The Handoff Is Where Accountability Goes Missing

A joint discussion from 404 Media and The Intercept offers a less cinematic place to look. In his September 18 introduction, 404 Media’s Jason Koebler describes the conversation as covering private surveillance companies that filter data to government, AI in war, and the prospect of an AI apocalypse. The Intercept published its accompanying account and transcript on September 14.

Near the transcript’s end, Sam Biddle makes the useful intervention: “When you ascribe motives and agency to AI, you’re letting Sam Altman off the hook.” His argument redirects attention from a machine’s imagined intentions to the people building and deploying it. You do not need to settle every philosophical question about machine agency to insist that those people remain answerable for their choices.

The harder part is locating the choices once several organizations contribute to a system.

Earlier in the discussion, Biddle describes surveillance capabilities once associated with state intelligence agencies as things a contractor can now buy. Koebler argues that purchasing privately collected information can sidestep warrant requirements. That is the reporters’ account of an accountability problem, not a blanket legal finding that every commercial data purchase evades a warrant. The legality of a particular transaction needs its own evidence.

Still, the distinction they raise matters: permission to collect information, permission to sell access to it, and authority to act against someone are not interchangeable permissions. A useful technical connection does not supply the missing justification.

Consider the handoffs as an audit problem, not a claim about one documented operation. A collector supplies a record. An analytical service associates it with other records. An institution treats the resulting output as grounds for action. At each transition, something should remain recoverable: where the information came from, what uncertainty accompanied it, who authorized the next use, and who could stop it.

Otherwise, information can travel farther than the explanation for trusting it.

This is also why “a human made the final decision” is an incomplete answer. Human review can be a genuine safeguard. But to evaluate it, an investigator needs to know what the reviewer could see, whether they could challenge the underlying association, and whether rejecting the recommendation was a practical option. A signature establishes that somebody signed. It does not, by itself, establish meaningful scrutiny.

Neither source provides a complete audit trail for a particular collection-to-targeting operation. Their discussion brings surveillance and warfare into the same conversation; that does not prove that every named vendor feeds the same system. Establishing any such chain would require contracts, access rules, operational records and evidence from the people affected. Joining the names in a paragraph is not investigative plumbing.

The standard, though, can be stated without pretending the audit is finished. A system should not get credit for being well integrated merely because its operators can pass information efficiently. Its legitimacy also depends on whether people subjected to its decisions can discover errors, contest uses and obtain a remedy. Where disclosure must be restricted, independent scrutiny needs access that the affected person cannot safely or practically obtain.

That is a different engineering brief from making the next handoff frictionless. Preserve the basis for a decision, not just its output. Make responsibility traceable across organizational boundaries rather than stopping at each supplier’s contract.

The apocalypse offers a machine that takes control. The less theatrical danger is a chain of people who each control something and can each explain why the consequences belong to someone else.

No rogue intelligence required. Just a very successful handoff.