Your Digital Agent: What Do You Delegate, and What Do You Keep?

In an American experiment, 543 people were invited to join a fictitious social network called NameDrop. Its terms of service included a clause making "your first-born child" the price of membership. Some 93% agreed to the terms, and about 98% failed to notice the planted clauses. They were not gullible: they spent about 51 seconds on the terms on average, while the researchers estimate that reading them takes between 15 and 17 minutes.

Delegation does not begin when we think it does

A new generation of digital assistants is described as an "agent": it does not merely answer your question but books, writes, compares and suggests. The first delegation its owner grants is not a considered decision but a click on "I agree". That is the lesson of NameDrop: we delegate by habit before we delegate by judgement. Once an agent acts in your name, you need to know what you actually allowed it to do, not what you assume you allowed.

What changes in our minds?

Evan Risko and Sam Gilbert define "cognitive offloading" as the use of physical action to alter the information-processing demands of a task so as to reduce mental effort. It is an ordinary everyday behaviour, such as setting a phone reminder, and not harmful in itself. In 2011 Betsy Sparrow and colleagues found that people who expect information to remain available remember where to find it better than they remember the information itself.

The evidence on modern AI assistants is still preliminary. Michael Gerlich found, in a survey of 666 participants in Britain, a negative correlation between frequent use of AI tools and critical-thinking scores. But he relied on self-report, and it is possible that people with weaker critical habits lean on tools more, rather than the reverse. In a smaller study (54 participants, posted on arXiv before peer review), the group that wrote with a large language model showed weaker neural connectivity on EEG than the search-engine group and the no-tool group. These are signals worth scrutiny, not final verdicts on "weakening brains".

Whom does it work for?

An agent that knows your calendar, messages and habits holds a picture of you that even those closest to you do not. You need to ask: who owns this agent? Where is my data stored? Can I review what it did in my name? The OECD AI Principles, updated in 2024, say that AI actors should implement mechanisms and safeguards, such as capacity for human agency and oversight, as appropriate to the context. A human in the loop is a written principle, but it is realised only when we practise it.

Practical rules for delegation

• Delegate reversible routine: scheduling an appointment, summarising a text, sorting a list.
• Keep what cannot be undone or is governed by values: a message of reproach, a large financial decision, a matter touching family and trusts.
• Read what you are asked to approve: access to your account, correspondence and payments first.
• Review the agent's log from time to time, and keep hours in which you think, write and choose for yourself.

In all of this, we take the means while relying on God; the trust of a decision is its owner's responsibility and is not handed to a programme. As for how these tools will develop in the years ahead, only God knows the details. What is certain is that every delegation we grant today is one we can withdraw if we read it well.