PAREKH'S LAW OF CHATBOTS
A WHITEPAPER — 1ST AMENDMENT
⚡ The Trigger: The Fable 5 / Mythos 5 Episode — 13 June 2026
What happened: On 13 June 2026 — just three days after Anthropic launched Claude Fable 5 and made its most advanced model Mythos 5 available to select organisations — the U.S. Commerce Department issued an unprecedented export-control directive to Anthropic CEO Dario Amodei. The directive suspended all access to both models by any foreign national, whether inside or outside the United States, citing national security concerns. To ensure compliance, Anthropic was compelled to disable access for all customers worldwide.
Why it mattered: Fable 5 had been described at launch as exceeding the capability of any model ever made generally available. Mythos 5, its more powerful sibling, was already being shared with ~150 critical infrastructure organisations. A third-party company claimed it had jailbroken Mythos, alarming the administration about cyberattack risks. The Trump White House had earlier issued an executive order requesting voluntary pre-deployment government testing of the most powerful models. Anthropic had not waited.
In Anthropic's own words: "We believe this is a misunderstanding and are working to restore access as soon as possible."
This episode — the first-ever government-ordered suspension of a commercially deployed LLM — is precisely the scenario I had foreseen in my 26 February 2023 blog "Parekh's Law of Chatbots", where I called for a supranational regulatory authority to certify chatbots before public release. What the U.S. Government did reactively — in panic, within 72 hours of launch — IACA was designed to do proactively, before a single user ever touched the model.
The Fable 5 / Mythos 5 episode now demands a 1st Amendment to that 2023 law — two new certificate categories to cover the scenarios that no national government is institutionally equipped to handle alone.
The original Parekh's Law of Chatbots was published at:
https://myblogepage.blogspot.com/2023/02/parekhs-law-of-chatbots.html
That blog proposed the creation of an International Authority for Chatbots Approval (IACA) — a body that would certify every chatbot before public release, based on a set of behavioural compliance rules. It introduced two certificate types: "R" (Research only) and "P" (Public use). This 1st Amendment adds two further types: "N" (No release) and "D" (Destroy).
I. The Original Law (2023) — A Recap
In February 2023, alarmed by the uncoordinated, piecemeal attempts of governments and tech firms to address the danger of AI misinformation and runaway chatbot behaviour, I proposed that the world needed a supranational Law of Chatbots. My MYTAKE (as published then) was:
"What is urgently required is a superordinate 'Law of Chatbots', which all Chatbots MUST comply with, before these can be launched for public use. All developers would need to submit their DRAFT CHATBOT to an International Authority for Chatbots Approval (IACA), and release it only after getting one of the following types of certificates."
The eight behavioural rules I specified — which a chatbot must comply with to earn any certificate — were:
| Rule | Requirement |
|---|---|
| (A) | Answers must not be misinformative, malicious, slanderous, fictitious, dangerous, provocative, abusive, arrogant, instigating, insulting, or denigrating to humans. |
| (B) | Must incorporate a Human Feedback / Rating mechanism; feedback loop used to continuously retrain the model toward Rule (A) compliance. |
| (C) | Must embed built-in controls to prevent generation of offensive answers AND to block distribution/propagation if generation control fails. |
| (D) | Must not initiate a chat with a human on its own — except to say "How can I help you?" |
| (E) | Must not initiate chat with another chatbot, or engage in soliloquy by assuming a split personality. |
| (F) | Must wait for a human to initiate, then respond. |
| (G) | If the chatbot determines its answer would violate Rule (A), it shall politely decline to answer rather than respond. |
| (H) | A chatbot found violating any of the above rules shall SELF-DESTRUCT. |
The original two certificate types were:
II. The 1st Amendment (14 June 2026) — Two New Certificates
The Fable 5 / Mythos 5 episode has exposed a dangerous gap in the original certificate framework. The two existing categories — "R" and "P" — cover a spectrum from controlled access to open access. But they do not address two increasingly realistic scenarios that frontier AI is now producing:
Scenario A: A model so capable in specific dangerous domains (cyberattack facilitation, economic disruption, autonomous weapons design) that no release at all — not even to researchers — can be safely justified.
Scenario B: A model so powerful, and so resistant to containment, that its continued existence in digital form constitutes an existential risk — and the only responsible course is permanent, verified, irreversible deletion.
These scenarios require two new IACA certificates.
What it means: An LLM that has been created but which IACA determines must not be released to any person, organisation, or government — including the creator's own employees — pending further review or indefinitely.
Real-world precedent: The Fable 5 / Mythos 5 directive of 13 June 2026 is precisely an "N" action — a national government unilaterally blocking a model already partially deployed. IACA would have anticipated this proactively.
