Alan Scott Encinas

Posted on Aug 15 • Originally published at alanscottencinas.com

The AI Reset

#humanaicollaboration #cognitiveoffloading #dataverification

There is a strange panic happening around AI right now.

Anthropic has begun embedding invisible watermarks into text generated by Claude models launched on or after August 2, 2026, with older models being transitioned later. The system is designed to survive copying, pasting, and some forms of editing, while supported files can also carry signed provenance metadata. The move comes as the European Union's transparency requirements for AI-generated content take effect, although Anthropic has chosen to apply these measures more broadly.

And judging by the reaction across Reddit, social media, and the conversations landing in my own inbox, you would think someone just announced the end of artificial intelligence. People are asking, "What are we going to do now?" I think they are asking the wrong question. The better question is: "What did we actually become while AI was becoming normal?"

Somewhere between the chatbot, the prompt box, the AI-generated essay, the automated marketing department, and the promise that anyone could make $100,000 a month with one keystroke, we crossed a line. We stopped using AI as a tool and started using it as a crutch.

That distinction matters.

AI became our universal encyclopedia. We ask it what to eat, how to exercise, whether our feelings are valid, how to handle a relationship, what to say to our boss, how to raise our children, how to care for our pets, how to lose weight, how to choose a career, how to write an email, how to start a company, how to code, how to research, and increasingly, what we should think about something. The problem isn't that AI can do these things. The problem is that we increasingly stopped doing them ourselves.

And 2026 has made that impossible to ignore.

We have watched AI move from being a productivity tool into an ambient layer sitting between people and information. At the same time, the internet has filled with something we now casually call "AI slop": fake articles, synthetic images, automated reviews, manufactured expertise, AI-generated news accounts, and entire social profiles that appear human until you spend more than thirty seconds looking at them.

But AI didn't create the content farm.

It industrialized it.

The problem isn't simply that some of this content is written by machines. The problem is that much of it has no reason to exist other than to capture attention. And that is a very different problem.

We have also created an entire economy around this phenomenon. The AI gurus arrived, and suddenly everyone was an expert. There were courses promising to replace your marketing team with AI, videos explaining how to build an entire business from your phone, consultants selling prompt formulas, and influencers promising passive income through automation. People were charging thousands of dollars to teach others things that could often be learned from documentation, experimentation, or a few hours of actually using the technology. AI became less about understanding a technology and more about selling the idea of having understood it.

That was never going to last. Eventually, the technology was going to collide with reality.

And now it is.

The Shift

The watermark conversation is one of the first visible signs of a larger transition, but it is important to understand what the technology actually does. Anthropic's watermark signals that text passed through Claude. It does not tell us whether Claude generated the entire passage, edited something written by a human, rewrote a section, translated it, or was used somewhere in between. That distinction is critical.

A person can write 90 percent of an article and ask Claude to clean up the grammar. Another person can ask Claude to generate the entire article, change a few sentences, and publish it unchanged. If both pieces retain the signal, the watermark alone cannot tell us the difference. But there is another complication: they may not both retain it.

Anthropic has acknowledged that extensive rewriting, paraphrasing, translation, or mixing Claude's output with human-written material can weaken or remove the signal. Short passages can also contain too little information for reliable detection. The detection tools Anthropic is developing are therefore not a universal test for whether something was written by AI.

Technically, that makes the system imperfect. Humanly, it makes the underlying question even more interesting.

Because we are entering a world where the question isn't simply, "Was this made by AI?" It is: "What happened between the human idea and the final artifact?" That is a much harder question.

A watermark cannot answer it by itself. Neither can an AI detector. Neither can a "Made with AI" label. Authorship, contribution, editing, translation, research, judgment, and intent are different things. That is the conversation we should actually be having.

And Anthropic is not the only company exploring this territory. Google has already deployed SynthID across AI-generated text, images, audio, and video. At its developer conference in May 2026, Google reported that SynthID had marked more than 100 billion images and videos, along with the equivalent of 60,000 years of audio. That figure describes cumulative scale since the technology's launch in 2023, not a current weekly count.

