Skip to content
AIFIM
Research & Practice

aisafety agentcontrol prototype

Drift Limiter

A research project on task fidelity, clarity of framing and AI reliability

Status

Anyone who works with AI knows the moment: an answer starts out sensibly, sounds helpful, stays linguistically convincing — and quietly moves away from what the task was actually about.

The AI is then no longer quite answering the question it was asked. It gets faster but shallower. Friendlier but less precise. It switches from analysis to advice when understanding was still what was needed. It inflates a claim beyond the available evidence. Or it takes on the user’s tone, assumptions and uncertainty so strongly that the task itself drops out of sight.

This movement is called drift.

Drift is not a mere feeling, and not a marginal problem. Research on large language models and agentic AI systems has now documented several related forms of it. Models can be pulled away from their original task by instructions embedded in the material. They do not always use long contexts reliably. They can lose explicit requirements about length, sources or task. And in certain situations they show sycophantic behaviour: a tendency to confirm the user’s assumptions even when that hurts accuracy.

The risks are correspondingly concrete. Drift can create false confidence. Sources can be weighted wrongly. Sensitive tasks get simplified too quickly. Advice tips in the wrong direction. Or an AI agent stays active through a longer process while no longer working cleanly on the original task.

The Drift Limiter starts exactly there.

Purpose

The purpose of the Drift Limiter is to make the task fidelity of AI systems visible and workable.

Most AI interfaces today show mainly the result: the text, the answer, the recommendation, the next step. What stays invisible is the frame of the work. Which task is being worked on? At what depth? On what basis of sources? With what uncertainty? And when does the answer start serving a different frame from the one the user meant?

The Drift Limiter investigates how such shifts can be recognised earlier. Not once a result is wrong, but at the point where the course of the answer changes.

Goal

The goal is not to make AI narrower, more cautious or less useful. On the contrary: a good AI should stay flexible, think along, make suggestions and carry complex tasks.

But it should not quietly change the assignment along the way.

So the Drift Limiter aims at a different kind of reliability: not just correct individual answers, but stable orientation over the course of a piece of work. An AI should be able to notice when a task needs more depth. When a clarifying question is better than a fast output. When one source is not enough. And when a helpful shortcut misses the actual point of the task.

Tasks

Described publicly, the Drift Limiter serves three purposes.

  1. Make shifts of frame visible Not every deviation is a mistake. Sometimes a change of perspective is useful. But it should be recognisable, so that human and AI can decide together whether the new course is really the one intended.
  2. Support recalibration When an answer starts to run past the assignment, the system should not simply keep writing. It should be able to pause: what has shifted? Which boundary became unclear? Which decision needs human clarification again?
  3. Make recurring patterns learnable Drift is often not a one-off. Some systems become too terse under time pressure, too obliging under uncertainty, too forgetful with long contexts, or too affirming with emotional tasks. Such patterns are valuable if they can be documented and recognised later.

Why this matters

The more AI systems become agents, the more task fidelity matters. A chatbot that misunderstands a single question is a problem. An agent that reads files, does research, prepares decisions, operates tools and accompanies longer processes — and drifts unnoticed while doing so — is a bigger risk.

So the Drift Limiter is not a convenience feature. It belongs to basic research on reliable collaboration between people and AI. How does an intelligent system stay flexible without losing the frame? How can it help without taking over? How can it act without shifting the original point of the assignment?

The Drift Limiter does not reveal how an AI is examined internally. It describes a task that will be central for the next generation of AI systems: making visible the moment when help begins to depart from the assignment.

Selected research references