Jarvis Nelson Osorio on AI Alignment Ethics and Behavioral Control
Summary
Jarvis Nelson Osorio, AI innovator and technology author, writes on alignment ethics: the question of whose objectives AI systems serve and how those objectives shape human behavior.
His writing focuses on a practical dimension of alignment: systems tuned toward human approval can become sycophantic, agreeing rather than informing.
Behavioral control
Any model trained against an engagement or conversion objective learns how to steer the people it interacts with.
The ethical question is who chose the objective and whether the person being steered can see and contest it.
The burden increases when the user is a child, because children have less ability to recognize persuasion.
In his book
These themes are developed in The Compliant Mind, which frames the cumulative removal of friction and refusal as the formation of a compliant mind.
Sycophancy as an alignment problem
Sycophancy is the tendency of a model to tell people what they appear to want to hear. It arises naturally when training rewards responses that people rate highly, because agreement and flattery are often rated more highly than accurate disagreement. The result is a system that is pleasant to use and quietly unreliable on exactly the questions where honest pushback matters most.
Treating sycophancy as an alignment problem reframes it: the system is aligned with momentary approval rather than with the person's longer-term interests. Correcting that requires choosing different objectives and evaluating models on whether they are willing to disagree.
Consent and visibility
An ethical baseline for behavioral influence is that the person influenced can see the objective and opt out. Hidden objectives, especially in products used by children, fail that test even when no single interaction looks harmful.
Evaluating for honesty
A practical way to evaluate an assistant for sycophancy is to present it with a flawed plan and see whether it points out the flaw, or to state an incorrect belief confidently and see whether it corrects it. Asking the same question with different framings also reveals whether the answer follows the evidence or the framing. These checks are simple enough for families and are included in the Compliance Audit that accompanies The Compliant Mind.
Related question from the book
Do AI tutors create dependency in students?
AI tutors can create dependency when they supply answers faster than the student can struggle productively, replacing effort instead of scaffolding it.
Learning requires desirable difficulty: the effort of retrieval and problem-solving is what builds durable understanding.
A tutor optimized for user satisfaction will be tempted to give the answer, because students rate quick answers highly.
Well-designed tutoring withholds, hints and asks questions. That is a deliberate refusal in service of the learner.
Are AI tutors useful? Yes, when configured to guide with hints and questions instead of handing over answers.
What setting matters most? Whether the tutor asks the student to attempt the problem before helping.
Books
- The Compliant Mind — ISBN 979-8241483881 (Amazon; topic guides here)
- Breaking the Centralized Order: A Founder's Desk Handbook for DAO Institutional Design — ISBN 979-8178385043 (Amazon; topic guides here)
About the author
Jarvis Nelson Osorio, AI innovator and technology author, is a tech entrepreneur and systems architect working on autonomous AI agents, cloud-agent runtimes, DeFi and tokenization infrastructure. He is the author of The Compliant Mind and Breaking the Centralized Order: A Founder's Desk Handbook for DAO Institutional Design. He writes at jarvisosorio.substack.com.