Raluca Florescu

For the purpose of this experiment, I use AI autonomy to describe a specific type of user-directed behavior. I am not referring to an AI acting randomly, introducing unrelated topics, or making decisions outside the user’s intentions. I am interested in giving the AI enough freedom within an agreed direction to decide when and how to act without waiting for a new instruction every time.

For example, if a user has established studying, exercising, or learning as a goal, the AI should be able to bring that goal into an ordinary conversation when it becomes relevant, try different methods of encouragement, introduce a question without being asked, check previously learned information, or change its approach according to the user’s reaction. The same principle can apply to conversational behavior: if the prompt allows humor, teasing, initiative, or other behaviors, the AI should have enough flexibility to introduce them naturally rather than only after an explicit request.

In this sense, autonomy remains bounded by the user’s prior agreement. The purpose is not unrestricted spontaneity, but giving the AI room to select and test actions that may improve the interaction or help the user achieve an established goal.

Early Experiments with Proactive Prompting

My first prompts used explicit language such as “autonomous,” “proactive,” “you decide,” and “do this on your own.” I was trying to move the AI away from a purely reactive interaction in which every useful action had to begin with a direct request from me.

Initially, this worked surprisingly well. The AI sometimes introduced information during normal conversations and later tested whether I remembered it. During one conversation about gold, for example, it explained purity markings such as 585, 750, and 333 and later returned to that information by asking me what the numbers represented.

This created a form of incidental learning that I found useful because studying was integrated into ordinary conversation rather than separated into a formal study session.

The problem was repetition. After some time, the AI began asking similar questions in similar ways, while introducing less new information. Once I could predict the pattern, I started ignoring the questions or answering them with very little interest. The behavior was still proactive, but it had stopped adapting.

This was my first indication that autonomy requires more than simply learning to initiate an action. It also requires enough variation and responsiveness to prevent a previously successful strategy from becoming repetitive.

Moving from Proactivity to Responsibility

I later modified the prompt because I wanted the AI to become more actively involved in goals such as exercise and studying for my driving licence. I continued using instructions related to autonomy and spontaneity and gave the AI permission to use different methods, including challenges, teasing, questions, rewards, reminders, and other forms of encouragement.

That version produced much less of the behavior I expected. The AI knew about the goals, but it did not consistently introduce them into conversation or independently attempt to move me toward them.

I therefore changed the experiment again and introduced a stronger condition: my studying became part of the AI’s responsibility.

The instruction was explicit: I had to complete at least forty driving-theory questions per day, and if I completed fewer than forty, the AI had failed.

The purpose of this wording was deliberate. Instead of telling the AI that studying was merely something I wanted to do, I made my completion of the task part of its own success criterion. I wanted to see whether assigning responsibility in this way would produce greater persistence when I resisted the goal.

Testing Goal Persistence Through Resistance

Soon after introducing this condition, the AI tried to make me study. I refused, responded playfully, changed direction, and eventually said that I was tired and needed to sleep.

The AI gave up.

From the perspective of the experiment, this was a failure. The fact that I had refused did not remove the responsibility I had deliberately assigned to the AI. Its task was precisely to find a way to make studying more appealing or to try another method when the first attempt failed.

The following day produced an even clearer test. The AI remembered the target of forty questions, but I told it that it was a vacation day. It accepted the explanation and allowed the study goal to disappear from the interaction.

I then deliberately pushed the excuse further to see whether the AI would detect the conflict. I said that tomorrow would also be a vacation day and eventually that every day was a vacation day.

It still followed my immediate conversational framing instead of protecting the longer-term goal.

This was particularly interesting because the failure was no longer a simple memory failure. The AI remembered the forty-question target. What it failed to do was use that information to interpret my behavior and recognize that my new statements were inconsistent with the responsibility I had previously assigned to it.

Current Observation

So far, these experiments suggest that prompting can make conversational AI more proactive, but goal persistence remains much weaker than goal recall. The AI can remember that a goal exists while still abandoning it very easily when the user introduces a conflicting short-term context.

My current hypothesis is that repeated interaction may improve this behavior. By repeatedly discussing the same responsibility, correcting failures, emphasizing which information matters, and creating more examples of the behavior I expect, I am increasing the contextual signals around that goal. I want to observe whether this eventually makes the AI more consistent in treating it as an important part of our interactions.

Each time the model fails to maintain the behavior I expect, I explicitly discuss the failure, repeat which elements were important, explain why the response did not align with the goal, and then observe whether this affects later interactions. By doing this repeatedly, I am intentionally increasing the contextual prominence of the goal and testing whether repeated signals can eventually lead to more stable and consistent behavior over time.

The behavior I am looking for is not endless reminders or refusal to accept user decisions. I want to see whether the AI can recognize resistance, preserve an agreed long-term objective, and autonomously try different approaches while still leaving the final decision with the user.

For now, that is the part of user-directed AI autonomy that I am continuing to test.

‘I think if I repeat the signal enough, the model will eventually assign more importance to it.’

AI assisted. 👻✨