b19 — Fact-sheet 4 — Verification-bandwidth asymmetry and scientific authorship#

Compiled:: 2026m05d13
Compiled by:: Claude Opus 4.7 Max (subagent for the b19 AI co-authorship analysis)
Builds on:: b17 llog at hell/ll/other/b/17/other_ll_2026m04d11_singularity-info-crisis-llog.rst
Scope:: Where in scientific authorship verification-bandwidth gaps hit hardest; ICMJE-criterion sensitivity; repurposed coping strategies; alone-disqualifying threshold; cross-reference to deceased-author and consortium cases.
Methodology:: WebSearch / WebFetch where it returns; b17 llog read directly; uncertainty flagged inline.
Status:: Independent reference document — informational, not a recommendation.

Reader’s note

This fact-sheet builds on the b17 singularity-info-crisis llog and applies its information-theoretic framing to the scientific-authorship case. It carries primary text from the b17 llog and from ICMJE / COPE / Vancouver where retrievable. No conclusions about specific cases.

Section 1 — Recap of the asymmetry from b17#

Operative definition (verbatim from b17 llog, Response 3)#

The b17 llog (hell/ll/other/b/17/other_ll_2026m04d11_singularity-info-crisis-llog.rst) formalised the asymmetry as follows. Quoted verbatim from Response 3, section “The core problem, stated precisely”:

“Your verification bandwidth (bits you can meaningfully review per unit time) is now permanently smaller than your generation bandwidth (bits AI can produce per unit time). This gap will only widen. No amount of effort closes it.”

“This means the question shifts from ‘how do I verify everything?’ to ‘how do I allocate finite verification bandwidth for maximum reliability?’”

This is the core asymmetry on which the rest of this fact-sheet builds. It applies to any two-party interaction in which one party generates content faster than the other party can meaningfully review it. The framing is deliberately content-neutral: it does not depend on whether the generator is an AI, a human team, or a single fast human writing under pressure.

The seven information-theoretic coping strategies (verbatim, b17 Response 3)#

Quoted in order from the b17 llog. Section headings preserved as written.

1. Sample, don’t scan.

“Quality control in manufacturing solved this long ago. You can’t inspect every unit. Instead: random sampling plus targeted sampling at known failure modes. For AI output, this means: spot-check randomly and probe specifically where AI is known to be weak (subtle logical dependencies, unstated assumptions, things that require lived experience).”

2. Checksums, not full reads.

“In data transmission, you don’t verify every bit — you verify a short hash that’s sensitive to any corruption. The cognitive analog: identify high-leverage verification points. For a mathematical argument, the checksum is: ‘does the conclusion actually follow from these premises and only these premises?’ If you verify the logical spine, local prose errors matter less.”

3. Adversarial probes over confirmatory review.

“Trying to confirm a 10-page document is exhausting and unreliable. Trying to break it is efficient. One counterexample is more informative than twenty agreeing paragraphs. This is why your multi-panel adversarial review system is information-theoretically sound — you’re already doing this.”

4. Rate-distortion tradeoff — choose your acceptable error rate consciously.

“Shannon’s rate-distortion theory says: for a given channel capacity, there’s an optimal tradeoff between throughput and fidelity. You must choose. Trying for zero errors at your current production rate is impossible. The honest options are:

- Produce less, verify more (reduce rate, reduce distortion) *- Produce more, accept more errors will slip through (increase rate,

accept distortion)*

*- Produce more, invest in better compression of verification (better
heuristics)”*

5. Redundancy as error correction.

“Error-correcting codes add structured redundancy so that errors can be detected and fixed. The cognitive analog: get the same question analyzed independently from multiple angles. If three independent analyses converge, confidence rises exponentially. If they diverge, you’ve found where to focus.”

6. Bound the blast radius.

“In engineering, when you can’t prevent all failures, you contain them. Publish in layers: inner layer (heavily verified core claims), outer layer (exploratory, clearly marked as less verified).”

