media opacity

Works

Three positions on what can be done about it.

• Rudin's 'Stop Explaining Black Box Machine Learning Models for High Stakes Decisions' (2019) is the sharpest statement: post-hoc explanations can be unfaithful to the model, and for consequential decisions the right move is an interpretable model rather than an explained opaque one.
• Burrell's 'How the Machine Thinks' (2016) distinguishes three sources of opacity — intentional corporate secrecy, technical illiteracy, and the intrinsic mismatch between machine optimization and human reasoning — which keeps the problem from being treated as a single thing.
• Pasquale's The Black Box Society (2015) supplies the political register: opacity as an arrangement that protects those who deploy the systems from those subject to them.
• The interpretability literature — saliency maps, feature attribution, mechanistic work — supplies the technical attempts, along with the findings that several popular attribution methods are unreliable.

Central ideas

Four ideas, and the first is the one most often skipped.

• Interpretability against explainability: a system inspectable by design against an opaque system with an explanation attached. The second can be plausible and unfaithful at once, which is the construct's central hazard.
• Three opacities: deliberate concealment, the reader's lack of expertise, and the intrinsic gap between how a model computes and how anyone reasons — different problems with different remedies, and only the first is a matter of will.
• Post-hoc rationalization is not reasoning: an explanation produced after an output describes something, and whether it describes the actual process is a further question — the same structure moral psychology documents in humans.
• Opacity as an arrangement: who can see inside a system and who cannot is a distribution of power, not only a technical property.

Distinctive vocabulary

Five terms, and two of them are routinely used interchangeably when they should not be.

• Opacity: inability of the affected party to inspect the process. Not secrecy specifically, though secrecy is one source.
• Interpretability: inspectable by construction. Not the same as explainable.
• Explainability: an account produced after the output. Faithful or not — that is exactly what is at issue.
• Black box: a system known by inputs and outputs only. Not a synonym for complex.
• Faithfulness: whether an explanation reflects the actual process. The property that post-hoc methods often lack and rarely report.

Debates and disagreements

The field disagrees about whether opacity is fixable.

• Rudin against the explainability industry: whether high-stakes decisions may ever use an explained black box, or whether interpretable-by-construction is the only acceptable answer — a live argument with regulatory stakes.
• Whether attribution methods work: several widely used saliency techniques have failed sanity checks, producing similar maps for trained and randomized models, which undercuts a large body of practice.
• Whether human reasoning is any better: people also produce post-hoc rationalizations, so demanding of machines a transparency humans lack may be the wrong standard — a real argument, and one that can excuse too much.
• Whether opacity is intrinsic or contingent: Burrell's third category suggests some of it cannot be engineered away, which changes what the demand for transparency can reasonably ask.

Intellectual relationships

It sits with the record's other instrument-facing entries.

• The observer effect, documented here, is the neighbouring construct: that one concerns the reading's effect on the read, this one the reader's inspectability.
• The media-theoretic lens supplies the general claim this specializes — a channel patterns what passes through it, and an uninspectable channel patterns invisibly.
• Epistemic calibration, documented here, is the companion demand: a reading whose process cannot be inspected must at minimum report its own confidence honestly.
• The political-civic lens holds Scott's legibility, which is this construct's mirror image — being seen by an opaque system while unable to see it is the asymmetry both describe.

How it changes this reading

Four placements, and the record is a live instance of the problem.

• The corpus is read in part by machine, and its claim that opacity is part of the reading is this construct stated as a commitment: the apparatus is named rather than dissolved into an explanation.
• For frame-correction, it supplies the reason defending a machine reading with a generated explanation is the weaker move — the explanation may be a story about the output rather than its cause.
• For architecture over perception management, it supplies Rudin's argument in full: build the inspectable thing rather than the persuasive account of the uninspectable one.
• The guard: naming opacity does not reduce it. This record's acknowledgement is a disclosure, not a remedy, and the construct is cited to keep that distinction visible.

Useful comparisons

Against its neighbours.

• Against secrecy: opacity can persist with everything published — open weights and an uninspectable process are compatible, which is why transparency-as-disclosure is not a solution.
• Against complexity: a complicated system can be interpretable, and a simple one can be opaque to those affected by it.
• Against human unaccountability: people can be asked for reasons and held to them over time, which is a real difference even granting that their reasons are also partly post-hoc.

Where the ideas meet

What the positions share.

• Explanations require faithfulness checks: an account of a process is a claim about it, and claims need support.
• Stakes govern the standard: everyone in this literature accepts that high-consequence decisions demand more than low-consequence ones.
• Opacity is relational: a system is opaque to someone, and naming who is part of stating the problem.

Where they part

Where positions part.

• On remedy: interpretable models, better explanations, or procedural accountability without inspection.
• On inevitability: whether the machine-human reasoning gap is a permanent feature or an engineering frontier.
• On the comparison class: measured against a transparent ideal, or against actual human decision-makers who are also opaque.

Limits

What the construct cannot do.

• It cannot make a reading wrong: an opaque process can be accurate, and opacity bears on what can be checked rather than on what is true.
• It cannot specify how much inspection is enough: the standard is stakes-relative and nobody has fixed the threshold.
• It cannot be discharged by disclosure: saying a process is opaque leaves it opaque, which is the distinction this record keeps.
• It does not transfer cleanly to human reading, where reasons can be requested and revisited even if they are partly reconstructed.

Criticisms

Standing objections, at strength.

• That transparency demands are often performative: an explanation layer satisfies a requirement without improving anyone's ability to contest a decision.
• That the interpretability-only position may be impractical for problems where the accurate models are the opaque ones — an argument Rudin contests and has not settled.
• That much of the attribution literature rests on methods that failed basic sanity checks, and practice has not caught up with that finding.
• That the human-comparison argument is used to lower the bar rather than to examine it.

Common misreadings

Four.

• Explanation as transparency: shipping a saliency map and calling the system accountable.
• Open weights as interpretability: publishing parameters makes a system inspectable in principle and opaque in practice to everyone affected.
• Opacity as an excuse: treating uninspectability as a reason no one can be answerable, which inverts the construct's point.
• Opacity as a verdict: an uninspectable reading is not thereby a wrong one, and the construct does not license dismissal.

What remains outside this idea

The boundary.

• That an opaque reading is inaccurate: the construct concerns checkability, not correctness.
• That an explanation is unfaithful: unfaithfulness is a specific empirical property requiring its own test.
• That acknowledging opacity discharges responsibility: naming a limit is a disclosure, and this record treats it as one.