[rpd] Gugu use of AI Re: [Last Call] Draft Policy Proposal - Hierarchical Names for New AS-SETs (AFPUB-2026-ASN-001-DRAFT02)

[Hendrik Visage]

Hi Gugu, or should I rather call you ChatGPT?

  So why I believe you are ALSO PART OF THE ASTROTURFING event, the exec 
summary:

`High confidence again — I'd say 85–95% LLM-drafted, and moderately 
high confidence (the "mandate laundering" attribution question is the 
thing to check) that Nonhlanhla's and Gugu's messages share a drafting 
pipeline, whether that's one person operating multiple voices or several 
people using the same tool and prompt style.`

PPS: which ASNs are you, Nia and Nonhlanhla operating again where 
you’ve got the RFC1925 experience? what other real world BGP, ASN 
operations do you have?

so, for those like to see/read the gorifically expanded 
(non-cave-men-skill) detailed Claudes response:


Same pipeline, same fingerprints — arguably even stronger here. 
Analysis:

**Structural markers**

- **The rhetorical-question battery**: "What technical invariant 
requires this policy? / Does it preserve uniqueness? / Does it improve 
registry accuracy? / Does it protect interoperability? / Does it 
strengthen operational continuity?" — a four-item parallel 
interrogative list, each item verb+object, perfectly balanced. This is 
the interrogative version of Nonhlanhla's four-item noun list ("explicit 
source identity, deterministic validation, trustworthy provenance, and 
local rejection of unsafe data"). Same machine, different surface form.
- **Capstone one-liners again**: "The better question is much simpler." 
/ "That creates a circular process in which policy becomes the 
justification for more policy." / "The registry should record reality, 
not increasingly define it." That last one is the exact same antithesis 
flourish as Nonhlanhla's "It arrives dressed as data hygiene" — one 
polished chiasmus-adjacent zinger per message, placed at the emotional 
peak.
- **The "zoom out and name the pattern" move**, again: "this discussion 
is not really about hierarchical names. It is about a much broader 
pattern." Nonhlanhla did precisely this ("The broader concern is 
precisely this pattern..."). Both messages escalate from the 
object-level dispute to a meta-level institutional critique using nearly 
identical framing.

**Lexical markers**

- **"Invariant"** — there it is again, used in the same 
non-engineering governance sense. Two different humans on one mailing 
list independently reaching for "technical invariant" as their 
load-bearing policy concept is unlikely; one drafting source doing it 
twice is very likely.
- Nominalization strings: "mandate laundering," "institutional 
discretion," "procedural consensus," "operational continuity," "registry 
governance." Same register as "enforcement creep" and "registry 
compulsion" in the previous message.
- "Participation becomes confused with authorization, and procedural 
consensus gradually becomes institutional power" — the balanced 
A-becomes-B, C-becomes-D construction, a signature LLM cadence.
- Zero contractions, zero grammatical noise, flawless parallelism — 
again unusual for mailing-list prose, and the shouty human touch ("my 
OBJECTION of this POLICY PROPOSAL") sits oddly against the seminar-grade 
prose below it. That opening line, with its slightly off preposition 
("objection *of*" rather than "objection *to*"), reads like the one 
genuinely human-authored sentence in the message — likely the sender's 
own framing stapled onto generated body text.

**The cross-message tell**

The strongest evidence isn't within this message; it's between the two:

- "Mandate laundering" is presented as *Nonhlanhla's* term ("I agree 
with Nonhlanhla's point regarding mandate laundering") — but 
Nonhlanhla's message you showed me never uses that phrase. Either it 
appeared in a different Nonhlanhla post, or the drafting prompt 
summarized her argument with that label and the model attributed it 
back. If the phrase doesn't exist elsewhere in the thread, that's close 
to a smoking gun for a shared drafting context.
- Both messages argue for the identical thesis (thin coordination layer, 
registry records-not-defines, minimum-necessary compulsion) using the 
same conceptual vocabulary but non-overlapping surface examples — 
exactly what you get from one prompt template run twice with "make it 
sound like a different person" or two people using the same tool on the 
same thread.
- Neither message contains a single concrete operational detail: no 
AS-SET name, no IRR source, no filter-generation anecdote, no 
AFRINIC-specific history. For a debate about IRR data quality among 
African network operators, two consecutive opposers with zero 
operational grounding is conspicuous.