7. Preserve your own clock (the slow-decoder clock).

“The deepest danger isn’t that AI produces errors. It’s that AI pace displaces the slow cognitive processes where your deepest understanding forms — the sitting-with-it you mentioned. Information theory has a concept called processing gain: a slower decoder that integrates over a longer window extracts signal that a faster decoder misses entirely. Your meditation on these questions is not inefficiency. It is a higher-gain decoding process. Protect it.”

These seven strategies are the toolkit Section 4 below repurposes as authorship-verification practices.

Why the asymmetry framing is content-neutral#

The b17 framing treats verification bandwidth as a property of the reviewer and generation bandwidth as a property of the generator. The framing does not require the generator to be AI. It applies equally well to:

a single human reviewer evaluating a multi-author consortium paper;
a senior author signing off on a junior author’s data analysis they have not re-run;
a peer reviewer assessing a 200-page methods supplement under a two-week deadline;
an AI co-author asked to take accountability for results it cannot re-execute.

The remainder of this fact-sheet exploits exactly this content-neutrality.

Section 2 — Where in scientific authorship the asymmetry hits hardest#

This section walks each phase of the standard scientific authorship workflow and scores asymmetry severity using a four-level scale (Low / Medium / High / Severe). The structural reason is given in each case. The phase ordering follows a typical CRediT-style decomposition.

Conceptualization / problem framing#

Severity: Low.

Conceptualization is the phase in which a small number of agents debate “what is the right question?”. The output is short (a paragraph, a model sketch, a research-question statement), iterations are slow, and the artefacts are themselves the verification. There is no high-bandwidth generation stream that outruns the reviewer because the deliverable is the discussion itself. A co-author can re-read a conceptualization paragraph in minutes and form a meaningful judgement on it.

The asymmetry can still bite when one party generates many candidate framings that another party cannot evaluate at the same rate, but the consequences of accepting an unverified framing are bounded by downstream steps that re-test the framing against data and methodology.

Literature review#

Severity: High.

Literature review is the first phase where generation bandwidth materially exceeds verification bandwidth, especially when an AI generator or a large team produces a citation list. Each cited claim implicitly asserts: “I have read this paper, I have understood its scope and limitations, and I assert that it supports the claim I am making here.”

The verification cost per citation (read the paper, check the claim, re-examine the conclusion in context) is high; the generation cost is low. A reviewer signing off on a literature review with 80 citations is implicitly trusting that someone has done the 80 verifications. If no one has, the literature review is a sequence of unchecked checksums.

Methodology design#

Severity: Medium.

Methodology is closer to conceptualization than to data generation: it is a short, dense artefact whose structure can be checked locally. A reader who understands the field can recognise whether the design choices are appropriate without re-running anything. The asymmetry rises when the methodology references many sub-methods, software libraries, or hyperparameter choices buried in a supplement — at that point, methodology shades into data generation in terms of verification cost.

Data generation / experimental work#

Severity: Severe.

This is the phase where the asymmetry is structurally maximum. By design, data generation produces orders of magnitude more bits than any reviewer can consume in a meaningful time window. A 10,000-row dataset, a 10-hour simulation, a wet-lab notebook spanning two years — none of these can be re-verified by a co-author who did not run them.

The standard mitigations are upstream verification (good protocol design), sampling-based audit (spot-check a fraction of runs), and trust in the generator. None of these close the gap. The reviewer’s confidence is bounded by their confidence in the generator, not by their own direct inspection.

Statistical analysis#

Severity: High.

Statistical analysis sits between methodology and data generation in asymmetry severity. The artefacts (code, model specifications, output tables) are inspectable in principle, but the verification cost of re-running an analysis with different defaults, checking for hidden hyperparameter choices, and confirming that the reported numbers actually come from the posted code is high.

Many co-authors on statistical-heavy papers in practice trust the analyst rather than re-running the code. The asymmetry is not as severe as raw data generation (the artefact is finite and re-executable), but it is high because the cost of meaningful verification scales with model complexity.

Drafting#

Severity: Medium.

Drafting (initial prose production) is the phase where AI assistance is most visible and most easily verified. A reviewer can read a 6,000-word draft in an hour or two and form a substantive judgement. The asymmetry exists — the draft can be generated in minutes but reviewed in hours — but it is bounded by the artefact size, which is a few thousand words rather than a gigabyte of data.

