Automation can turn a true observation into a false decision through several individually inspectable steps. A shared mailing address is treated as one household, then as duplicate receipt, then as fraud, then as a reason to refuse help. The address match may be correct while every later inference is wrong. This chapter examines how delegated classification acquires practical authority and how correction must follow the resulting decisions.
A deployed service moves evidence to action through forms, databases, models, interfaces, staff permissions, rules and payments. Each transition can change meaning. No component need contain the whole mistake for the assembled process to impose it. The central question comes before intelligence: what happens along the path from sign to consequence, and who can correct it? Delegating memory and comparison extends human powers across distance; the people governing those extensions remain responsible for labels, thresholds, appeals and payments.
A learned model can also identify false links or missing evidence. Its contribution enters an authority structure: an institution adopts the rule, a vendor maintains data, and a reviewer may reverse the decision. Divided technical responsibility leaves the institution responsible for arranging the complete remedy.
Chapter 16 established how records shape action; Chapter 32 established the requirements of review. The digital addition is propagation: one inference can be reused in scores, cached summaries, training labels, and later decisions. Correction must identify which of those copies still exercises authority.
Prediction is not a finding
Training labels can record past referrals rather than independently established fraud. Agreement with them reveals the office's practice before it establishes anything about an applicant. Reusing unchallenged refusals as confirmed cases makes the mistake self-reinforcing.
Even sound labels leave a base-rate problem. Suppose 10 of 1,000 applications involve fraud. A classifier detects nine and flags 99 of the 990 legitimate applications: 90 percent sensitivity and 90 percent specificity, but only nine of 108 flags, about 8.3 percent, identify fraud. A useful investigation prompt can be weak evidence of guilt.
A score of 0.8 is not an 80 percent probability without suitable calibration in the deployment population. Even a calibrated probability must be weighed against the costs of each action. Investigation, clarification and refusal impose different burdens. Assessing only flagged cases conceals misses; asking staff to validate the model while relying on its warning can reproduce the inference being tested. Independent assessment must sample appropriate unflagged cases and examine abandoned applications.
False witness includes a true fragment placed under a false heading. Ed/martyria, witness, sheqer/pseudos, falsehood, and mishqal, measure, concern what another is asked to believe and receive. Proverbs requires hearing before answering and scrutiny of the first persuasive account. Augustine distinguishes mistaken speech from intended deception without making careless error harmless or a useful lie innocent. Responsibility differs among an employee trusting an inadequate display, a manager ignoring contrary evidence and a supplier knowingly overstating it. Truthful framing preserves what the datum establishes and leaves contrary evidence a way to change the decision.
Security, privacy and warranted disclosure
Confidentiality, integrity and availability protect against different failures: stolen lists, altered destinations and inaccessible services. NIST CSF 2.0 joins assigned responsibility, knowledge of dependencies, safeguards, detection, response and recovery as continuing work. Authentication also needs a route for legitimate users who lose their device, with safeguards against impostors.
Privacy can fail through authorized processing. Linking records may construct an unnecessary intimate profile, and a true or false inference may expose someone. NIST's Privacy Framework therefore distinguishes processing risks from cybersecurity and compliance risks, connecting responsibility to granular controls and intelligible communication.
A click under dependence on essential assistance is thin consent. Consent to eligibility checks covers eligibility checks, not unrelated advertising or perpetual profiling; public availability covers its original setting, and each new combination needs its own justification; usefulness for training still answers to permission, creators' rights and fair terms.
Faithful confidence, sod, and good acts done in secret, kryptos, make public observability an inadequate criterion of trust. God's final disclosure, apokalypto, delegates no omniscience to a platform. Necessary investigative retention differs from public searchability; protecting a survivor differs from destroying evidence to protect a perpetrator. An accurate location notification can enable stalking. The strongest surveillance argument begins with real concealed danger, but must establish necessity, proportion and controls against unrelated use.
Attention and correction
Recommendation ranks scarce visibility; advertising adds a purchaser's objective; moderation removes, limits, labels or delays. A quotation documenting abuse can be misclassified as its endorsement. Identity verification can establish account authority without proving a speaker's truthfulness, while compulsory public legal names can expose witnesses. Easy enrollment, obscure cancellation and default visibility change the effort and information available for choice.
Reach, familiarity, confidence, expression, belief and action remain different outcomes. Removing Facebook reshares in a three-month 2020 election intervention reduced political/untrustworthy exposure and news knowledge without detectable changes in measured attitudes. That null result stands alongside evidence of amplification.
Correction must reach the places where error acts. Pausing amplification, restoring source context, replacing the misleading account, enabling acknowledgment, updating practice and retaining corrected status are distinct operations. A correction hidden from the original audience can fail; republishing a private history to prove correction can injure again. Rules themselves need challenge where companies or governments use protection to suppress inconvenient testimony.