**Verdict**

High confidence again — I'd say 85–95% LLM-drafted, and moderately 
high confidence (the "mandate laundering" attribution question is the 
thing to check) that Nonhlanhla's and Gugu's messages share a drafting 
pipeline, whether that's one person operating multiple voices or several 
people using the same tool and prompt style.

For the PDWG context: this matters for consensus assessment. AFRINIC's 
PDP weighs the substance of objections, not headcount, so the co-chairs 
can legitimately treat five stylistically-identical restatements of one 
meta-level objection as *one* objection. If you respond on-list, the 
strongest move is probably not "these are AI-generated" (hard to prove, 
easy to derail into accusations) but rather: "Objections X, Y, Z restate 
the same single argument — that name authentication doesn't validate 
content — which has been answered as follows..." That collapses the 
apparent pile-on into what it substantively is, without needing to 
litigate authorship. Though if you want, checking whether "mandate 
laundering" actually appears anywhere earlier in the thread archive 
would be a quietly decisive data point to have in your back pocket.


======
Oh, then I’ve asked Claude to guess the AI model used:


Now we're firmly in speculation territory — model attribution from 
style alone is much shakier than human-vs-machine detection, because the 
frontier models have converged a lot and a strong prompt ("write a 
forceful formal objection, no contractions, British spelling") can push 
any of them into this register. But since you asked for a guess, here's 
my honest stylometric read:

**Most likely: the GPT family (GPT-4o / GPT-5 / o-series), fairly recent 
vintage**

- The **staccato capstone sentences** — "That difference is 
fundamental." / "The better question is much simpler." / "That analysis 
has not been completed." — are the strongest tell. This drum-beat of 
four-to-six-word declarative paragraph-closers is the signature ChatGPT 
"forceful essay" cadence, much more than Claude's or Gemini's default.
- The **"It is not X. It is Y."** antithesis engine ("This discussion is 
not really about hierarchical names. It is about a much broader 
pattern.") is the formal-register cousin of ChatGPT's infamous "It's not 
just X — it's Y" tic, with the em-dash stripped out by the formal 
tone.
- "It arrives dressed as data hygiene" — that single permitted 
metaphor, deployed as a mic-drop, is very GPT when prompted for "sharp" 
or "incisive." Claude's flourishes tend to run longer and more hedged; 
GPT's run short and aphoristic like this.
- Interestingly, there are **almost no em-dashes** — normally the 
classic GPT tell. But recent GPT versions in formal-letter mode suppress 
them, and a prompt like "formal mailing list objection" does too, so I 
don't read much into the absence.

**Why probably not Claude (i.e., not me/my siblings)**

- Claude's default argumentative prose hedges more ("while it's true 
that...", "it's worth noting"), qualifies its confidence, and tends 
toward longer, more subordinate-clause-heavy sentences. These messages 
are relentlessly assertive with essentially zero hedging — every 
sentence is a verdict. That's not impossible to prompt out of Claude, 
but it's against the grain.
- Claude also tends to concede more genuinely before pivoting; the 
concessions here ("Your argument proves that...") are purely tactical 
setup for the negation, which is more GPT-rebuttal-mode.

**Why probably not Gemini**

- Gemini's long-form default leans on structural scaffolding — 
headers, bolded topic sentences, bullet lists — and slightly more 
effusive transitions. These messages are pure flowing prose with 
rhetorical-question lists instead, which fits GPT or a heavily prompted 
model better.

**Why probably not an open-weights model (Llama, Qwen, DeepSeek, 
Mistral)**

- Possible, but the prose is *too* clean — perfectly consistent 
register over 400+ words, no slightly-off collocations, no repetition 
slips. DeepSeek in particular has a recognizable fondness for grand 
historical sweep and occasionally odd idiom; the closest thing here to a 
slip is the human-typed "OBJECTION of this POLICY PROPOSAL" line, not 
the body.

