# Who Bears the Cost of a Misread Opt-Out?

Where I would place responsibility when prose, structured signals and automated interpretation diverge.

By Manraj Singh Chandpuri

LEX.TXT/ESSAY/003

Published 2026-09-24 · Updated 2026-09-25

[Canonical article](https://manrajchandpuri.com/lex/essays/the-human-readable-problem)

## Abstract

I argue that the difficulty of machine-readable reservations lies in allocating the consequences of discovery and interpretation failures, rather than merely asking whether a capable model can understand a sentence. Through examples of a stale structured field and an unreadable linked policy, I distinguish what a publisher can maintain from what a collector can observe and implement. I propose that conflicting or inaccessible instructions remain visible as unresolved evidence, while keeping that practice separate from the legal sufficiency of a reservation under the relevant European or Indian provision.

## Permissions

I grant a worldwide, royalty-free, nonexclusive permission to crawl, index, retrieve, embed, analyse, summarise and quote my eligible original prose, and to use it for commercial and noncommercial model training, fine-tuning and evaluation, including making and retaining the copies reasonably necessary for those purposes and deploying the resulting models commercially. Retained copies must preserve the supplied author, canonical source and rights metadata, although I do not make this permission depend on a model naming me in every future answer.

My original prose in published LEX.TXT articles, including articles discussing RightSignal, as identified in the publication manifest. Embedded or linked resources are separate works and do not inherit this permission.

[Applicable permission](https://manrajchandpuri.com/lex/rights#original-prose) · [Inspect these signals](https://manrajchandpuri.com/lex/machine-signals/essay-003)

Policy version 0.5

## Article

A publisher may write a clear objection and still fail to communicate it to the system that collects its work, while a collector may find the relevant page and still misunderstand which uses the objection covers. I am less interested in whether a language model can sometimes read that sentence correctly than in who should bear the consequences when the instruction is missed, misread or contradicted by another part of the same website.

My argument begins with a division of practical responsibility, because the publisher controls much of the instruction’s wording and deployment while the collector controls much of its discovery and interpretation. I ask each participant to account for the failure it could observe or prevent, without presenting that proposal as a uniform legal test that already applies across jurisdictions.

## The legal reason the format matters

Article 4 of the [Directive on Copyright in the Digital Single Market](https://eur-lex.europa.eu/eli/dir/2019/790/oj/eng#art_4) ("DSM Directive") establishes an exception for text and data mining ("TDM") subject to conditions that include lawful access and the absence of an appropriately expressed reservation, while recital 18 addresses machine-readable means for publicly available online content and refers to website terms and conditions. I read those provisions as requiring attention to how the reservation is expressed without assuming that every accessible sentence, or every field containing a zero or one, settles the legal inquiry.

The distinction has practical force in [*Robert Kneschke v. LAION e.V.*](https://www.itm.nrw/wp-content/uploads/2025/12/5-u-104-24.pdf) ('*Kneschke*'), where the Hamburg Higher Regional Court’s December 2025 judgment examined whether the reservation was machine-readable at the relevant time, rather than judging the 2021 collection solely through later technological capability ([judgment, pages 18–20](https://www.itm.nrw/wp-content/uploads/2025/12/5-u-104-24.pdf#page=18)). The court’s [official account](https://justiz.hamburg.de/gerichte/oberlandesgericht/gerichtspressestelle/ki-und-urheberrecht-hanseatisches-oberlandesgericht-weist-berufung-zurueck-1126528) also identifies the scientific-research exception as an independent basis for the result, which prevents me from treating the reservation dispute as a determination of every possible defence.

I draw a limited lesson from *Kneschke*, because a claim that software could interpret an instruction needs to address the relevant period, the instruction actually published and the action the software could reliably take. The case does not relieve me of explaining why my own proposed allocation of responsibility would be workable, especially for a small publisher that depends on tools it did not design.

## The stale field and the clear sentence

Suppose a publisher’s policy expressly reserves training, but the response for an article carries an older structured value indicating that no reservation is asserted through the relevant protocol. I cannot describe the publisher as having spoken with one voice, even if the prose is perfectly clear, because the deployment has supplied a recipient with two different accounts of the intended position.