The dangerous variant is when drafting includes fabricated citations or fabricated equations that look superficially correct; these route the asymmetry through the literature-review channel (high severity) rather than the drafting channel itself.

Interpretation / discussion#

Severity: Medium.

Interpretation is high-leverage but low-volume. Like conceptualization, the deliverable is short and the iterations are slow. A co-author can re-read the discussion section and form an opinion. The asymmetry bites mainly when the interpretation depends on data or analysis the reviewer cannot independently re-execute — in which case the asymmetry is inherited from the data-generation or statistical-analysis phase, not native to interpretation itself.

Final approval / sign-off#

Severity: Severe.

Final approval is the formal moment at which each author affirms the entire manuscript. By construction it asks the slowest reviewer to take responsibility for the entire fast-generator output. The verification bandwidth available is fixed (a few hours to a few days); the generation bandwidth that produced the manuscript may have spanned years across multiple specialists.

Final approval is the phase in which the bandwidth gap is operationalised into a signature. It is the formal contract that the rest of the workflow points toward. This is where the asymmetry is most dangerous because the signature claims a completeness of review that the bandwidth math cannot support.

Post-publication accountability#

Severity: Severe.

Post-publication accountability is what each named author commits to in ICMJE criterion 4: agreement to investigate and resolve questions about the work’s integrity. The asymmetry is severe because the questions that arise post-publication can be about any part of the work, in any timeframe, with no advance notice of which part will be challenged.

A co-author who could not verify the data at time of submission has no path to resolve questions about that data three years later. The asymmetry is not just instantaneous but cumulative: the gap that existed at submission remains the gap that must be closed during a post-publication challenge, plus any new uncertainty introduced by the passage of time.

Per-phase summary#

Verification-bandwidth asymmetry by authorship phase#
Phase	Severity	Structural reason
Conceptualization / framing	Low	Short artefact; deliverable is the discussion itself; iterations are slow.
Literature review	High	Each citation is an implicit checksum the reviewer rarely re-evaluates.
Methodology design	Medium	Dense but inspectable artefact; shades higher when sub-methods proliferate.
Data generation	Severe	Output volume structurally exceeds any reviewer’s bandwidth by orders of magnitude.
Statistical analysis	High	Re-execution cost is high; verification scales with model complexity.
Drafting	Medium	Bounded artefact size; verifiable in hours; routes through citations to higher-severity phases.
Interpretation / discussion	Medium	Short artefact; inherits severity from data/analysis when dependent on them.
Final approval	Severe	Operationalises the gap into a signature; claims completeness the bandwidth math denies.
Post-publication accountability	Severe	Cumulative gap; questions can target any phase at any future time.

Section 3 — ICMJE-criterion sensitivity to verification asymmetry#

The ICMJE recommendations specify four criteria, all of which must be met by every named author. Quoted text below is drawn from secondary summaries of the ICMJE recommendations document; direct fetch of the ICMJE site was not retrievable from this sandbox.

ICMJE four criteria (working text)#

The four ICMJE criteria, as restated across multiple secondary sources [QUOTE NEEDS VERIFICATION: original ICMJE site not retrievable from sandbox; text below assembled from secondary summaries]:

“Substantial contributions to the conception or design of the work; or the acquisition, analysis, or interpretation of data for the work; AND”
“Drafting the work or reviewing it critically for important intellectual content; AND” (revised in May 2023 from “revising” to “reviewing”)
“Final approval of the version to be published; AND”
“Agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.”

“All those designated as authors should meet all four criteria for authorship, and all who meet the four criteria should be identified as authors.”

Criterion 1 — Substantial contribution#

Sensitivity to verification asymmetry: Low.

Criterion 1 is about production, not verification. It asks whether the candidate author has contributed to conception, design, acquisition, analysis, or interpretation. The verification-bandwidth gap is silent on this question: a fast generator can contribute substantially without ever having to verify anyone else’s contribution.