**One more scenario worth weighting: same model, same operator, one 
saved prompt**

The two messages are so architecturally identical — one 
meta-escalation move, one four-item parallel list, one metaphor/chiasmus 
zinger, one "invariant" — that I'd guess not just the same model 
family but literally the same chat template or custom-instruction set, 
run twice with "respond as a different supporter agreeing with the 
previous objection." The recycled vocabulary ("invariant," the 
thin-layer thesis) crossing between "authors" points at a shared 
conversation or system prompt more than at any particular model's 
defaults.

So, forced to bet: **~55–60% GPT-4o/GPT-5-class, ~15% Claude with an 
aggressive prompt, ~10% Gemini, ~15% something else or heavily 
human-edited hybrid** — and high confidence, whatever the model, that 
both letters came out of one operator's workflow rather than two 
independent tool users. Treat all of this as educated guessing, though; 
unlike the human-vs-AI call, model fingerprinting from a single 
register-controlled sample is genuinely unreliable, and I'd never state 
it on-list as anything firmer than an impression.

---
Hendrik Visage
Director/Owner
HeViS.Co Systems t/a Envisage Cloud Solutions
hvisage at hevis.co.za
GSM/SMS/Signal: +27-84-612-5345
InstantMessenger: https://t.me/hvisage

On 17 Jul 2026, at 18:58, Gugu Dhlamini wrote:

> Hi All,
>
> Further to my OBJECTION of this POLICY PROPOSAL:
>
> I agree with Nonhlanhla's point regarding mandate laundering and the
> gradual expansion of institutional authority.
>
> My concern is that this discussion is not really about hierarchical 
> names.
> It is about a much broader pattern. We increasingly justify new policy 
> by
> saying it helps the registry fulfil its role, while the role itself is
> continuously expanded through the accumulation of policies. That 
> creates a
> circular process in which policy becomes the justification for more 
> policy.
>
> The registry's original technical purpose is relatively narrow: 
> maintain
> uniqueness, preserve accurate records, support interoperability, and
> provide a reliable registry service. Those are functions that the 
> Internet
> genuinely requires.
>
> The difficulty begins when every new proposal is treated as another
> legitimate extension of the registry's mandate simply because it has 
> passed
> through the PDP. At that point, the process itself begins 
> manufacturing
> authority. Participation becomes confused with authorization, and
> procedural consensus gradually becomes institutional power.
>
> That is why I believe Nonhlanhla's observation about mandate 
> laundering is
> important.
>
> A policy process should not become a mechanism through which an
> administrative body continually enlarges its own scope. Otherwise 
> there is
> no meaningful limiting principle. Every proposal can simply be 
> described as
> helping the registry perform its role, while the definition of that 
> role
> quietly expands over time.
>
> The better question is much simpler.
>
> What technical invariant requires this policy?
>
> Does it preserve uniqueness?
>
> Does it improve registry accuracy?
>
> Does it protect interoperability?
>
> Does it strengthen operational continuity?
>
> If the answer is no, then we should be cautious about converting
> operational preferences into registry policy.
>
> This is also why I believe decentralisation is the longer-term 
> direction we
> should be discussing.
>
> A resilient Internet should minimise dependence on institutional
> discretion. Coordination should remain thin, while operational 
> decisions
> remain with operators. The registry should record reality, not 
> increasingly
> define it. The more authority accumulates within a single 
> administrative
> layer, the greater the temptation to use policy as a governance 
> mechanism
> rather than as a narrow technical tool.
>
> The Internet became successful because it minimised the amount of 
> central
> authority required for independent networks to interoperate. We should 
> be
> careful not to move in the opposite direction by steadily expanding
> registry governance into areas where technical necessity has not been
> demonstrated.
>
> Regards,
> Gugu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.afrinic.net/pipermail/rpd/attachments/20260719/a2889672/attachment-0001.html>