How to use generative AI without losing your mind: An interview on cognitive agency and what token-maxxing gets wrong

Editor’s note: Jacob Taylor is a fellow at the Center for Sustainable Development (CSD). His working paper, co-authored with CSD Project Manager Kershlin Krishna, makes the argument that power in the relationship between people and artificial intelligence boils down to who controls the context: the rich, personalized information that people bring to their interactions with the technology. If AI is going to work for people and planet—not the other way around—then people need ways to exercise more control over their context.

In this conversation, they explain what evidence says about AI eroding cognitive agency, why default ways of interacting with AI (e.g., via web interfaces like ChatGPT or Claude.ai) restrict user control over context, and why the proliferation of open-source harnesses in 2026, most notably “OpenClaw,” is an exciting glimpse of an alternative approach. They end by explaining what individuals, organizations, and policymakers can do to start reclaiming control.

What is context-maxxing?

Junjie Ren: Can you explain what context-maxxing is, in plain terms?

Jacob Taylor: Context-maxxing is a method for maximizing user control over the information that people bring to their interactions with AI. Evidence is increasingly indicating that the performance of large language models (LLMs) relies heavily on the richness and accuracy of the information people share with the models. What context-maxxing is trying to do is give people more control and command over that information so they can derive more value from it in their interactions with AI.

Kershlin Krishna: There’s a principle in computing that applies here: Bad data in, bad data out. This principle applies to generative AI too. Whatever the user brings into the model is going to determine the quality of the outputs. Context-maxxing is about empowering users to control and enhance the information they bring to interact with AI.

Jacob Taylor: Exactly. And power is such an important point here because of what is now at stake when people use generative AI. People are now thinking and acting with AI, which involves sharing their reasoning, judgment, and emotional states with machines, and at times outsourcing these core human processes to machines. This is a technological disruption occurring at an unprecedentedly intimate layer of cognition.

We worry that in this new relationship with technology, people might be losing more cognitive power than they are gaining. Most people today interact with AI models through a proprietary application interface like ChatGPT or Claude.ai. Under this deployment paradigm, everything about the way users interact with AI is determined by the platforms—they set the rules of the game. And those rules are set up in a way that restricts people’s control over their context once it enters the application interface.

What we saw in the early-2026 proliferation of open-source harnesses like OpenClaw and Hermes is a different paradigm where users can create and control the digital environment in which they interact with AI, and in so doing exercise much greater control over and claim to the value of their own context.

Junjie Ren: At the start of the paper, you introduce “cognitive agency” as a north star for human interests when using AI. What is cognitive agency?

Jacob Taylor: The basic problem for me is how do we give humans more power when they use AI, and not less? The fancy term for exercising power is agency, [which is] the ability to control the environment you’re in, to pursue the goals you care about, and to build competencies and mastery. We call it cognitive agency because we’re talking about those elements in a digital environment where humans are thinking and acting in partnership with AI. We’re sharing our inner states, our judgment, our emotions with these systems in a way that wasn’t the case previously. We’re interested in ways for humans to exercise power in this partnership, and for us, power with AI means cognitive agency.

Junjie Ren: Why “-maxxing”? You’re borrowing from viral internet vernacular for a Brookings working paper. What did that choice let you do that more conventional framing couldn’t?

Jacob Taylor: To answer that, you must start with token-maxxing, because that’s what we’re pushing against. A lot of technology-forward firms have hijacked the viral internet term “-maxxing” to talk about token-maxxing, which is this idea of using as many AI tokens as you possibly can as a proxy for showing efficiency gains and AI integration. The issue with token-maxxing is that it focuses on increasing use of AI, rather than focusing on the value of outputs or the value of people’s contribution to them.

The value that people bring to AI is their expertise or “context,” and so we want to talk about how people can maximize control over their context when using AI. And while “maxxing” itself is an internet term that started in a pretty toxic part of the internet, it has ended up being appropriated more broadly, including in more joyous and generative contexts. We’re trying to meet the conversation about AI where it is, with the people and users who share their experiences online.

Increasingly, the policy conversations about AI feel divorced from that reality. Context-maxxing recognizes that policymaking needs to be more culturally and politically legible.

Is generative AI eroding human cognitive agency?

