Manifestos ask for adherence. Questions ask for answers, case by case. Here are the five questions I use as a field test for algorithmic accountability, and the reason the last one cannot be faked.
It is tempting to end any argument about technology and power with a manifesto. Ten principles, seven rights, twelve recommendations: tidy numbers give the impression that the world has finally accepted the structure of the slide deck.
I prefer questions, because they do not ask for adherence, they ask for answers, and they can be addressed to a concrete thing: a technology, an institution, a service, a decision.
Call it algorithmic accountability if you need a label for the procurement document; I think of it as a field test. These are the five questions I propose, and they apply equally to a citizen in front of a platform state, an organization in front of a supplier, and a society in front of its infrastructure.
The five questions
1. Can we understand it?
Not every technical detail; enough to know what we depend on, which data is used, which objective is pursued, and which limits exist.
Transparency does not coincide with publishing thousands of incomprehensible pages, and a system can be formally open and substantially opaque.
The operational test: if nobody can explain why a decision was made, who is actually governing it?
2. Can we contest it?
Every system that produces consequences should provide for the possibility of error: showing that the real case does not match the representation, correcting a record, obtaining a recognizable responsibility.
A power without appeal turns probability into a verdict.
The quality of a system is not measured only by how often it is right, but by how it treats those who have suffered its error.
3. Can we leave it?
Can we move the memory, keep operating, choose another solution, or does the infrastructure grant freedom only to those who stay inside?
The possibility of leaving disciplines the present relationship: a supplier who knows it can be replaced behaves differently from one that considers us technically imprisoned.
Real competition is not the number of logos on the market; it is the concrete possibility of changing.
4. Can we forget and correct?
A healthy memory does not keep everything the same way.
Accountability requires some records to survive; freedom requires others to expire.
The test has two halves: who decides what must be forgotten, and who verifies that the promised forgetting has actually happened, across copies, indexes, backups, and derived models.
5. Can we still surprise it?
A society that knows us through the past must leave room for the fact that we can change.
A person totally reduced to a prediction is no longer treated as a subject: they are a trajectory.
The right to surprise the system is the right not to coincide, forever, with one’s own data.
Algorithmic accountability is a ladder, not a checklist
Put in sequence, the five are not a flat list but a ladder that climbs through different planes.
Understanding is epistemic. Contesting is political. Leaving is economic.
Forgetting and correcting is temporal. Surprising is almost ontological: it concerns what we are, for the system.
And the first four already have regulatory anchors, imperfect but real: transparency duties in the GDPR and the AI Act.
The contestation of automated decisions in article 22, portability and the exit strategies now imposed on critical services by the European rules on digital resilience, rectification and erasure as written rights.
That is the ladder algorithmic accountability actually has to climb.
How these anchors are landing in national law is a story of its own; I looked at the Italian case in my analysis of Draft Decree No. 418.
The fifth question cannot be faked
The fifth has no anchor, and it cannot have one, because unpredictability is not proceduralizable.
No regulator can audit a right to be surprising.
This is exactly what makes it the most important of the five: the first four can be satisfied formally until they are empty, the compliance theatre every practitioner recognizes, while the fifth admits no theatre.
It is the part of algorithmic accountability that cannot be delegated to an auditor.
Either the system leaves room for those who do not resemble its past, or it does not.
A system that passes residency clauses, transparency reports, and audit checklists but fails the fifth question is compliant, and captured.
Running the algorithmic accountability test
How do you use this in practice? In a vendor review or a procurement round, ask the five questions in order and demand artifacts, not adjectives.
An explanation a non-specialist can follow, a contestation path with a named owner, a tested exit plan with formats and timelines, a verified deletion procedure, and a demonstrable margin for cases that do not match the past.
Algorithmic accountability is cheap to claim and expensive to show; the questions exist to move the conversation from the first to the second.
Treat every unanswered one as a finding, not a footnote.
The question that is missing
There is an objection I take seriously: all five questions are defensive.
They measure freedom as the capacity to resist a system, to understand it, contest it, exit it, escape its memory.
The affirmative question, can we build it, can we co-write it, is not on the list.
I think it exists, and it lives elsewhere in the same argument: in imagination as the power to open futures rather than defend against them.
Whether that is a deliberate choice or the sixth question waiting for a second edition, I leave open, which seems appropriate for a framework whose whole point is to end with better questions rather than cleaner answers.
Where this comes from
The five questions close chapter 13 of Who Owns Memory Owns the Future, a narrative essay on how the infrastructures of memory, from fire and clay tablets to databases and machines that remember, have always shaped power and freedom.
The book develops each question against four thousand years of precedents, from the keys of the granary to the permissions of an AI agent, treating algorithmic accountability as a very old problem wearing new clothes.