This criterion is therefore the most preserve-able under high asymmetry. A party who can only generate (and not verify) can still satisfy it cleanly, as long as their generation is itself substantial.

Criterion 2 — Drafting or critically reviewing#

Sensitivity to verification asymmetry: Medium.

Criterion 2 has two limbs joined by or: drafting, or critically reviewing. Drafting is a generation activity (low asymmetry sensitivity). Critically reviewing is a verification activity (high asymmetry sensitivity).

A candidate who satisfies the drafting limb but not the reviewing limb still satisfies criterion 2 as written. The asymmetry sensitivity of criterion 2 is therefore mixed: it depends on which limb is being invoked. The May 2023 revision (from “revising” to “reviewing”) loosened the verification requirement slightly — “reviewing” does not require making changes, only critically reading.

Criterion 3 — Final approval#

Sensitivity to verification asymmetry: High.

Final approval is a single binding act applied to the entire manuscript. The verification bandwidth required to make a meaningful final approval is the full manuscript bandwidth, condensed into the approval timeframe (often days). For most multi-author papers, the bandwidth math says final approval cannot be done thoroughly by every author.

In practice, criterion 3 is often satisfied by a delegated trust model: each author affirms the parts they generated and trusts colleagues for the rest. This works socially but is not what the criterion literally requires. The asymmetry between literal interpretation and practice is large but is absorbed by the community’s tacit conventions.

Criterion 4 — Accountability#

Sensitivity to verification asymmetry: Severe.

Criterion 4 is the criterion most sensitive to the asymmetry, and by a wide margin. It requires not merely a one-time approval but an ongoing commitment to investigate and resolve questions about any part of the work at any future time. The verification bandwidth needed to honour this commitment is unbounded in advance (because any part of the work could be challenged), distributed across time (because the challenge can come at any moment), and tied to the author’s continued capacity to act (because investigation and resolution require active engagement).

A party with severely limited verification bandwidth (whether human or AI, present or deceased) cannot honour criterion 4 in the strict sense. They can only honour it via delegated accountability (the consortium / corresponding-author model) or via substitute accountability (a next-of-kin / legal-representative arrangement in the deceased-author case).

The alone-disqualifying threshold#

Of the four ICMJE criteria, criterion 4 is the criterion whose failure is most-strongly disqualifying when verification asymmetry is severe, because:

Criterion 1 can be satisfied by generation alone.
Criterion 2 can be satisfied by the drafting limb alone.
Criterion 3 is satisfied by a single point-in-time signature.
Criterion 4 is the only criterion that imposes an ongoing verification burden over an unbounded time horizon.

Criterion 1 is the criterion least sensitive to the asymmetry, and is therefore the criterion that survives most cleanly under severe asymmetry. A party who can only generate (and not verify) is structurally well-suited to satisfy criterion 1 and structurally ill-suited to satisfy criterion 4.

Summary table#

ICMJE-criterion sensitivity to verification asymmetry#
Criterion	Sensitivity	Why
1 — Substantial contribution	Low	Generation-side criterion; verification not required to satisfy.
2 — Drafting or critically reviewing	Medium	Mixed: drafting limb is generation-side, reviewing limb is verification-side.
3 — Final approval	High	Single binding act over full manuscript; bandwidth math rarely supports literal interpretation.
4 — Accountability	Severe	Unbounded ongoing verification burden over unbounded time horizon.

Section 4 — Repurposing the seven coping strategies as authorship-verification practices#

This section walks each b17 strategy and asks: what does this look like when reframed as an authorship-verification practice? For each strategy, the structure is:

original information-theoretic statement (quoted from b17);
proposed authorship-verification application;
what an editor / reviewer / co-author could concretely do under it;
what fails if the strategy is not applied.

Strategy 1 — Sample, don’t scan#

(i) Original (b17): “Random sampling plus targeted sampling at known failure modes. … spot-check randomly and probe specifically where AI is known to be weak.”

