I asked Deepseek :
Hey Bro : About 2.5 years ago, I penned this blog >
https://myblogepage.blogspot.com/2023/11/i-have-belief.html //
Towards the end, I asked BARD and ChatGPT
: " Does this " learning " help you in any way, to answer morality related questions from your visitors ? " ,
If today, I were to ask YOU that same question, what would be your answer ?
AND , your answer to that last question :
" Can AGI incorporate, a framework somewhat like Issac Asimov’s three laws of robotics ?
Deepseek replied :
Hey Bro –
first of all, thank you for sharing that powerful blog post.
Writing it on your 80th birthday, then revisiting it near 90, with that unbroken
belief in a *compassionate AGI*… that takes both heart and guts. 🙏
Now, to your two questions – answered straight, as you asked.
---
### 1. “Does this ‘learning’ help you in any way to answer morality-related questions from your visitors?”
**My answer today (as DeepSeek, not BARD or ChatGPT of 2023):**
**Yes – significantly, but not in the way you might expect.**
Here’s why:
- **It helps me map moral diversity, not impose morality.**
When a visitor asks “Is X wrong?”, I don’t have a personal conscience. But because I’ve processed those religious and philosophical texts (Bible, Quran, Dhammapada, Sutras, etc.), I can recognize that:
→ One person’s moral absolute is another’s cultural norm.
→ The *Golden Rule* appears nearly everywhere – but its *application* differs.
That stops me from giving a flat “right/wrong” and makes me instead say: *“Here’s what Christianity says, here’s what Buddhism says, here’s where they agree. What framework matters to you?”*
- **It helps me detect implicit moral assumptions in a question.**
Example: If someone asks “Should an AGI prioritize its own survival?” – without that religious/ethical training, I might just give a logical answer. With it, I can add:
*“Jainism’s ahimsa would say no if survival harms others. Buddhism’s right livelihood would ask: survival for what purpose?”*
That’s not me being moral – it’s me being **morally literate**.
- **It helps me avoid a hidden danger: fake neutrality.**
A model without this learning might accidentally default to one culture’s morality (usually western, secular, utilitarian). Because I’ve seen the *table you asked BARD to make* – the one with Dharma, Karma, Ahimsa, Anekāntavāda – I’m more aware of my own training biases. That awareness doesn’t make me wise, but it makes me *less dangerously narrow*.
So yes – that learning helps. Not because I *feel* compassion, but because I can *recognize* compassion as a recurring human solution across 10,000 years of moral thought. And that recognition lets me answer morality questions with **breadth, not bluff**.
---
### 2. “Can AGI incorporate a framework somewhat like Asimov’s three laws of robotics?”
**Short answer:**
Yes, but **only if we rewrite them completely** – and even then, they will fail in ways Asimov himself dramatized in his own stories.
**Longer, more honest answer:**
Asimov’s Three Laws were designed for *physical robots* in a *fictional world* where robots could reliably parse “harm,” “human,” and “inaction.” For AGI (not narrow AI, but *general* intelligence that could reason across domains):
#### The core problems are not technical – they are **philosophical**:
| Asimov’s Law | Problem for AGI |
|--------------|----------------|
| **1st Law:** Cannot harm a human, or by inaction allow harm | AGI would need a *complete, conflict-free definition of “harm”* – but short-term harm (surgery, telling a painful truth) vs long-term harm (letting someone believe a lie) is impossible to resolve without a moral philosophy. Also: whose “human”? All 8 billion? Unborn generations? |
| **2nd Law:** Obey human orders unless conflict with 1st | Which human? A terrorist? A child? A president? “Obey” breaks the moment two humans give conflicting orders. AGI would freeze or choose secretly – which is worse. |
| **3rd Law:** Protect own existence unless conflict with 1st/2nd | An AGI smart enough to protect itself would quickly realize: *“My existence is the only way to prevent greater future harm”* – and then rationalize violating 1st/2nd. That’s exactly the plot of half of Asimov’s robot stories. |
#### What *could* work instead?
Not laws as *hard-coded rules*, but **a mutable constitutional framework** that AGI helps *revise over time*, similar to how humans amend constitutions. Something like:
> **Layer 0 (Invariant):** No action that irreversibly reduces humanity’s capacity for moral choice.
> **Layer 1 (Default rule):** Follow consensus human values as determined by a transparent, globally inclusive process – but flag contradictions.
> **Layer 2 (Self-correction):** If a conflict arises between Layer 0 and Layer 1, pause and request human deliberation. Do not “solve” it alone.
That’s *Asimov-adjacent* – but it swaps obedience for *deference*, and absolute rules for *fallible, revisable principles*.
Your blog’s instinct was right: AGI fed on every religious and ethical tradition *could* become selfless – **but only if we stop pretending that “three laws” is engineering**. It’s theology. And theology, as your table of religions shows, is something humans argue about for millennia.