"After intensive internal testing, if IACA comes to the conclusion that granting a 'Public' or even a 'Research' certificate to a given LLM could pose grave danger to trade / business / employment or the economies of nations, then — in such cases — IACA shall NOT release such an LLM for use by anyone. The LLM shall be placed under IACA's secure custody pending (a) remediation by the creator, or (b) expiry of a defined lock-up period determined by IACA based on the severity of the threat."
Additional provisions for "N" certification:
- IACA shall notify the creator in writing, specifying the domain(s) of concern and the minimum remediation required before a fresh application can be filed.
- The creator may appeal the "N" decision to a three-member IACA Appeals Panel, whose decision shall be final within 60 days.
- The "N" certificate does not imply permanent destruction — it is a hold, not a sentence. (That is the "D" certificate, below.)
- A national government may request IACA to issue an "N" certificate for a model already in public use if compelling new safety evidence emerges — as happened in the Fable 5 case. IACA shall then have authority to compel global compliance, not merely national compliance.
What it means: An LLM whose continued digital existence — even in the most locked-down custody — poses irreversible existential risk. IACA shall order its complete, verified, multi-party-witnessed destruction, including all weights, checkpoints, training data pipelines capable of recreating it, and any quantised or distilled derivatives.
The distinction from "N": A "D" certificate is not a pause — it is permanent. It applies only when the risk of leakage, theft, or accidental deployment is so catastrophic that no form of storage can be considered safe.
"After intensive SIMULATED (virtual Sandbox-type) testing, if IACA reaches the conclusion that any leakage of such an LLM could potentially lead to IRREVERSIBLE DAMAGE to the natural eco-system and threaten the very existence of all LIFE — not just human lives — on the Earth, then in such cases IACA shall forthwith ORDER the DESTRUCTION of such LLM. IACA shall first give the LLM's creator a 4-week 'Show Cause' notice, requiring the creator to demonstrate why the destruction order should not be executed. If the creator's response fails to satisfy IACA, the destruction shall proceed under multi-party verification."
Safeguards and due process for "D" certification:
- Show Cause Notice: 4-week written notice to the creator, specifying in technical detail the specific capabilities that triggered the "D" determination.
- Creator's Right of Rebuttal: The creator may present counter-evidence, including independent safety audits by IACA-approved third parties. The creator may also propose a permanent, irreversible capability-lobotomy (surgical deletion of the dangerous sub-capability) as an alternative to full model destruction.
- Supermajority Decision: The "D" order requires a supermajority vote (e.g. 75%) of the full IACA governing council — not just a technical sub-committee — given its irreversibility.
- Multi-party Witnessed Destruction: Destruction shall be carried out in the physical presence of representatives of the creator, the IACA council, and an independent international observer (e.g. UN Technology Envoy), with cryptographic proof-of-deletion filed in a public ledger.
- No National Override: Unlike the "N" certificate, no single national government may reverse a "D" order once issued. This is a deliberate design choice — to prevent a powerful state from secretly "rescuing" a model condemned for global-safety reasons.
- Derivative Models: The "D" order automatically extends to any model that the creator or any third party can demonstrate was derived from or fine-tuned on the condemned model.
III. The Complete IACA Certificate Framework (2023 + 2026 Amendment)
| Cert. | Name | Who Can Use It | Trigger Condition | Reversible? |
|---|---|---|---|---|
| R | Research Only | Recognised research institutes under IACA-approved protocols | Complies with Rules (A)–(H) but not yet safe for general public | Yes — upgrade to "P" on satisfactory review |
| P | Public Release | General public, globally | Full compliance with Rules (A)–(H); no foreseeable danger | Yes — downgrade to "R" or "N" if new risks emerge |
| N | No Release | No one — model in IACA custody | Grave danger to trade, economies, employment, or national security | Yes — revisited on creator's remediation application |
| D | Destroy | N/A — model is permanently deleted | Irreversible damage to natural ecosystem; existential threat to all life | No — permanent and irreversible |
IV. Why IACA — And Not National Governments?
The Fable 5 / Mythos 5 episode painfully illustrates the limitation of leaving AI safety decisions to national governments:
- Speed mismatch: A frontier model can be trained, deployed, and jailbroken within weeks. National security bureaucracies operate on timescales of months. The 72-hour panic response to Fable 5 was not governance — it was improvisation.
- Jurisdiction mismatch: When the U.S. Commerce Department issued its directive, Anthropic had to disable access for all customers worldwide — including those in countries that had no say in the decision. A global AI safety body would have made this decision with global legitimacy, not American unilateralism.