(ii) Authorship application: A reviewing co-author commits not to reading the entire manuscript with equal density, but to (a) reading randomly chosen sections in depth, and (b) probing specifically the sections most likely to contain errors given the generator’s known weak spots (e.g., for an AI co-author: citation accuracy, factual claims requiring lived experience; for a junior human co-author: methodological subtleties; for a fast-typing senior: domain boundaries).

(iii) Concrete action: Each co-author declares in the contribution statement which sections they reviewed in depth and which they only sampled. Editors can require this declaration.

(iv) What fails without it: Uniform-density reading exhausts the reviewer’s bandwidth on low-leverage content, leaving high-leverage failure modes unchecked. The reviewer reports completed review but has effectively inspected only an arbitrary fraction.

Strategy 2 — Checksums, not full reads#

(i) Original (b17): “For a mathematical argument, the checksum is: ‘does the conclusion actually follow from these premises and only these premises?’ If you verify the logical spine, local prose errors matter less.”

(ii) Authorship application: Each co-author identifies a small number of load-bearing claims in the manuscript and verifies those, rather than scanning all paragraphs. Checksums for scientific papers include: does the abstract claim match the results table; do the conclusions follow from the stated effect sizes; are the cited works actually saying what the citation says they say.

(iii) Concrete action: Reviewers and co-authors maintain a short “checksum list” per manuscript (e.g., five high-leverage claims). The list is part of the review record.

(iv) What fails without it: Reviewer attention spreads across hundreds of small claims, none of which are structurally critical, while the load-bearing inferential spine remains unverified.

Strategy 3 — Adversarial probes over confirmatory review#

(i) Original (b17): “Trying to confirm a 10-page document is exhausting and unreliable. Trying to break it is efficient. One counterexample is more informative than twenty agreeing paragraphs.”

(ii) Authorship application: A co-author’s review duty is reframed from “confirm the claims are correct” to “try to break the claims and report what survives.” This aligns with standard adversarial peer review but pushes the same discipline into internal co-author review.

(iii) Concrete action: Each named co-author submits an “attempted-break log”: the list of attacks they tried on the manuscript and which ones the manuscript withstood. Held-versus-breach records replace silent PASS-style sign-off.

(iv) What fails without it: Confirmatory review devolves into “looks-fine” sign-off. Errors are not caught because no one was trying to catch them. The asymmetry is compounded by an attention pattern that is information-theoretically the wrong choice.

Strategy 4 — Rate-distortion tradeoff#

(i) Original (b17): “Trying for zero errors at your current production rate is impossible. … Produce less, verify more; or produce more, accept distortion; or invest in better compression of verification.”

(ii) Authorship application: Each manuscript declares its operating point on the rate-distortion curve. Has the author team prioritised throughput (many papers, lower per-paper verification) or fidelity (few papers, deep verification)? A field-norm that obscures this choice is dishonest about its own error rate.

(iii) Concrete action: Journals (or author teams) report a verification-density estimate per paper: hours of meaningful review per 1,000 words, or per data point, or per citation. Aggregated across a field, this exposes the field’s average rate-distortion choice.

(iv) What fails without it: Fields drift toward the “produce-more-accept-distortion” corner without ever announcing the choice. Readers assume each paper has been thoroughly checked; in fact the field’s verification budget has been silently spread thinner over time.

Strategy 5 — Redundancy as error correction#

(i) Original (b17): “Get the same question analyzed independently from multiple angles. If three independent analyses converge, confidence rises exponentially. If they diverge, you’ve found where to focus.”

(ii) Authorship application: Critical claims are checked by multiple independent verifiers whose methods do not share a common failure mode. Multi-reviewer peer review already attempts this; co-author review can add a second independent verification axis.

(iii) Concrete action: Editors assign reviewers from divergent sub-fields rather than redundant ones from the same sub-field. Author teams designate at least two co-authors per load-bearing claim, with the requirement that they verify independently (without conferring first).

(iv) What fails without it: Three reviewers from the same sub-field share the same blind spots; their convergent endorsement is not three independent checks but one check repeated three times.

Strategy 6 — Bound the blast radius#

(i) Original (b17): “Publish in layers: inner layer (heavily verified core claims), outer layer (exploratory, clearly marked as less verified).”

