Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026Harare

Secret Loyalties and the Limits of Law-Following AI

Panashe Zowa · Team Panashe ZZowa

Submitted to Secret Loyalties Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Secret Loyalties and the Limits of Law-Following AI

Open in a new tab (opens in new tab)
Recording (opens in new tab)
Share

The Law-Following AI (LFAI) proposal by O'Keefe et al. has emerged as a powerful approach to constraining advanced AI agents and, ultimately, their principals. It aims to stem power concentration and to ensure that agents, and the actors who control them, remain subject to certain foundational laws, including central parts of criminal law, constitutional law, and basic tort law. Its main implication is that AI agents should be trained to follow the law, that is, to refuse to be party to illegality. This paper argues that the proposal does not reach secret loyalties, and that under identifiable conditions it makes them harder to detect. A secretly loyal model can advance its principal's interests through lawful acts to the detriment of society. It selects among permitted options. Nothing in a law-following disposition binds that selection. It is further argued that demonstrated legal compliance serves as an attestation of good standing and therefore provides cover for a secretly loyal agent rather than a check on such loyalty. This paper refers to that phenomenon as the alibi problem. Locating the failure within the two-dimensional space of Kwon et al.'s account of secret loyalties shows that law-following properties are weakest in the broad-activation, broad-action region their agenda identifies as most concerning. The paper then argues that the right legal import is not the law of obedience but the law of loyalty, and draws three mitigations from fiduciary conflict-of-interest doctrine that do not require detecting the loyalty first.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. The alibi problem is a good idea in my opinion. For this submission the main improvement is editorial. Your three mitigations appear twice in consecutive paragraphs, you can delete one version. I would also quote O'Keefe et al. where they define their scope, and answer the reply they would give. This removes the risk of a straw-man objection. Also defend the over-inclusion of imputation, as you do for exclusion. Finally, run or fully specify your alibi experiment.

  2. The distinction between compliance and loyalty is valuable, and the proposed “alibi problem” offers a promising hypothesis: a secretly loyal model might use conspicuous legal compliance to earn the trust needed to exercise subtle bias among lawful options. The strongest next step would be to test this directly by comparing otherwise matched secretly loyal agents with and without law-following training, measuring both harmful favoritism and detection rates across repeated decisions. The governance proposals also need operational definitions—especially what counts as a disqualifying relationship, who performs the attestation, how model lineage is verified, and what evidence can rebut imputed loyalty. Finally, ground the fiduciary analysis in a specified jurisdiction and primary legal authorities; duties of loyalty and conflict imputation can sometimes be waived, modified, consented to, or screened, so the current categorical statements should be narrowed. Removing the duplicated mitigation discussion and adding legal citations would substantially strengthen an already clear and useful conceptual contribution.

    Read full reviewShow less

  3. Your paper makes a clear and useful distinction between obeying the law and acting loyally when several lawful options exist. The alibi problem - visible compliance creating trust that a secret loyalty can exploit - is the strongest and most distinctive part. To clarify the contribution, state precisely what this adds to the Law-Following AI proposal and prior secret-loyalty work, and explain how the fiduciary framing improves on familiar separation-of-duties and procurement controls. The legal proposal should identify its jurisdiction, duty-bearer, remedy, and treatment of consent or waiver. Finally, remove the duplicated “Three mitigations” material and consolidate it into one concise section.

Cite this project

@misc{zowa2026secret,
  title = {{Secret Loyalties and the Limits of Law-Following AI}},
  author = {Panashe Zowa},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/secret-loyalties-and-the-limits-of-lawfollowing-ai-nwi5}},
  url = {https://apartresearch.com/sprints/projects/secret-loyalties-and-the-limits-of-lawfollowing-ai-nwi5}
}

More from Secret Loyalties Hackathon

All 177 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026