependencies change and make sure it continues doing what users expect.

If Lovable becomes the place where people can reliably run and improve the businesses they build, it may have a strong position even as the underlying models improve.

If its main advantage remains making the first build easier, the model companies could eventually offer much of that themselves.

The company has real demand, enormous growth and another $400 million to build a deeper relationship with its users. It also operates in a market where some of the companies helping make its product possible could become its most dangerous competitors.

That makes the funding both exciting and risky. Investors are not only betting that more people will build software. They are betting that Lovable can continue owning the customer relationship as the technology underneath it becomes more capable.

AI is reducing the cost of turning an idea into software. It is not reducing the difficulty of choosing a useful problem, reaching customers, maintaining the product or earning their trust.

Lovable has already shown that people want an easier way to start. Now it has to prove that it can remain useful after starting becomes easy everywhere.

Read the Lovable report →

Claude is trying to identify where AI participated in the work. Twitch is using creator output to improve another part of Amazon’s business without offering creators an obvious share of the value. YouTube is trying to protect quality by requiring more commitment before new creators can enter its primary revenue-sharing programs. Lovable is making software creation accessible to almost anyone while facing the possibility that larger AI companies eventually pursue the same customers directly.

None of these companies has an easy problem to solve. Open participation can lower quality, invite abuse and produce enormous amounts of work that nobody wants. The answer cannot simply be removing every standard.

But I still believe making the first dollar attainable is the better long-term strategy.

It gives serious creators proof that they can build something of their own before financial pressure pushes them toward an employer, sponsor or platform deal that changes the work. The standard should be whether the work creates real value, not only whether the person survived a long unpaid audition.

That balance is difficult. Platforms need to make earning possible without making low-quality production attractive.

I would rather see them do the difficult work of finding that balance than continue moving the starting line further away.

Thanks, as always, for reading.

Darwin