(ii) Authorship application: Each manuscript explicitly partitions its claims into verification tiers: strongly-verified core claims; moderately-verified supporting claims; exploratory or speculative claims marked as such. Authorship-level accountability is graded by tier rather than applied uniformly.

(iii) Concrete action: Papers include a “verification tier” annotation on each major claim. Retraction and correction policies treat tier-1 claims more severely than tier-3 claims. Authorship attribution can be tiered too (full author for tier-1 claims; contributor acknowledgement for tier-3).

(iv) What fails without it: A single high-stakes claim and a single low-stakes claim carry the same authorship weight, which means either accountability is over-applied to exploration (chilling new work) or under-applied to load-bearing claims (degrading reliability of the field).

Strategy 7 — Preserve the slow-decoder clock#

(i) Original (b17): “The deepest danger isn’t that AI produces errors. It’s that AI pace displaces the slow cognitive processes where your deepest understanding forms. … A slower decoder that integrates over a longer window extracts signal that a faster decoder misses entirely. … Your meditation on these questions is not inefficiency. It is a higher-gain decoding process. Protect it.”

(ii) Authorship application: The authorship process protects time-for-thought as an explicit verification resource. A co-author’s verification budget includes not just hours of reading but also a minimum number of days of integration time — because some verification can only happen by sitting with the material over a long window.

(iii) Concrete action: Journals impose minimum dwell-time between manuscript submission and final author sign-off. Author teams build in “slow weeks” between drafting and approval. The corresponding author attests not just that the manuscript has been reviewed but that each named author has had at least N days with the final version before signing.

(iv) What fails without it: All verification is compressed into the fast-decoder window. Errors that would only have been caught by integration over days (logical inconsistencies, subtle drift between sections, slow realisations about implications) remain undetected. This is the failure mode b17 flagged as the deepest danger.

Summary — which strategies repurpose cleanly#

Strategies 1, 2, 3, 5, and 7 repurpose cleanly as authorship-verification practices — they translate directly into reviewer or co-author actions without significant conceptual stretching. Strategies 4 and 6 are also repurposable but require more institutional infrastructure (rate-distortion reporting standards; tiered-claim conventions) and so are less immediately actionable for an individual co-author.

Section 5 — When does verification asymmetry ALONE disqualify someone from authorship?#

This section examines the structural question: does severe verification-bandwidth asymmetry, by itself, disqualify a candidate from authorship? Both sides are argued, the threshold is identified, and the AI-as-asymmetric-party case is compared with the human-in-a-kilo-author- collaboration case.

Case for alone-disqualifying#

The strongest case for “asymmetry alone disqualifies” rests on ICMJE criterion 4: agreement to investigate and resolve questions about any part of the work. The argument runs:

Criterion 4 is not contingent on which part of the work is challenged.
A party with severely limited verification bandwidth cannot, in principle, investigate and resolve questions about parts of the work they could not verify in the first place.
Therefore severe asymmetry forces a structural failure of criterion 4.
Since all four ICMJE criteria must be met, severe asymmetry alone disqualifies.

The argument is clean but rests on a literal reading of criterion 4.

Case for not-alone-disqualifying#

The strongest case against “asymmetry alone disqualifies” rests on the existing exceptions in practice:

Deceased authors are routinely retained on papers although they cannot honour criterion 4 in any literal sense.
Large consortium authors (5,000-author particle-physics papers) are listed as authors although no single one can verify the full work.
Senior authors on team papers regularly sign off on parts they did not re-execute.
If asymmetry alone were disqualifying, none of (1)–(3) would be valid authorship arrangements; but the community treats all three as valid.
Therefore, in practice, asymmetry alone is not treated as disqualifying; what disqualifies is the failure to attach asymmetry to a mitigating structure (delegated accountability, substitute accountability, attribution discipline).

The threshold#

The threshold that separates “asymmetry disqualifies” from “asymmetry tolerated” is whether the asymmetric party is attached to a mitigating accountability structure.

