AI can clear an inbox, summarise a meeting and prepare a first draft before you have finished your coffee. The awkward part comes later: what did you understand, what did you merely approve, and which decisions quietly moved out of your hands?
That question has started to appear inside AI products themselves. On 9 July, Anthropic announced a beta feature called Reflect with Claude. It lets users review topics and patterns across one, three, six or twelve months of activity. It can also ask which tasks a user wants to keep doing personally, and it offers quiet hours and break reminders.
The feature is interesting because it asks an AI to help its user decide when not to use AI. That sounds almost comic, but it points to a serious design problem. A capable personal assistant should save time without making delegation invisible.
Faster work is not the same as better thinking
Productivity is easy to notice. A finished email or a compressed document gives immediate feedback. The cost of cognitive offloading is harder to see because it often appears later, when someone needs to explain a decision, catch an error or work without the tool.
Research does not support a simple claim that AI makes people less intelligent. It does show that the way people use AI matters.
A Microsoft Research study presented at CHI 2025 surveyed 319 knowledge workers and collected 936 examples of AI-assisted work. Higher confidence in generative AI was associated with less reported critical-thinking effort. Higher confidence in one's own ability was associated with more. The researchers also found that critical thinking did not simply disappear. It shifted towards checking information, integrating responses and supervising the task.
That study was based on self-reports, so it cannot prove that AI caused the change. A more controlled Anthropic experiment on coding-skill formation found a narrower but useful result. Developers learning a new Python library with AI assistance scored 17% lower on an immediate mastery test than the group working by hand. The sample was small and the task specific. More revealing was the variation inside the AI group: people who asked for explanations and used the assistant to build understanding did better than those who gradually handed over the whole task.
The practical lesson is modest. Delegation is not automatically harmful, but passive delegation can leave a gap between output and understanding.
Separate delegation from adoption
When people talk about trusting AI, they often combine two different decisions.
The first is delegation: should the system do this task before you see the result? The second is adoption: once the system offers an answer, should you accept it?
A 2026 study of human-AI question-answering teams treats these as separate choices. In its experimental game, human-AI teams did better than either humans or AI working alone, yet participants still relied too much in some cases and too little in others. The study is a preprint with a small, artificial setting, but the distinction is useful well beyond the experiment.
You might happily delegate the search for three flight options, then inspect the proposed itinerary before booking. You might ask AI to summarise a contract, yet refuse to adopt its interpretation of a liability clause without reading the source or consulting a qualified adviser. You might delegate the formatting of a board paper while keeping the recommendation and its consequences entirely human.
Once those two decisions are separated, “use AI” stops being a yes-or-no choice.
A four-part practice for using personal AI
The following framework is deliberately simple. It works for a chat assistant, an agent that can use tools, or AI built more deeply into a phone.
1. Retain the judgement you need to own
Keep direct control when a decision carries accountability, depends on tacit context or develops expertise you cannot afford to lose.
Before involving AI, write one sentence that states your current view. For a hiring decision, that might be the capability you think the team lacks. For an investment memo, it could be the assumption most likely to break. For a difficult email, it may be the outcome you are trying to protect.
This small act gives the assistant something to challenge instead of letting its first answer become your starting position.
2. Delegate the work, not the standard
AI is well suited to gathering, sorting, comparing, transcribing and preparing options. Delegating those steps can preserve attention for the parts that need context and judgement.
Define the acceptance test before the assistant begins. Ask for sources, dates, uncertainties and rejected alternatives. If an agent can take actions, state which ones require confirmation. “Find three options under these constraints” leaves ownership with the user. “Handle the trip” hides too many decisions inside one instruction.
3. Inspect the reasoning behind the answer
Fluent output is easy to approve. Inspection should make the rough edges visible again.
Check the source behind a material claim. Ask what evidence would change the recommendation. Request the strongest case against the preferred option. For work that builds skill, ask the assistant to explain a step, quiz your understanding or leave part of the task for you to complete.
The goal is not to repeat everything manually. It is to keep enough contact with the reasoning that you can spot a wrong turn and explain the final decision.
4. Review and revoke
Personal AI becomes more useful when it remembers preferences and connects to other services. That usefulness creates a maintenance job. Old access can outlive its purpose. Saved context can become inaccurate. A workflow that once needed autonomy may later deserve tighter limits.
Once a week, review what the assistant handled, which recommendations you accepted and where you corrected it. Then inspect connected apps, active permissions and stored memory. Remove what no longer helps.
Reflection without the ability to change a boundary is only a report. A well-designed personal AI should let the user act on what the report reveals.
What reflective AI should look like on a phone
Phones hold messages, calendars, documents, travel plans and the small decisions that connect them. An AI agent close to that context can be far more useful than a chatbot in an isolated window. It also needs clearer limits.
Four design choices follow from the practice above: memory should be curated rather than mysterious; significant actions should wait for confirmation; access boundaries should be visible; and permissions or memories should be revocable.
These principles are part of how VERTU describes Hermes Agent, its private AI agent for selected VERTU devices. The official product information describes curated memory, human-guided execution and confirmation for significant actions. It also says users can review or withdraw permissions, integrations, sessions and memories. Supported apps and features vary by device, service, configuration, software version and user authorisation.
Hermes Agent is not claimed to include Anthropic's Reflection feature. The connection is conceptual: both point towards personal AI that makes the user's role explicit. Capability matters, but so does the ability to see, approve and reverse what the system is doing.
Readers who need a more detailed access-control checklist can use VERTU's guide to AI assistant privacy, permissions and memory. That guide covers the technical controls; this article stays with the human habit of deciding what to hand over.
A ten-minute weekly AI review
Set aside ten minutes and answer five questions:
Which task did AI complete that genuinely saved me time?
Which output did I accept without checking?
Where did I correct the system, and should that correction become a rule or saved preference?
Which decision helped me build judgement and should remain mine next time?
Which permission, integration or memory no longer needs to exist?
The answers do not need to produce a perfect personal policy. They simply turn invisible habits into choices.
AI will keep getting better at finishing work. Users will need to get better at deciding what “finished” means, what deserves inspection and where accountability stays. The most useful personal AI will not remove the human from that loop. It will make the loop easier to see.
Frequently asked questions
Does using AI reduce critical thinking?
Not inevitably. Current research suggests that outcomes depend partly on how the tool is used. Passive reliance can reduce cognitive effort or short-term mastery in some settings, while explanation-seeking, verification and active supervision can preserve more understanding. The evidence is still limited and should not be generalised to every person or task.
What tasks should not be delegated to AI?
Keep control of decisions where you carry the accountability, where important context is not available to the system, or where doing the work is necessary to maintain expertise. AI can still gather evidence or challenge your reasoning without making the final choice.
How often should I review an AI assistant's access?
A brief weekly review is useful for active workflows. Also review permissions after a project ends, when a connected service changes, or when the assistant has handled unusually sensitive information.
