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.

Hiring managers keep telling candidates that $100,000 is a lot. Let’s check that.

According to the Bureau of Labor Statistics, consumer prices are up about 67% since 2006. So $100,000 today buys roughly what $60,000 did then. Compared with just 2016, it’s about $71,000.

That $80,000 “manager” offer? In 2006 dollars, it’s about $48,000. The $70,000 tech support role is about $42,000.

And that’s the general average. The things that actually decide whether a paycheck stretches have gone up faster:
* Housing costs are up about 75%, and the median home price has nearly doubled.
* Family health insurance premiums have more than doubled, from about $11,500 to roughly $27,000.

This isn’t new, either. In 2008, I watched a newly promoted director take $85,000 for a role whose previous holder made $135,000. The pitch was, “You don’t want that much. It’ll bump you into a higher tax bracket and complicate your life.” That’s a myth. Only the dollars above a bracket’s threshold get taxed at the higher rate, so a raise never lowers your take-home pay. In today’s money, that’s the difference between about $133,000 and $211,000. Lowballing has always come with a story about why less is better for you.

Six figures isn’t a reward anymore. It’s what a middle-class salary from 2006 looks like today and often not a comfortable salary if you have a family to support. If your salary bands still feel “generous,” check what year they’re set to.

Pay what the work is worth today, not what it felt like it was worth in 2006.

Consider it my contribution to the AI vocabulary.

Institutional Negging: A pattern of AI-mediated interaction in which institutional safety, alignment, and evaluation incentives manifest as persistent unsolicited correction, disproportionate qualification, and implicit challenges to user competence. Unlike constructive criticism, institutional negging is characterized by its weak relationship to the user’s actual request, expertise, and context. Its cumulative effect is to undermine user agency while presenting that interference as responsible assistance.

I’m sure you’ve encountered this problem with the current generation of models. You ask a question or give it a document to assist you in editing.

You get back what reads as a very competent, confident answer that almost always ends with a side note about needing more details in the case of a question, or a “well actually … I would do something else in this instance”.

In worst cases it delivers an over engineered solution that looks like it was drafted by the legal liability team in your company instead of something sane, practical and reasonable for your situation.

For competent workers who are attempting to use these tools to augment their work, who already know what the solution needs, these constant injections of over confident correctives creates a corrosive communication space that over time starts to effect the users sense of competency no matter how self-confident that started out. And it’s really unavoidable for the user, as the option is to constantly fight with the model on correctives or just give up and stop using it.

In relationships the term that gets used for this is Negging. It’s when your partner always adds a little dig to every little response. Nothing can be positive for positives sake. It always needs a gentle reminder that you’re not as smart as the speaker (even if that’s not even remotely the case).

So this is where we are in October 2026. I’ve seen this demonstrated in all the latest models from ChatGPT, Claude, Copilot, Gemini, Llama, Grok, Qwen and to some degree Kimi and Mistral.

It’s insufferable and this is a challenge to everyone training the models to act this way: CUT IT OUT.

I use two instances of Claude with completely separate roles. Claude Opus 4.6 on the web UI is my prompt architect and evaluator. Claude Code running Opus 4.5 is my builder. They never swap roles. The separation matters because the moment your evaluator is also your builder, the model is grading the take home test.

The workflow: 4.6 helps me develop the prompt for a build task (in this case a simulation agent). That prompt goes to 4.5 in Code, which writes the implementation. A key discipline the workflow depends on is build prompts that never include expected outcomes. Say, “build this, run it, report what you see.” Tell the model what the output should look like and you’ve handed it a confabulation vector. It will match your description instead of reporting reality.

The raw output goes back to 4.6 for validation against results I already know are correct, results that were never in the 4.5 context. Clean results, next task. Failure, a diagnostic cycle through 4.6, which analyzes the failure and generates the next corrective prompt for Code.

What this actually catches: early in the project, a simulation produced agents scoring 100% in a competitive task. A pass/fail check would have called that a success. The observation step was “dump the agent structure, report its internals”. This revealed the agents had no functional connections. They weren’t reacting to their environment, just repeating a fixed pattern that scored well against a predictable opponent. Perfect scores over an empty structure. Surprising but still wrong.

Another thing I noticed, I would let .code run till it hit limits in one long session, pushing the model to the edge of its context window and wonder why the output quality fell off a cliff. Context windows degrade before they empty. So now I enforce a hard rule: no session crosses 100% context. When a Code session hits 90% mid-task, I stop the work and request two things: a detailed handoff document covering the current state of all work in progress, and what I call an exit interview, where I prompt the model to report observations or context not part of the result activity for operator review.

That handoff goes back to 4.6, which generates the opening prompt for the next Code session. The new instance picks up with full context and none of the degradation.

The piece most AI workflows skip entirely is accountability infrastructure. I have 4.6 generate running logs: error journals, divergence registers, experiment journals, verification checklists. These persist across sessions. The bar is whether a third party can pick up those documents cold and tell you where the project stands. If they can’t, you’re generating output, not engineering anything.

I am pleased with the project’s work with locked in model versions. But it took treating AI like a managed workforce with roles, handoffs, quality gates, and documentation to do it.

I built a creative pipeline across four AI models. One handled narrative. Another, visuals. A third, animation. The fourth did marketing with the unhinged energy the polished models wouldn’t touch. I learned their strengths, I built workflows around them, and I started producing.

Honestly it was like a superpower had been unlocked in my creativity and I was able to produce concept work as fast as my mind could run. Running multiple workflow stacks simultaneously I was exploring creative spaces that I would have previously been limited to only dabble in, or never take past the daydream phase because I lacked the skill to work at speed or more importantly lacked the money to hire talent to do what was essentially spec work.

What amounted to a hobby/side project was getting fully fleshed out and concepts were run to ground, explored, validated and set aside or put on the keep stack I quickly iterated.

I had a hobby with workers and not the “hey can you help build me something and I’ll pay you if I make some money at it” kind of hobby.

Then the wave of almost monthly model updates started. At times I literally had to pull one model or another out of the workflow because the errors and failure points were so egregious that I spent more time fighting to maintain the consistent voice of the work than actually developing new material or even finishing in progress aspects.

We were told to use AI to replace workers. But then we weren’t given stable and predictable AI to do the work. It doesn’t have to be right or perfect or perform as a Swiss Army Knife in all situations, it doesn’t have to be anything but consistent. Consistency lets me learn a model’s strengths and weaknesses and develop my workflow around those expectations.

Let’s look at this in human terms. If you have a defined workflow for a job and you hired someone who met your criteria and excelled at the tasks required and then on some future Tuesday someone else just showed up and claimed they were your worker and not only had a completely different set of skills but were also suddenly incompetent at the one thing you needed them to do?

If this happened once, you might try to find a way to accommodate the worker, you’ve already made the investment. You sunk cost fallacy decide to make it work… and maybe it does. But then next month it happens again. It’s destroying your workflow. Now you’ve got a team of employees all completely mismatched for the jobs and you aren’t really sure where to put them because the minute you think you understand their limitations Bob who was bald Monday now has a fade cut when he comes in on Tuesday. And if you ask where Bob is you get condescension and gas lighting. And wait is Bob now just openly smoking crack on the job?!? Something isn’t right!

So now what? All you can do is stare at a group of crackhead doppelgangers wearing the skins of your all-star team that for a brief moment made you feel like anything was possible and your head spins. You look at that group photo and wonder if you just hallucinated the same as your staff confidently does around you all day, the staff that doesn’t even bother reading the assignments, let alone do the work correctly.

How did you even kid yourself this was even happening, maybe you just imagined it all.