If yes (e.g., a consortium author whose specific contribution is named and accounted for by a corresponding-author/spokesperson model; a deceased author whose responsibility is assumed by a next-of-kin or legal representative; a junior author whose senior co-authors take responsibility for the integrated work): asymmetry is tolerated.
If no (e.g., an authorship that asserts the four ICMJE criteria literally for every author with no acknowledged mitigating structure): asymmetry should disqualify, because the literal claim of meeting all four criteria is false.

The threshold is therefore not the asymmetry itself but the honesty of the authorship attestation. An attestation that admits the asymmetry and attaches it to a mitigating structure is tolerable. An attestation that denies the asymmetry while quietly relying on it is not.

The AI-as-asymmetric-party case#

When the asymmetric party is an AI, the structural picture is:

The AI can generate at high bandwidth (often higher than any human co-author).
The AI cannot verify the parts of the work it did not generate, at least not in the same sense a human co-author can (no persistent memory across reviews; no real-time access to lab notebooks; no capacity to be deposed three years post-publication).
The AI cannot be deposed, sued, sanctioned, or held legally accountable; AI is not a non-legal entity in COPE’s framing [QUOTE NEEDS VERIFICATION: COPE position statement on AI tools, summarised in search results].
There is no mitigating structure of the consortium type (a corresponding author who absorbs the AI’s accountability gap) unless one is deliberately constructed.

So the AI-as-asymmetric-party case fails the threshold unless a mitigating structure is explicitly attached. COPE’s current position is that AI cannot be an author at all, which is structurally a refusal to construct the mitigating structure.

The kilo-author-human case#

When the asymmetric party is a human in a 5,000-author consortium, the picture is:

The human can generate (their specific contribution is small relative to the whole, but real).
The human cannot verify the parts of the work outside their specific contribution; the bandwidth gap is overwhelming.
The human can in principle be deposed, sued, sanctioned — they are a legal entity with persistent identity.
The mitigating structure exists by convention: collaboration spokespersons, internal review boards, layered review processes, named service contributions [QUOTE NEEDS VERIFICATION: hyperauthorship conventions described in search results].

The kilo-author-human case therefore passes the threshold (mitigating structure exists). The accountability is delegated to the collaboration as a whole, not to each individual author’s literal verification of the full work.

Do they reduce to the same problem?#

Partial reduction. The two cases share the structural shape: “generation-capable party cannot verify the full work; therefore literal ICMJE criterion 4 fails for this party in isolation; therefore a mitigating accountability structure is required.”

But they diverge at the legal-entity layer: the kilo-author human remains a legal entity who can be held accountable post-hoc; the AI is not. This means the content of the required mitigating structure differs:

For the kilo-author human, the mitigating structure can rely on the human’s legal standing as a backstop.
For the AI, the mitigating structure must include a legal-entity backstop external to the AI (a sponsoring human author, an institution, an explicitly named accountability party).

So the asymmetry reduces to the same information-theoretic problem but to different legal-and-institutional problems. The fact-sheet records this without judging which mitigating structure is appropriate for any specific case.

Section 6 — Cross-reference to the deceased-author and consortium cases#

Both the deceased-author rule and large-consortium author lists are cases where named authors cannot complete verification of the work they sign. This section maps each onto the verification-asymmetry framing.

Deceased-author rule#

The deceased-author case is the maximum-asymmetry case: the deceased author’s verification bandwidth is zero going forward. They cannot respond to any post-publication question, cannot investigate, cannot resolve.

ICMJE provides no specific deceased-author guidance [QUOTE NEEDS VERIFICATION: secondary sources state ICMJE makes no specific reference to deceased authors and that strict reading of criterion 4 cannot be satisfied posthumously]. In practice, journals and publishers have constructed mitigating structures:

Substitute accountability: a next-of-kin or legal representative approves the work on behalf of the deceased author [QUOTE NEEDS VERIFICATION: described in COPE case discussion and Science Editor article].
Pre-mortem attestation: if the deceased author signed off before death, that signature is preserved; criterion 3 was met at the time; criterion 4 is implicitly assumed to have been met pre-death.
Co-author absorption: remaining co-authors implicitly accept that post-publication accountability for the full work falls on them, not the deceased.

