X is retiring Revenue Sharing on September 7 and opening applications for Original Content Rewards on September 8, gated on Premium, 500 verified followers and 500,000 verified-user impressions. The originality standard sits entirely in editorial guidance with no stated detection method — a rule written against reposters, arriving in a year whose problem is generators.
X announced on 8 August that it is retiring its creator Revenue Sharing program and replacing it with something called Original Content Rewards. The easy version of this story is that it is the fourth revision to X's creator payouts in a year, and that Musk-era policy churn continues. That version is accurate and not very useful.
The more interesting observation is what X is attempting. It is trying to price originality — and originality has become computationally hard to verify in precisely the period X has chosen to start paying for it.
The dates are firm and worth writing down if you earn on the platform.
Eligibility requires three things: an active X Premium subscription, at least 500 verified followers, and at least 500,000 qualified Home Timeline impressions from verified users in the preceding 90 days. Impressions on replies do not count.
Read that list again with the program's name in mind. Every gate is a reach threshold. None of them measures originality. A creator qualifies by demonstrating that verified users saw their posts half a million times in three months — which is a statement about distribution, not about provenance.
The originality test sits entirely in editorial guidance. X says qualifying content includes original reporting and analysis, photos and videos created by the poster, and memes or graphics they designed themselves. Commentary qualifies only if it adds what X calls meaningful original value. Disqualified: copied posts, re-uploads, and reposts without meaningful transformation. There is no stated detection method attached to any of it.
Look at the disqualification list closely. Copied posts. Re-uploads. Reposts without meaningful transformation. Every item on it describes a reposter — an account that takes existing content and redistributes it. That was the 2024 problem, and it was a real one: engagement-farming accounts scraped viral posts, republished them at volume, and collected payouts on impressions they had contributed nothing to generating.
The 2026 problem is not reposters. It is generators. And the category X most needs to exclude — high-volume synthetic content that has no human provenance — passes the originality guidance as written, because it does not trip any of the disqualifiers.
A generated image is not copied from another account. A generated commentary is not a repost. Under the rule as published, both are original.
This is not a loophole in the sense of something a lawyer found. It is a definitional consequence. Originality in the disqualification list means not-taken-from-someone-else. Originality in the aspirational half of the guidance — original reporting, photos you shot, memes you designed — means produced by a human with intent. Those are different concepts, and generated content sits exactly in the gap between them: novel by the first definition, absent by the second.
The failure mode is therefore not that a bit of AI slop sneaks through an otherwise sound filter. It is that the filter was not built to detect this at all, and the eligibility thresholds actively favour whoever can produce at volume. Half a million verified-user impressions in 90 days is roughly 5,500 a day. Volume helps. Marginal production cost near zero helps more.
There is a second approach to the same problem, and the contrast is instructive because both are being tried at once by different parts of the industry.
X's approach is account-level eligibility: decide who is allowed to be paid, using thresholds and guidance, and enforce after the fact when something looks wrong. The alternative is artefact-level provenance: attach machine-readable evidence to the content itself at the moment of creation. Suno announced on 6 August that it will begin watermarking generated songs. C2PA content credentials do the same thing for images and video by embedding a signed manifest describing how a file was made and edited. The EU AI Act's transparency obligations, which entered enforcement on 2 August, impose a machine-readable marking duty on providers of generative systems.
The structural difference is what travels. Provenance metadata attaches to the artefact, so it survives being reposted, embedded, or reuploaded by a different account — at least until something strips it, which is the standing weakness of the approach. Eligibility rules attach to the account, so they say nothing about any individual piece of content and everything about who posted it. Only one of those survives contact with a generator, and it is not the one X has chosen.
To be fair to X: artefact-level provenance is not solved either. Watermarks can be removed, C2PA adoption is thin outside a few professional tools, and a system that only labels content from cooperating generators tells you nothing about content from uncooperative ones. Neither approach works well yet. But one of them at least targets the object in question.
X's diagnosis of its own program is correct, and the company said so plainly. Allegra Jacchia, who leads creator products at X, said the old program had reached a point where its incentives were misaligned, and that creators should be focused on bringing net new content to the platform instead of maximizing payouts.
That is a fair description of what happened. The patch history supports it: X reduced payments to clickbait and aggregator accounts in April 2026, and Musk paused a local-audience weighting change in March 2026 after creator backlash. A payout formula that has accumulated that many corrections in a year is a formula that is being gamed faster than it can be repaired, and restarting from a clean rule set is a defensible response rather than an evasive one.
It is also worth noting that the reach thresholds, whatever they fail to measure, do exclude a lot. Five hundred verified followers plus half a million verified-user impressions is a meaningful bar, and it filters out the long tail of low-effort accounts that made the old program expensive. Whether it filters out the high-effort synthetic ones is the open question.
Here is what would have to be true for the new program to work as described. X would need a detection capability it has not announced, applied at the account level with enough accuracy to survive appeals, and enforced consistently against accounts that are by construction profitable to the platform's engagement metrics. Or it would need to adopt artefact-level provenance signals from the generators themselves, which requires cooperation X does not control.
Absent one of those, meaningful original value is a phrase, not a test — and phrases are adjudicated by whoever is reviewing the appeal that week.
Nothing here has run yet. Every characterisation above is X's own description of its own unlaunched program, and it deserves to be judged on what it produces rather than on what it promises. Applications open 8 September. The thing to watch is not the payout rates. It is whether X publishes anything at all about how it intends to tell the difference.
Applications open 8 September 2026, rolling out first to existing Revenue Sharing participants. You need an active X Premium subscription, at least 500 verified followers, and at least 500,000 qualified Home Timeline impressions from verified users over the preceding 90 days. Impressions on replies are excluded from that count.
DrafterDaily covers the rules that decide who gets paid on the internet — and whether they can actually be enforced.
Subscribe freeCVE-2026-0768 is an unauthenticated root RCE in Langflow. It was disclosed in January, the fix has shipped through seven releases, and attackers are hitting it in September — because low-code AI middleware became critical infrastructure without acquiring a patch owner.
Opposition to data centres is not a referendum on AI. It is a cost-allocation dispute — and PJM's capacity auction is the mechanism that turned an abstract argument into a number on 67 million households' bills.
OpenAI's first published benchmarks for its custom inference chip are real, from a public benchmark, and normalised on nameplate TDP rather than measured power. OpenAI disclosed that itself, in a sentence nobody is quoting.