I measured what watermarking actually does to AI writing. The answer isn’t what either side thinks.

Anthropic started watermarking Claude’s text output in August and extended it to additional older models on November 3rd per EU AI Act compliance. SynthID-Text. No opt-out.

Some users claim watermarking degrades output quality. Anthropic says zero impact. I think we’re somewhere in between, and I have the data to show it.

I run RVI (rvi.run), a tool that measures linguistic variance in text. Sentence length distribution, vocabulary diversity, and structural patterns. THIS IS NOT an AI detector, it’s more like a thermometer. It doesn’t care who wrote the text, it just measures how much variety is present.

In August, before watermarking, I ran 100 structured prompts (10 genres × 10 each) across five Claude models via API. Clean baseline. On November 5th I ran the same prompts against seven models, including ones I had baselines for.

Sonnet 4.6, the cleanest before-and-after:
Aggregate RVI: 59.0 (August) -> 59.43 (November). Median: 59 -> 59. No change.

Anthropic’s claim holds… On the surface.

But broken out by category:
• Scientific review: −3.3 points
• Business writing: −2.4
• Narrative fiction: +6.2
• Personal essay: +2.6

The November outputs showed less variety in factual writing and more in creative writing. The aggregate looked flat because the effects canceled out.

This makes mechanical sense. SynthID-Text biases word choices using a keyed random pattern. In factual text there’s often only one right word… the watermark constrains an already narrow field. In fiction, where “overcast,” “grey,” “leaden,” and “dim” are all valid, the keyed randomness adds variation on top of the model’s natural distribution.

The watermark isn’t neutral. Aggregate metrics completely miss the story.

If you’re generating technical documentation or research, watermarking measurably constrains your output variety. Not by a lot, but not zero. If you’re writing fiction or personal essays, it may actually be adding variety.

The dataset is 1,100 scored essays across 11 Claude models and growing. It’s all measured through RVI at rvi.run. I’ll be adding analysis results to the site soon along with other models.

Fun fact, Adam Smiths’ “Wealth of Nations” is the lowest scoring document in my human test set at a score of 33… not a lot of information in that work it seems.

The Y2K “bug” was a serious data and infrastructure issue impacting nearly all technology coming into the end of the 1990’s.

The main reason is that from the 1960’s and forward memory wasn’t just expensive, it just didn’t exist for systems in the scale that you may be used to today. In the mid-to-late 1980’s 64k of RAM was a standard for desktop computers and if you wanted to expand that you might be able to buy 1MB and by 1990 2MB-4MB was an exotic premium.

Beyond cost and construction the #1 limiting factor was the CPUs on these early systems. They literally didn’t have the capability to address RAM in large quantities, so early on you augmented the RAM with tricks that offloaded information to the computers storage medium (we still do this even today with Swap memory on modern PCs but that’s a practical solution to not enough RAM, not the CPU anymore).

Also storage mediums were a huge expense too. In the late 80’s a 10MB (MegaByte! Not GigaByte) hard-drive for an Apple //e was well over $1000, and even with the best hardware you couldn’t actually utilize all of the 10MB as one volume. Enterprise HDs for servers were in the 5-6 figures, easily.

So when you wrote a program that needed a database, every single byte cost you hard dollars. Either in the RAM or storage. So doing things like only storing the last two digits of the year was THE standard practice up until the early 90’s as a practical cost solution. Why spend money on storing the bytes “1” “9” on every single year when that’s just wasted space.

So, all of a sudden all these systems were looking at a situation where when the year 2000 hit they would all roll back to 1900. This wasn’t speculation, this was the core crisis. Suddenly untold number of programs doing calculations on “date greater that 19XX” would suddenly report, NO. This impacted banking, finance, government records, *everything*. And many times it was not just software but embedded systems and that hardware just had to go in the trash.

The industry pulled together behind the scenes and globally replaced hardware and software across the spectrum to meet the challenge. Staggering amounts of money were spent and people put in 80-100+ hours a week of labor to fix it and the fact that we passed over that threshold without serious issues should be celebrated not minimized as a “scare”.