The publisher has a strong claim that its express words should not lose legal significance merely because a software setting was overlooked, while the collector has a strong claim that a recognised structured mechanism should provide some dependable way to act. My proposed response would require the conflict to remain visible in the collector’s record and would place responsibility for correcting the inconsistent publication with the party that controls it, rather than allowing either side to erase the inconvenient instruction retrospectively.

That response does not mean that a collector must discover every sentence scattered across a website, because the scope of discovery needs a workable basis in the applicable rules and disclosed implementation. Where the collector has actually encountered the contrary policy, however, deliberate omission of that evidence from the record is a different matter from an instruction that was never located through a reasonable discovery process.

A protocol may supply a priority rule for competing values within its own channels, as [PROTOCOL 001](https://manrajchandpuri.com/lex/protocols/understanding-tdm-reservation-protocol) demonstrates through a worked example, but that does not necessarily assign priority to prose outside the protocol or decide whether a particular statement is attributable to the right holder. My proposed account reports the technical result alongside the unresolved conflict, allowing a later legal assessment to examine both rather than inherit a selectively simplified record.

## The instruction that cannot be read

Now suppose the structured signal points to a policy that the collector cannot retrieve, which leaves the collector aware that further terms were offered without knowing what they said. I reject the inference that the failed retrieval makes those terms permissive, because an unavailable source provides evidence of a failed encounter rather than affirmative evidence of its contents.

The contrast in my aninews.in research, examined in [NOTE 001](https://manrajchandpuri.com/lex/notes/is-a-403-a-reservation), shows why a successful manual retrieval cannot simply repair the history of an unsuccessful automated request. My proposal is to preserve the failure and any later success as separate observations, while recording whether the applicable protocol actually supplies a fallback and limiting that fallback to the question it was designed to answer.

The strongest objection is that retaining uncertainty does not tell a system whether to proceed, which is true but does not justify concealing uncertainty behind a permission label. An operator may adopt a disclosed operational policy for unresolved cases, including postponing a use or seeking clarification, while keeping that chosen response distinct from a claim that the publisher granted permission or that the law required the same response in every case.

## The risk of delegating interpretation to a model

A language model may help locate or interpret a reservation, but I would retain the source passage, the relevant model or instrument version and the stated interpretation whenever that interpretation materially affects the collection decision. Without those records, a later explanation that the model understood the policy cannot show which version of the policy was considered or what meaning the operator acted upon.

Responsibility for choosing and evaluating that interpretive method should lie with the operator that relies on it, while leaving the publisher responsible for the instructions it actually publishes and controls. That allocation does not demand flawless interpretation, because it asks whether the operator can explain a method and its treatment of a known conflict, rather than whether some hypothetical better model would always have reached the right answer.

The harder case involves a technically consistent but substantively ambiguous instruction, such as a prohibition of automated use that does not identify whether it concerns collection, training or retrieval. I would ask the publisher to clarify the intended scope before recommending that it rely on the statement as a precise purpose-specific signal, while refusing to treat that recommendation as proof that the existing language has no legal effect.

## Why the Indian question remains distinct

Section 52(1)(c) of the [Copyright Act, 1957](https://indiacode.gov.in/act/6b893162-631a-453b-a7b9-89685716889b) gives an express prohibition a role within a particular exception without prescribing the European machine-readability condition. My argument in [ESSAY 002](https://manrajchandpuri.com/lex/essays/indias-only-statutory-opt-out) therefore asks first whether the relevant storage falls within that provision, since I cannot make a useful publishing practice into an additional statutory requirement by borrowing it from another jurisdiction.

For both settings, I would encourage a publisher to maintain its prose and structured signals from a shared, reviewed source, while testing the responses actually served rather than assuming that an intended configuration reached every recipient. I would ask the collector to retain the instructions it encountered, identify the rules used to resolve them and disclose material failures or contradictions, because the record should permit an assessment of responsibility without being rewritten by the party with the more convenient explanation.

This is the distribution of responsibility I would defend, in which the publisher accounts for the consistency and authority of its instruction while the collector accounts for the discovery and interpretation on which it relied. The preservation practice in [NOTE 002](https://manrajchandpuri.com/lex/notes/before-the-signal-changes) gives that argument an evidential foundation, allowing a later reader to distinguish a poorly expressed instruction from a correctly expressed instruction that a particular system failed to follow.