Since then, verification has also been moving beyond Google's standalone detector. Google has been integrating SynthID and C2PA verification into products including Search and Chrome, while C2PA verification has also been added to the Gemini app. The standalone SynthID Detector remains in limited testing, but the broader direction is clear: provenance is moving closer to the platforms where people actually encounter digital content.

Google has also begun extending SynthID beyond its own ecosystem. OpenAI is adopting SynthID for images created through ChatGPT, Codex, and the OpenAI API, with other technology companies including Kakao and ElevenLabs participating as well. That matters because the industry may be moving toward greater interoperability around provenance standards. But interoperability is not the same thing as universal detection.

SynthID can only identify content carrying the relevant watermark from participating systems, and a negative result does not mean something was created by a human. Open-weight models can generate content without these specific signals. Content can also move through multiple models, be mixed with human writing, translated, paraphrased, edited, or transformed until the original signal is weakened or disappears.

There may eventually be broad convergence around a handful of provenance standards. That would be useful. But it still wouldn't answer the larger question of authorship.

That is why I don't think the future is going to be about a single detector that tells us whether something is "AI" or "human." It is going to be about increasingly layered evidence: Where did this come from? What tools touched it? What sources informed it? Was it edited? Was it translated? Who made the decisions? How much of the final artifact reflects human judgment?

Those questions are harder. But they are also more useful.

The Cognitive Cost

This is where I think the current panic misses the larger transition. The goal shouldn't be to eliminate AI from creative work. It should be to make human involvement meaningful again. And ironically, I think that is exactly where this technology is taking us.

The evolution of AI has moved through distinct phases. First came access. Everyone suddenly had a writing assistant, programmer, researcher, designer, analyst, and strategist. Then came automation, as companies started asking how many people they could remove from a workflow. Then came optimization, as people began figuring out where AI actually worked and where it failed.

Now we are entering something different: verification. Not simply, "Did AI make this?" But, "Can you show me how you got here?" That is going to be uncomfortable.

It will expose companies that assumed AI could simply replace expertise. It will expose creators who built entire identities around generating content rather than developing ideas. It will expose fake experts and synthetic media farms. And it will probably expose a lot of people who have become so dependent on AI that they are no longer comfortable thinking without it. That last part may be the most important.

Because the biggest risk of AI was never that machines would think too much.

It was that humans would think too little.

We already see the consequences of outsourcing cognition. Students use AI for work designed to develop their reasoning. Professionals use it to avoid writing. Executives use it to summarize things they should probably read. Millions of people ask machines questions that, not long ago, would have forced them to investigate, experiment, talk to someone, or simply sit with uncertainty.

There is research on this. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks they had used AI for. Confidence in the tool tracked with less critical thinking, while confidence in their own ability tracked with more. The same study found that AI does not remove critical thinking so much as relocate it, toward verifying information, integrating responses, and stewarding the task.

AI didn't destroy critical thinking.

It made avoiding critical thinking incredibly convenient.

We built a remarkable technology and then discovered that convenience has a cost. That doesn't mean the technology is bad. It means we are finally learning how to use it.

The Information Loop

There is another problem that has received far less attention. AI-generated information can become input for more AI-generated information.

A model summarizes an article. Someone uses that summary to create another article. Another model scrapes it. Someone turns that into a social post. Another system summarizes the social post. Eventually, the information may travel through several layers of transformation without anyone returning to the original source.

The information doesn't necessarily become false at every step. It becomes increasingly detached from where it came from.

That distinction matters.

Misinformation doesn't always look ridiculous. Sometimes it looks perfectly reasonable. It has a headline, statistics, citations, professional language, and a confident conclusion. Everything about it can feel credible while the underlying chain of information has quietly degraded.

AI didn't invent this problem either. Humans have been copying, distorting, and republishing information for centuries. What AI changes is the scale and speed. We can now manufacture enormous amounts of plausible information at almost no marginal cost.