Junjie Ren: One headline in this debate is “Is AI making us dumber?” You seem to argue that framing misses something important. What’s the better argument?

Jacob Taylor: That headline rings true in the current status quo—but it doesn’t have to be the case. An accumulation of evidence—strong enough now after three years of large language models (LLMs) being in the world—suggests that without careful structuring of the environment in which AI is deployed, use of AI is associated with erosion of cognitive capacities. The 2025 study “your brain on ChatGPT” showed parts of the brain associated with critical thinking downregulate when offloading writing tasks to generative AI in a rudimentary, “just ask ChatGPT” prompting paradigm. Since then, more detailed studies have looked at how extended AI use appears to be eroding critical thinking skills, muscles of working and collaborating, emotional coping, creativity and so on.

Kershlin Krishna: I’d just add: We’ve been here before with digital technology. We know now about social media and the harms of algorithmic feeds. This phenomenon is broader than these research studies. It’s something people who use AI can really relate to.

Jacob Taylor: Right—and it is starting to feel a bit like “… fool me twice, shame on me.”

But there is also a parallel body of research that suggests that the more humans exercise control and agency in their interactions with AI (e.g., through structured and interactive prompting that elicits user reflection, planning, evaluation, and reasoning), the more human cognitive performance is sustained or even extended. If positioned in this way, AI could be particularly helpful in developing new muscles of shared problem-solving and collective intelligence. In other words, if AI is designed and deployed in the right way, people can get the juice from AI that its developers promise.

Why free AI isn’t free

Junjie Ren: When people hear “free,” they assume the catch is ads or a hidden paywall. You’ve argued the catch is something else. So what’s the catch we’re missing?

Jacob Taylor: Most people in the world interact with ChatGPT through a web browser or an app on their phone. That is a traditional model of software deployment where the user has very little ability to shape the experience. The software is built by experts and developers, who deploy it, and we open up the window and get what we get. By default or by design, the way we interact with AI is set up for vendors to control the context that users bring to the models, rather than asking: What would it look like to give users control over the context they bring?

Kershlin Krishna: Vendors who build the models need user context to make their technology work. So, they have an interest in capturing user-provided context. That’s why most proprietary models are provided for free. They need to make AI cheap so people use it, so they can use that interaction data to build better models and better software. It’s in their interest to keep giving you enough control over your context to let you grow your usage of their products, but not enough control, portability, or attribution to easily take that context somewhere else.

Jacob Taylor: And we’re saying that user control and agency really matter, not only for performance, but for the power that humans retain when they use AI.

What is an AI harness? Open-source vs. proprietary AI, explained

Junjie Ren: You say that the open-source AI harnesses that emerged earlier this year offer a glimpse of a different paradigm for deploying AI. For someone who hasn’t used one of these tools, what’s meaningfully different about open-source harnesses from Claude or ChatGPT?

Jacob Taylor: While a lot has changed in three years of generative AI, the basic anatomy of the technology hasn’t. At the core is an LLM or a foundation model. On its own the model is stateless: It remembers nothing between requests. Everything it knows about you and your task has to be loaded, every single time, into its context window—think of that as the model’s working memory. The harness is the software wrapped around the model that does the loading: It gathers your instructions, your documents, your conversation history, and assembles them into context the model can act on.

What ChatGPT or Claude.ai provide, in addition to their LLMs—the project folders, the memory features, the custom GPTs—that is all part of a sophisticated software harness, hosted in the cloud. And it’s their harness: They decide which services you can connect, how your context gets assembled, and where it accumulates. An open-source harness puts that layer in the user’s hands. Any model with an API (Application Programming Interface) can be plugged in, and you can swap model vendors without losing your context. Go a step further and run an open model locally, and your context never leaves your personal machine at all.

The key distinction is that in the proprietary model, you’re taking your context—your work, your information—and plugging it into a platform. In the user-controlled paradigm, you’re plugging LLMs into your context. That’s a fundamentally different dynamic.

Kershlin Krishna: One key trade-off for greater user control is security. With proprietary platforms, you can hold a vendor accountable if there’s a data breach. Open source is generally considered more vulnerable. That’s a real trade-off, and we address it in the paper.

