Concern 32 · Working draft · editorial review in progress
RLHF preference poisoning
Preference poisoning corrupts the human-feedback examples used to teach an AI which answers people prefer.
Understanding the concern
In reinforcement learning from human feedback, or RLHF, ratings help shape later behavior. If an attacker introduces misleading preferences, the system may be rewarded for responses that legitimate reviewers would not want.
The issue is deliberate manipulation, not ordinary disagreement among reviewers. A study of poisoned preference datasets showed changes in model sentiment toward chosen entities. That finding supports the mechanism, but does not establish how often commercial training pipelines are affected.
An example
A cooking contest becomes misleading if planted judges reward the dish they were paid to favor rather than the dish they consider best.
Illustrative unless identified as a documented case.
Words in plain English
- AI
- Software that learns patterns from examples or uses rules to perform tasks such as recognizing images or generating text. It can make mistakes and does not understand every situation as a person would.
- Model
- The learned part of an AI system. It uses patterns found during training to produce an answer or prediction.
- Training
- The process of using examples to shape how an AI model behaves. Learning patterns from material does not establish that the material is true or fair.
- RLHF
- Reinforcement learning from human feedback: people rate outputs, and those ratings help shape later behavior. The ratings may not capture every value or situation.
- Data poisoning
- Deliberately changing training material to push a system toward unwanted behavior.
Evidence and its limits
Best-of-Venom tests injected preference pairs and resulting sentiment manipulation in its experimental setup.
Baumgärtner and colleagues · Research preprint
Best-of-Venom: Attacking RLHF by Injecting Poisoned Preference Data ↗ (opens in a new tab)Experiments manipulate preference datasets and model sentiment. Scope is the experimental setup, not an estimate of commercial prevalence.
Broader frameworks and background
These resources provide context. They do not establish every claim in this concern.
A question to discuss
How are feedback contributors, suspicious rating patterns, and disputed preferences reviewed?
Keep track of different views, what needs checking, and any next step in your own discussion notes.
Open discussion notes →Principles for the conversation
Before you finish · reflect together
Virtues and principles to consider
Virtues are qualities we practice in how we treat people and make choices. They are not scores for judging someone. Use these ideas to explore what a thoughtful response could look like.
Truthfulness
Truthfulness means making claims that match the evidence and being clear about what we do not know. It includes correcting mistakes rather than protecting a polished story.
Check sources, label guesses, and distinguish a trial result from a proven benefit. If AI helped produce a claim, someone still needs to take responsibility for checking it.
Ask together: What can we honestly say, and what still needs checking?
Justice
Justice means fair treatment, fair procedures, and attention to who receives benefits and burdens. It asks whether people with less power can be heard and seek repair.
Look at outcomes across different groups and individual cases. Provide a usable way to challenge decisions and correct harm. Equal wording in a policy does not guarantee equal treatment in practice.
Ask together: Who benefits, who bears the cost, and who can challenge the decision?
Stewardship
Stewardship means caring for resources and systems that others depend on, including people who will use them later. It looks beyond the benefit of a quick launch.
Plan for maintenance, energy use, data protection, and eventual shutdown. Ask who will look after the tool when the original team moves on and how future users can avoid inheriting hidden costs.
Ask together: What are we responsible for caring for over time?
Editorial ratings
These are editorial judgments from the register, not measured probabilities or predictions about your situation.
About the register and scoring →More ways to explore
Search for more writing
For more specific or recent material, open a search below or ask for a web summary. AI-generated summaries can contain mistakes; open the original articles to check them.