Staying in Control as AI Becomes More Capable

I use AI considerably more than I did a year ago, yet the conditions under which I am willing to trust it have become narrower, more explicit, and much more dependent on the nature of the work.

When I reread the material I developed for Practice Safe AI in 2025 and Stay in Control When AI Joins Your Work earlier this year, the underlying principles still held: protect your data, protect your judgment, and protect your human-ness. The limitations were less about those principles than about the operating assumptions beneath them, because my guidance largely presumed an interaction in which I initiated the work, supplied the context, received an output, and remained responsible for deciding what happened next.

My own AI use now looks quite different. I rely heavily on AI for research and brainstorming, increasingly begin searches through AI rather than traditional search, use persistent memory when continuity adds value, create specialized AI partners for different roles, and ask AI to challenge assumptions rather than merely answer questions. At the same time, I remain reluctant to give autonomous agents broad access to applications or data, continue to validate important sources independently, and have become much more sensitive to the point at which AI assistance begins to dilute my own reasoning or voice.

My earlier AI use case patterns often began with a prompt, followed by several rounds of refinement and whatever useful rabbit holes emerged along the way; increasingly, I begin by defining the job to be done, determining where AI can materially improve the work, deciding where I need human checkpoints, and only then engaging the tool. Recent research on human-AI collaboration reinforces the value of this upstream discipline, arguing that as AI assumes more of the production of knowledge work, explicit goals become increasingly important to planning, delegation, coordination, and evaluation.

Four areas in particular have evolved in my thinking about staying in control: when control needs to begin, what data protection now encompasses, where human judgment needs to enter the work, and how deeply I need to participate in the thinking if I want to retain the knowledge afterward.

Control starts before delegation

Prompt quality still matters, although increasingly capable models make sophisticated prompting less central than it once appeared. A beautifully constructed prompt can improve execution while leaving untouched the harder questions about whether the task should have been delegated, whether the problem has been framed correctly, which constraints deserve scrutiny, and whether the person using AI has enough expertise to recognize a plausible but incomplete answer.

For consequential work, I now spend more time defining the outcome before deciding how AI should participate. In some situations AI becomes a researcher or analyst; in others it is more useful as a critic that challenges my assumptions, introduces perspectives I have not considered, or interrogates reasoning I have already developed. Sometimes I intentionally keep it out of the earliest stages because I do not want its first formulation to become the anchor around which my own thinking develops.

This distinction becomes particularly important as AI moves beyond conversational assistance into longer-running, increasingly agentic work. Stanford researchers examining roughly 250,000 real AI conversations in 2026 found that people still directed most collaborations, while the inevitable friction of working with AI often improved the outcome because users had to clarify intentions, correct misunderstandings, and remain actively engaged rather than simply accepting the first result.

Data protection now includes permission architecture

I still deidentify information when appropriate, avoid placing personal financial or health information into general AI systems, and think carefully before sharing any personal information. Those habits address only the information I consciously provide.

Connected AI introduces another problem: information may already be accessible because of permissions established for an entirely different technology environment.

I encountered this at work when AI made information discoverable through existing SharePoint permissions that had not been created with AI-mediated retrieval in mind. The AI had not granted access that did not exist; it had dramatically lowered the effort required to find, combine, and make use of information that was technically available already.

That experience made data protection much less about the contents of a prompt and much more about the underlying architecture of access. Before connecting an AI system, I increasingly want to understand not only what it can see, but which permissions it inherits, what it can remember, whether its access can be limited to read-only, which actions it can initiate, and where explicit approval is required.

As autonomous agents become practical rather than experimental, this area concerns me considerably more than prompt privacy alone.

Human review depends on where the human participates

The common recommendation to “keep a human in the loop” has also become inadequate for me because it says almost nothing about the intellectual contribution of the human.

A person who reviews a completed recommendation after AI has defined the problem, selected the evidence, developed the alternatives, and drafted the rationale is unquestionably in the workflow, yet may have participated in very little of the reasoning that produced the result.

I have seen the limitation most clearly when using AI against complex engineering problems. Models can generate several technically plausible approaches quickly, while still missing physical constraints, organizational history, tacit knowledge, or an unconventional reframing that an experienced engineer recognizes because the problem is understood in a context the AI cannot fully reconstruct.

Human judgment therefore matters well before approval. Problem definition, deciding which constraints are actually fixed, determining what evidence would invalidate an approach, creating alternatives that do not follow the obvious pattern, and recognizing when the premise itself deserves reconsideration may represent the most important human contributions.

The more useful question for me has become less about whether a human remains somewhere in the loop and more about whether human judgment enters the workflow early enough to alter its direction.

Cognitive depth requires participation

The least visible risk has been emerging through my own learning.

AI allows me to cover far more material than I could otherwise absorb in the same amount of time, particularly when I am entering an unfamiliar technical domain or conducting broad research. Yet I have also noticed that when AI performs too much of the initial reading, organization, and synthesis, I can recognize the conclusion later without being able to reconstruct the reasoning that led to it, retell the argument with the same richness, or recall the underlying evidence well enough to teach the topic without returning to the generated artifact.

Possessing the answer and possessing the understanding are not equivalent.

Recent research is beginning to examine this distinction more carefully. Microsoft Research is actively studying how AI can function as a “tool for thought” that provokes critique and metacognition rather than merely completing work, while a 2026 randomized study of software developers found lower mastery among participants using AI assistance, with substantially better outcomes among those who used AI to ask conceptual questions and build understanding rather than simply generate code.

Writing exposes the problem in a different form. I have reread posts that AI helped refine and encountered phrases that were grammatically polished and perfectly serviceable, yet unmistakably not mine. The problem was not factual accuracy; I had allowed enough of the formulation to move outside my own reasoning that the resulting prose had lost the idiosyncrasies through which my point of view normally becomes visible.

Sometimes the friction I am tempted to eliminate is performing useful intellectual work.

My current control check

For work that matters, I increasingly ask five questions before and during an AI-assisted workflow:

1. What job am I actually trying to accomplish?
2. What contribution do I want from AI, and which parts of the work need to remain mine?
3. What information, memory, applications, or permissions does the AI genuinely need?
4. Where will I independently challenge the reasoning, verify the evidence, or introduce expertise the AI does not possess?
5. If the AI disappeared after producing the result, could I explain it, defend it, modify it, and account for the decisions embedded within it?

While rereading my earliest DeeplyProduct posts for this article, I returned to a piece I wrote in 2020 about accountability for AI applications. At the time, I was thinking primarily about the responsibilities of the people defining, designing, developing, and deploying AI systems, and I ended the post questioning whether accountability itself was the right long-term problem to solve.

Six years later, I see another layer of that same problem. Responsibility increasingly sits with the people adopting these systems as well, particularly when an apparently successful outcome can obscure weaker understanding, inherited access that was never reconsidered, diminished authorship, or a human review process that exists formally without contributing much judgment.

My AI use will almost certainly continue to deepen, which makes staying in control less a matter of limiting the technology than of becoming increasingly deliberate about the intellectual work, access, authority, and accountability I am willing to hand over.