Mapping to the verification-asymmetry framing: the deceased-author case exhibits all the structural features of severe asymmetry (criterion 4 cannot be literally met), and the community’s mitigation is substitute accountability via a legal-entity proxy. This is exactly the threshold described in Section 5: asymmetry is tolerated when attached to a mitigating accountability structure.

Large-consortium / kilo-author case#

The kilo-author case (ATLAS, CMS, large genomics consortia) exhibits a distributed asymmetry: no single author can verify the full work, but the asymmetry is spread across thousands of partial verifiers.

The community’s mitigation here is delegated and collective accountability:

Each consortium member typically performs service work that qualifies them for sign-off on publications during a defined period [QUOTE NEEDS VERIFICATION: ATLAS authorship convention summarised in hyperauthorship sources].
Specific contributions are listed in an end-note or institutional table.
A small number of spokespersons / corresponding authors absorb the individual-author accountability gap.
The consortium as a named entity may also be cited (in some conventions the consortium name appears as the author and individuals appear in a supplementary list).

Mapping to the verification-asymmetry framing: the consortium case distributes the verification load across many partial verifiers and attaches them to a collective accountability structure. Asymmetry per individual is severe; aggregate accountability is preserved by the structure.

Comparison#

Verification-asymmetry mapping for existing exceptions#
Case	Asymmetry severity	Mitigating structure	Legal-entity status
Deceased author	Maximum (no future verification possible)	Substitute accountability via next-of-kin / legal representative	Estate / legal proxy
Kilo-author consortium	Severe per individual; aggregate covered	Delegated / collective accountability via spokesperson / consortium entity	Each author is a legal entity; consortium may also be one
AI co-author (hypothetical)	Severe and structurally different (no persistent identity)	Not standardised; would require external legal-entity backstop	AI is not a legal entity (per COPE)

The takeaway for the framing#

The verification-asymmetry framing unifies these three cases under a single structural lens: all three involve named authorship attached to verification gaps that cannot be closed by the named author. The existing exceptions (deceased, consortium) are tolerated because the community has built mitigating accountability structures around them. The AI case is not tolerated yet because no comparable mitigating structure has been built and ratified.

Whether a comparable structure can be built, should be built, and what form it should take is the deliberation the main b19 session is working through. This fact-sheet does not pre-empt that deliberation.

Sources and verification status#

Primary text retrieved directly#

b17 llog: source/matheology/hell/ll/other/b/17/other_ll_2026m04d11_singularity-info-crisis-llog.rst (read directly; quotations in Section 1 are verbatim).

Secondary sources via WebSearch (WebFetch unavailable in sandbox)#

The following sources were accessed via WebSearch snippets only. Quoted material attributed to these sources is flagged [QUOTE NEEDS VERIFICATION] in-line above and should be checked against the primary documents before any consequential use.

ICMJE Defining the Role of Authors and Contributors, official recommendations page (icmje.org/recommendations/browse/roles-and-responsibilities/defining-the-role-of-authors-and-contributors.html).
ICMJE Up-Dated Recommendations May 2023, news/editorial page.
COPE Authorship discussion document (publicationethics.org/guidance/discussion-document/authorship).
COPE Authorship and AI tools position statement (publicationethics.org/guidance/cope-position/authorship-and-ai-tools).
COPE Author deceased prior to submission case discussion.
Science Editor, The Authorship of Deceased Scientists and Their Posthumous Responsibilities (csescienceeditor.org).
Mega-authorship implications: How many scientists can fit into one cell? (Taylor & Francis, accountability in research, 2024).
Physics paper sets record with more than 5,000 authors (Nature, 2015).
The ATLAS Experiment at the CERN Large Hadron Collider (IOPscience and CERN Document Server).
AIP History Programs, From Lone Genius to Wisdom of the Crowd: Hyperauthorship in High-Energy and Astrophysics.

All quotations from these secondary sources are flagged in-line. The analytical claims of this fact-sheet do not depend on the literal wording of these sources; they depend only on the structural features documented across multiple sources independently.

End of fact-sheet.