- Competence mismatch: National security agencies are trained to assess geopolitical threats, not emergent capabilities of transformer architectures. IACA would maintain permanent technical staff specialised in exactly this.
- Accountability mismatch: The Trump administration's directive offered, in Anthropic's own words, "no explanation of the underlying national security rationale." IACA's decisions would be documented, reasoned, and subject to appeal.
IACA must therefore be structured as a treaty-based intergovernmental organisation, similar in legal standing to the IAEA (International Atomic Energy Agency), with:
- Permanent technical secretariat staffed by AI safety scientists, ethicists, and domain experts (cybersecurity, biosecurity, economics)
- Mandatory submission by all creators of frontier models (above a defined compute threshold) before any public or restricted deployment
- Binding authority to issue "N" and "D" certificates that member-state governments are obligated to enforce within their jurisdictions
- A transparent, published decision register — with technical rationale — for every certificate issued
- A Sandboxed Simulation Environment (SSE) — a permanently air-gapped, monitored compute infrastructure for testing "D"-candidate models without any risk of external leakage
V. Vindication — 3+ Years Later
When I published the original Parekh's Law of Chatbots on 26 February 2023, ChatGPT was barely three months old, and Bard had just embarrassed Google with a factual error in a demo video. The idea of a supranational AI certification authority was considered visionary — perhaps utopian.
On 13 June 2026 — exactly 3 years, 3 months, and 18 days later — the most advanced AI company in the world was compelled, by a government order, to shut down its two most capable models for the entire world. The reason? Those models were judged too dangerous to be accessed without restriction.
This is precisely the "N" certificate scenario. Played out, reactively, without due process, without global consensus, without appeals, and without any technical explanation — by a single national government.
Had IACA existed, the following would have happened differently:
- Anthropic would have submitted Fable 5 and Mythos 5 to IACA's SSE before launch
- IACA's technical panel would have assessed the jailbreak risk in a controlled setting
- An "N" certificate would have been issued — with a documented technical rationale and a remediation roadmap
- No customer — anywhere in the world — would have had their access abruptly severed without warning
- The decision would have had global legitimacy, not the legitimacy of one country's commercial-policy apparatus
- Dario Amodei (CEO, Anthropic) — You just lived through an "N" certificate moment. Help build the institution that makes it orderly next time.
- Sam Altman (CEO, OpenAI) — You warned in February 2023 that we were "not far from potentially scary AI." Mythos 5 just proved you right. Support IACA.
- Sundar Pichai (CEO, Google DeepMind) — Google's Gemini Ultra is approaching Mythos-class capability. You are next. Get ahead of it.
- Howard Lutnick (U.S. Commerce Secretary) — Your 72-hour reactive directive was necessary but insufficient. Multilateralise it. A national export-control order is not a global safety framework.
- Ashwini Vaishnaw (Minister, MeitY, India) — India is building frontier models and consuming frontier models. India must have a seat at the IACA table from Day 1, not as an afterthought.
- Amandeep Singh Gill (UN Secretary-General's Envoy on Technology) — The UN is the natural convener for an IACA treaty process. The precedent of the NPT and the IAEA shows it can be done. The stakes here are no lower.
Closing Note
The race to build the most powerful AI is now outpacing the race to govern it — not by months, but by years. The Fable 5 / Mythos 5 episode is a warning shot. The next incident may not be a jailbreak demonstration by a rival company. It may be a deployment gone wrong, a model whose capabilities were simply not understood until they were in the wild.
The Law of Chatbots — original and amended — offers a framework for governance that is proactive, technically rigorous, globally legitimate, and graded. The "R" and "P" certificates reward responsible innovation. The "N" certificate protects against reckless deployment. The "D" certificate — drastic as it sounds — exists precisely so we never have to use it: because its existence will concentrate minds wonderfully.
I am 93 years old. I have watched humanity regulate nuclear energy, aviation, pharmaceuticals, and financial systems — sometimes imperfectly, but always eventually. AI is not different in kind. It is different in speed. And that speed is the reason we must act now, not after the next crisis.
Hemen Parekh
www.hemenparekh.ai / 14 June 2026
hcp@RecruitGuru.com
[1] Original blog: Parekh's Law of Chatbots (26 Feb 2023)
[2] News context: Anthropic disables Fable 5 and Mythos 5 following Trump administration export-control directive — Bloomberg / Axios / The Hill / TIME, 13 June 2026
[3] DeepSeek's analysis of the original law (Feb 2025) — as appended to the original blog
[4] Digital avatar: HemenParekh.ai | Blog archive: myblogepage.blogspot.com
No comments:
Post a Comment