That is why provenance matters, but not because a watermark magically tells us what is true. It doesn't.

A provenance signal can tell us something about origin. A citation can tell us something about evidence. A source history can tell us something about how information moved. Human judgment is still required to determine whether the underlying claim is actually correct.

Those are different layers. And we are going to need all of them.

From Content to Capability

The most interesting people I know aren't using AI to avoid thinking. They are using it to think further.

They use an IDE instead of a prompt box. They build systems instead of asking for answers. They write code, test hypotheses, run models, interrogate datasets, challenge assumptions, prototype ideas, break things, rebuild them, and use AI inside that process. The difference is enormous.

One approach asks, "What should I say?" and produces content. The other asks, "Here's what I'm trying to build. Help me find the weaknesses." and produces capability. That distinction is going to define the next phase of AI because there is a fundamental difference between using AI to produce an artifact and using AI to expand what you are capable of doing.

The first makes you faster. The second makes you more capable. And that is where I think the real opportunity is.

The engineers, scientists, researchers, and creators using these systems effectively aren't using AI as a substitute for curiosity. They use it as an accelerator. They bring the problem, the judgment, the context, and the willingness to be wrong. AI helps them explore the possibility space faster. That is a very different relationship with the technology.

AI was never meant to be our identity. It was never meant to be our encyclopedia, our therapist, our conscience, or our substitute for curiosity. It is a tool, an extraordinarily powerful one, but still a tool.

The Reset

Perhaps that is what this moment in 2026 is giving us: a reset. The technology is maturing, regulations are catching up, the internet is saturated with synthetic noise, companies are discovering that automation does not automatically equal competence, and people are beginning to realize that having access to intelligence is not the same thing as possessing judgment. That realization is going to change how we use these systems.

The next generation of AI users won't just know how to prompt. They will know how to architect. They won't just generate. They will verify. They won't just ask. They will investigate. They won't use AI to replace their thinking. They will use it to extend it.

And perhaps that is the irony of this entire moment. The technology that made it easier than ever to avoid thinking may ultimately force us to become better thinkers.

The first phase of AI was about making intelligence available. The next phase is going to be about learning what to do with it.

That is not the death of AI.

It is something much more useful.

It is the moment we stop treating AI as an answer machine and start treating it as what it should have been all along: an instrument for human capability.

Frequently asked questions

Does an AI watermark prove that AI wrote something?

No. Anthropic's own documentation says a detected mark indicates the content may have been processed by Claude, and does not on its own confirm the full provenance of that content. Someone can write a passage themselves and ask Claude to proofread, translate, or summarize it, and the result may still carry the mark. The watermark is evidence that a model touched the text, not evidence of who authored it.

Can an AI watermark be removed?

It can be weakened or lost, though not reliably on purpose. Anthropic says heavy editing, paraphrasing, translation, or mixing Claude's output with other writing can strip the signal, and very short passages may not carry enough of it for reliable detection. That is a property of the system rather than a loophole in it.

Does the absence of a watermark mean a human wrote it?

No, and this is the more dangerous mistake. A negative result only means that no participating system's watermark was found. Open-weight models can generate text carrying no such signal at all, and content that has passed through several models, been translated, or been heavily edited may have lost it along the way.

What is the EU rule behind this?

Article 50 of the EU AI Act, whose transparency obligations apply from 2 August 2026. It requires providers of AI systems that generate synthetic audio, image, video, or text to mark those outputs in a machine-readable format and make them detectable as AI-generated. Anthropic has chosen to apply its marking worldwide rather than only to users in the EU.

What is the difference between using AI as a tool and using AI as a crutch?

A tool extends what you are already capable of doing. A crutch replaces the part you were supposed to do yourself. The practical test is what you bring to the exchange: if you bring the problem, the judgment, the context, and a willingness to be wrong, AI is an accelerator. If you bring only the request, it is a substitute, and the thinking that would have been yours never happens.

DEV Community