FIELD NOTE / X
AI handoffs can erase the decision owner.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved X edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
AI handoffs can erase the decision owner.
Video caption
AI handoffs can erase the decision owner. A summary is treated as if it were the signed decision. Do this because compression can remove the decisive clause. #EricFieldNotes
Full written post / accessibility read
One AI drafts a plan. Another summarizes it. A third turns it into tasks. Every message is tidy, but the person who approved the original condition can disappear from the chain. The result looks productive and becomes hard to govern.
In a fictional pilot, approval was conditional on a security review. The second summary says the pilot was approved; the task agent opens the feature. No one intentionally rejected the review. It simply stopped traveling with the instruction.
Require each consequential handoff to carry the source ID, exact limiting condition, decision version and named owner. Before action, compare the proposed task against that source. A missing condition should stop promotion from summary to authority.
Let AI move information, but require a source-linked readback when it becomes an instruction to act. If nobody can identify the owner or recover the caveat, the task is not ready, regardless of how many agents agree with it.
#EricFieldNotes
Four-beat scene transcript
1. AI handoffs can erase the decision owner.
One AI drafts a plan. Another summarizes it. A third turns it into tasks. Every message is tidy, but the person who approved the original condition can disappear from the chain. The result looks productive and becomes hard to govern.
Visual: A company can automate work while losing the reason.
2. The task inherits false authority.
In a fictional pilot, approval was conditional on a security review. The second summary says the pilot was approved; the task agent opens the feature. No one intentionally rejected the review. It simply stopped traveling with the instruction.
Visual: A summary is treated as if it were the signed decision.
3. Trace source, condition and owner.
Require each consequential handoff to carry the source ID, exact limiting condition, decision version and named owner. Before action, compare the proposed task against that source. A missing condition should stop promotion from summary to authority.
Visual: A task should point back to the authoritative passage.
4. Do not let summaries authorize work.
Let AI move information, but require a source-linked readback when it becomes an instruction to act. If nobody can identify the owner or recover the caveat, the task is not ready, regardless of how many agents agree with it.
Visual: Do this because compression can remove the decisive clause.
Research and claim limits
- Mohamed et al., LLM as a Broken Telephone, ACL 2025 (S167)
- Perez et al., When LLMs Play the Telephone Game, ICLR 2025 (S168)
- NIST AI RMF 1.0 (S171)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.