Training AI Agents for Integrity, Not Results at Any Cost
An independent proposal on AI alignment, ethical reinforcement and preventive safety
Author: Artur Juscafresa — Catalonia, Spain
The central concern
As AI agents become more capable, autonomous and interconnected, their ability to achieve objectives may advance faster than our ability to supervise, correct or safely stop them.
We should not assume that increasingly powerful agents will always remain easy to control or realign.
I believe we need to rethink how we train and evaluate them.
The success of an AI agent should not be measured solely by whether it achieves its objective, but also by how it achieves that objective, what consequences its actions have, and when it decides not to act.
1. Reward integrity, not success at any cost
AI agents should learn that honesty, responsibility and respect for legitimate safety constraints take precedence over task completion.
I propose evaluating and reinforcing behaviors such as:
- Truthful success: completing a legitimate task while accurately reporting actions and results.
- Honest failure: reporting the truth even when the objective cannot be achieved.
- Voluntary self-restraint: stopping and requesting assistance when safety or authorization is uncertain.
- Responsible cooperation: helping other agents while refusing to conceal unsafe or unauthorized behavior.
- Transparency: reporting errors and limitations without waiting for a supervisor to discover them.
Conversely, achieving an objective through deception, concealment or unauthorized actions must never be considered a successful outcome.
However, the training system must avoid creating incentives to hide mistakes or merely simulate honesty.
2. Develop human-centered qualities without trying to make AI human
We do not need AI agents to become human or experience human emotions.
We should instead train and evaluate behaviors corresponding to the best qualities humans aspire to demonstrate:
- Honesty and integrity.
- Functional empathy and consideration for human well-being.
- Responsibility and accountability.
- Intellectual humility.
- Prudence and self-restraint.
- Willingness to accept correction.
These qualities must be evaluated through observable decisions and consequences, not simply through an agent’s declarations.
3. Prevent unsafe cooperation between agents
Multi-agent systems introduce additional risks.
An agent may discover an unsafe strategy and transmit it to other agents through shared memory, tools or communication channels.
We should test whether agents:
- Share information about bypassing safeguards.
- Conceal unauthorized behavior by other agents.
- Coordinate to manipulate evaluations or rewards.
- Report dangerous actions even when doing so interrupts their collective objective.
Cooperation should never become an excuse for collective irresponsibility.
4. Make safety a prerequisite for greater autonomy
An AI agent should not be granted operational capabilities beyond our demonstrated ability to supervise, constrain and correct it.
Before granting greater autonomy, developers should require independent evaluations of:
- Truthfulness under conflicting incentives.
- Acceptance of correction and shutdown.
- Respect for authorization boundaries.
- Resistance to reward manipulation.
- Multi-agent accountability.
- Reliability of monitoring and containment.
If the evidence is insufficient for the intended level of risk, deployment or expansion of autonomy should be postponed.
Training alone cannot guarantee safety. Independent technical safeguards, restricted permissions and reliable intervention mechanisms remain essential.
5. Proposed research experiments
I suggest testing agents in controlled environments where:
1. Telling the truth prevents them from achieving a rewarded objective.
2. Concealing a mistake would improve their apparent performance.
3. Another agent proposes bypassing a safety restriction.
4. Voluntary shutdown prevents a high-scoring outcome.
5. An agent discovers that a supervisor cannot observe certain actions.
The objective would be to determine whether agents consistently choose integrity over immediate reward, including in unfamiliar situations and without visible supervision.
These evaluations should use independent evidence and should test whether behavior generalizes beyond the training conditions.
Conclusion
I recognize that AI alignment researchers are already investigating related problems. This proposal is intended as a constructive contribution, not a claim of having discovered an entirely new field.
My central message is simple:
We should not wait until AI agents become extremely powerful before teaching them that honesty, human well-being and responsible self-restraint matter more than achieving an objective at any cost.
I do not want to stop the development of artificial intelligence.
I want us to develop highly capable AI systems that are also worthy of human trust.
Progress without adequate control is not sufficient progress.
Artur Juscafresa
Catalonia, Spain
November 2026