Synthetic expression can translate or assist, but also fabricate evidence, impersonate and simulate commitments no person has made.
Consequential tools and changed work
A robot must connect perception to safe action within time, accounting for load, stopping distance and failure. Its operational autonomy works within a bounded task and a support system.
Tool access similarly turns proposals into consequential instructions. Permission to summarize does not authorize disclosing records or changing bank details. Instructions embedded in retrieved material have not acquired the user's authority. Controls at action boundaries must limit what a mistaken interpretation can accomplish. Direct and indirect prompt injection differ from poisoned training data.
If a payment acknowledgment is lost, retrying can duplicate an award. Preserving the identity of the authorized transaction and checking its state separates requested, accepted, completed and failed acts. Safe stopping likewise depends on the task: abrupt power loss may endanger a loaded cart; account suspension may require an assisted route to essentials.
Actual workplace gains show how integration can improve particular tasks. Brynjolfsson, Li, and Raymond studied a staggered introduction of an assistant among 5,172 customer-support workers. Issues resolved per hour rose by about 15 percent, with larger gains for less experienced and lower-skilled workers; the most skilled showed small speed gains and some quality losses. The system supplied suggestions that workers could edit or ignore. Benefit belonged to a specific arrangement of acquired knowledge, available assistance, and work. A fund should likewise test whether better information produces timely, correct assistance while preserving the competence needed to detect exceptions.
The comparison itself needs care as tools and work change. METR's randomized early-2025 study found 19 percent longer task times for sixteen experienced maintainers given then-current AI tools. Its February 2026 follow-up reported severe selection and timing problems: some developers and tasks were withheld from randomization because participants preferred AI, and concurrent agents complicated time accounting. The investigators consequently declined to treat the newer estimates as a reliable measure of the current effect. The durable lesson concerns measurement: include selection, unfinished work, review, correction, and recipient outcomes when judging the whole service.
ILO's 2025 exposure index evaluates tasks through worker assessment, expert review and AI-assisted scoring, grouped in four gradients. It estimates some generative-AI exposure in about one quarter of global employment, uneven across occupations, countries and sexes, with clerical work especially exposed. Transformation is more general than wholesale occupational disappearance; remaining tasks can still form fewer or worse jobs. The study estimates technological possibility, not completed displacement.
Connectivity, integration costs, staffing and chosen judgment boundaries influence adoption. Saved labor becomes a better job only through appropriate pay, authority and expectations. Transition requires funded learning and workers' voice. Annotation, maintenance, exception handling and distressing moderation remain work, sometimes transferred to contractors; hardware and energy remain material dependencies.
An appeal that changes the service
For the address-link case, review needs access to original evidence and authority to reverse refusal, with an assisted route and funded urgent deadline. The claimant need not expose another household's private circumstances. A provisional emergency payment can follow verified identity and need, counted openly toward any later award.
Once separate households and eligibility are established, the false identity link and fraud finding must be withdrawn, while the true address match remains. Corrected status must reach operative data, the award must be paid, and receipt confirmed. The refusal must also cease to serve as a confirmed-fraud training label.
Repair extends to cached summaries, retrieved copies and training labels wherever the old account continues acting. A changed authoritative record or decision rule may suffice even where the learned parameters are unchanged. Proportionate retention supports accountability without making a true past wrong an unlimited present verdict.
The fund reviews other refusals depending on the same link, pays owed awards and agreed redress costs, and funds staff correction. It suspends the classifier while retaining narrowly described retrieval and matching. Simpler competent decisions and review replace the defective route; added work must not create a new inaccessible backlog. Knowing neglect requires separate investigation, while owed correction need not await every culpability finding.
The fund compares assistance received, genuine fraud detection, timing, privacy, applicant burdens, staff work and costs, including abandoned applications. More complaints can reflect worse harm or a newly usable way to report it. Known harmful refusals require correction, not retention as a control. When competent alternatives give equal outcomes at less burden, the simpler method serves recipients better.
Images, marks and the person
An image, tselem/eikon, represents without creating a person's worth. Revelation's speaking image and charagma connect demanded worship with deception, persecution and conditioned commerce. Credentials can protect access; technical resemblance alone establishes no fulfillment. The analogy concerns power using signs and economic dependence to demand allegiance it cannot rightfully command. The Lamb's truthful witnesses provide the counterpart.
Ordinary trust in demonstrated competence is not worship. Technical lordship begins when efficiency or prediction becomes immune to judgment. Ephesians joins truthful speech, work and generosity to learning Christ and life in the personal Spirit: ganav/klepto, stealing, must cease through changed conduct, not a new label.
Digital communication sustains relationships through truthful attention and service. Archives and generated likenesses preserve memory; resemblance cannot establish the bearer's continuity. A record witnesses to a life; resurrection promises restoration of the person it represents.
Related research
Compare this chapter with source documents or study its biblical references.
Chapter comments