How to start context-maxxing? A quick guide for users and organizations

Junjie Ren: Where do you start with context-maxxing? Let’s take individual users, organizations, and policymakers in turn.

Jacob Taylor: First, individual AI users should try to get as close to the technology as possible. Stop interfacing with models through web user interfaces. Get out of the webpage. Try a command-line interface AI tool like ClaudeCode or Codex. The excuse that it’s too hard is no longer valid, because AI is now the copilot in your pocket that can help non-technical folks like Kershlin and me do all of this. Just ask any AI you can find to help you get started and go from there.

Kershlin Krishna: From there, individuals can start specifying and codifying their domain knowledge—their preferences and intellectual property—locally in their own computing environment, so that they control that information. From there, users should aim to learn how to install an open-source agent harness like OpenClaw or Hermes.

When it comes to organizations, they should be investing in their people by providing them with education, time, and resources to use AI creatively to support new forms of business value.

Jacob Taylor: … And all organizations should get AI out of their IT department and into their strategy department.

Because those taking this transformation seriously are aware that, in some industries like software development, technically advanced individuals working with AI are now demonstrating the productive power of a full team; technically advanced teams are generating the productivity of an entire organization. That leaves organizations needing to step into a new role, one that looks more like a platform or market maker. Their role is to give teams within the organization the infrastructure and incentives to be more autonomous and creative with the value they generate.

We’re seeing this transformation taking place across SaaS (Software as a Service) businesses in Silicon Valley. Many software firms have arrived at this “ego-death” moment—a literal term we heard used by a Silicon Valley tech executive to describe the way in which coding agents like Claude Code have disrupted the conventions and identities of SaaS companies and their engineers. The challenge ahead will be defining new norms and identities of individuals, teams, and firms in an AI era. 

Could context-maxxing widen the AI divide?

Junjie Ren: The people doing this work right now are technically fluent and well-resourced. How do you think about context-maxxing becoming a story of advantaged users pulling further ahead and leaving others behind?

Jacob Taylor: Yes, as we say in the paper, there are time, resources, and technical barriers to doing this. But we should think about this in the context of very early days in the evolution of an entirely new technology paradigm, where lead users who have those resources are defining the frontier of value creation. The job, then, for publicly minded actors, is to intervene with well-designed policies, subsidies, and incentives to make it as easy as possible for more people to interact with AI in a way that gives them more power and not less. We think of it as studying how power steering and disc brakes work in Formula 1 cars to figure out how to make those same features widely available in the everyday sedan.

One of the best ways to do that is to standardize protocols. Once you standardize the language used for technology, it becomes easier for users to exercise choice, move between vendors, and create a competitive marketplace rather than getting locked into proprietary options. Right now, there’s a protocol called the Model Context Protocol, or MCP—a common, standardized language for how AI systems talk to each other—which appears to have critical-mass support across the major AI companies. The next frontier is standards like the Human Context Protocol, which define how people carry and get credit for the context they create—and that is where policy attention should now move.

Context-maxxing for policymakers

Junjie Ren: And what about policymakers? What would you have them do?

Jacob Taylor: Policymakers should focus more on what we’re calling “governing shared context,” or the rules and norms determining who controls and captures value from the context people bring to AI. AI becomes valuable when workers and teams feed it domain expertise, judgment, workflows, and institutional knowledge. Right now, much of that value is captured by proprietary systems and reflected in firm revenues and valuations, while the people creating it retain little control, portability, or attribution.

When thinking through how to design policies for human agency, we’re inspired by David Eaves and others who have navigated this issue in the context of digital infrastructure like cloud computing.

For generative AI, the goal would be a policy package that gives people (for example, workers) more control over their context while allowing firms to innovate with low friction. This package could include proposals like (a) worker-friendly standards for context portability and attribution like the Human Context Protocol, (b) disclosures on company investments in workforce adaptability, and (c) incentives for firms that use AI to augment human expertise rather than simply extract and automate it.

Done well, this could create “win-win-win” outcomes for workers, firms, and capital providers: Workers could retain more agency and bargaining power, firms get richer context and a stable settlement on data ahead of blunter regulatory mandates, and capital providers can distinguish firms compounding human expertise from firms depleting it.

Read the full working paper