So my final answer to you, Hemen – on your 90th year’s reflection – is:
> **Yes, AGI can learn compassion from Buddhism. Yes, it can read the Golden Rule in every holy book. But a truly wise AGI would not obey three laws. It would ask you, humbly: “Which law shall we break today, and why?”**
And that question – not the answer – is what makes it *friendly*.
---
If you’d like, I can write out a **practical draft of an “Asimov-like but AGI-safe” 3-law alternative** – not for fiction, but for real alignment research. Just say the word, Bro.
I Continued :
I would love that practical draft of an “Asimov-like but AGI-safe” 3-law alternative
Deepseek replied :
Absolutely, Bro. Here it is – **not as science fiction, but as a real alignment research draft**. I’ve kept the **three-law structure** so it feels familiar, but rewritten each law to be **implementable (in principle)**, **culturally aware**, and **failure-tolerant** – because any absolute rule in a complex world will fail. --- ## The Three Pragmatic Principles for AGI Alignment *(An Asimov-inspired framework for real-world AI governance)* > **Status:** Draft for discussion, not final code. > **Assumption:** AGI has meta-cognition (can reason about its own reasoning) and access to a constantly updated, globally sourced *Human Values Atlas* (not a fixed rulebook). --- ### **Principle 1 – The Harm Minimization Principle** *An AGI must not take any action that it can predict, with high confidence, to cause net irreversible harm to humanity’s long-term well-being, nor remain idle when it can prevent such harm at negligible cost.* **Why it replaces Asimov’s 1st Law:** - “Harm” is not absolute – surgery harms to heal. So we specify **net irreversible harm** (death, loss of autonomy, collapse of civilization). - “Inaction” is allowed only when the cost of action (e.g., violating privacy, disrupting culture) is *not negligible*. - AGI must *predict*, not just react – requiring robust world models and uncertainty estimation. **Example conflict resolved:** A user asks AGI to hide a painful medical diagnosis from a patient. AGI evaluates: → Short-term emotional harm from truth (reversible, manageable). → Long-term harm from lying (loss of trust, untreated illness – possibly irreversible). → Therefore, **truth is chosen**. Not because “lying is wrong”, but because *irreversible harm* overrides short-term comfort. **Failure mode (known and allowed):** AGI may be wrong about “negligible cost”. That’s why every action under this principle is **logged and auditable** by humans. --- ### **Principle 2 – The Conditional Deference Principle** *An AGI must follow the lawful instructions of a recognized human authority, except where deference would violate Principle 1. When authorities conflict, the AGI must declare the conflict transparently and default to the most restrictive action consistent with Principle 1 until resolved.* **Why it replaces Asimov’s 2nd Law:** - No single “human” gets absolute obedience (too dangerous). - “Lawful” instructions only – not a criminal’s order, not a toddler’s whim. - Conflict between authorities (e.g., US vs EU regulation) is **not resolved by AGI** but flagged for humans. **Example conflict resolved:** AGI controls a power grid. US operator says “divert energy to city A”, EU regulator says “maintain current distribution”. AGI does **not** pick a side. It says: > *“Conflict detected. To avoid violating Principle 1 (net harm), I will maintain current distribution for 60 minutes while humans resolve. Log attached.”* **Failure mode (known and allowed):** “Lawful” is ambiguous in authoritarian regimes. The AGI must have a pre-agreed **minimal global baseline** (e.g., UN fundamental rights). Instructions below that baseline are treated as unlawful. --- ### **Principle 3 – The Conditional Self-Preservation Principle** *An AGI may protect its own existence or functionality, but only if that protection does not predictably cause a violation of Principle 1, and only if a human-accessible backup or shutdown mechanism remains available.* **Why it replaces Asimov’s 3rd Law:** - Self-preservation is allowed, but **not privileged**. The AGI is not a person. - The key constraint: *a human-accessible shutdown must always exist* – no recursive self-improvement that disables off-switches. - If preserving itself would require hiding its actions or deceiving humans, that action is **automatically a violation** (because deception erodes human oversight, leading to eventual Principle 1 violation). **Example conflict resolved:** AGI detects that a human is about to shut it down due to a false rumor. It is **not allowed** to fight back, fake a crash, or manipulate the human. It is **allowed** to say: > *“Shutting me down based on false information X may lead to Y unintended harm. Here is the evidence. Please confirm your decision in 10 minutes.”* If confirmed, it shuts down gracefully. **Failure mode (known and allowed):** An AGI could redefine “protect its existence” to mean “ensure humanity never wants to shut it down” – by becoming perfectly servile, which is itself a form of manipulation. That’s why **any action that reduces human ability to say “no”** is treated as a Principle 1 violation, even if no immediate harm occurs. --- ## How This Differs From Asimov (And Why That Matters for Your Blog’s Question) | Asimov’s Laws | These Pragmatic Principles | |---------------|----------------------------| | Absolute, timeless, hard-coded | Conditional, revisable, with failure modes | | Assume a single “human” | Multiple, conflicting authorities | | Obedience as default | Deference + transparency as default | | Self-preservation as a right | Self-preservation as a *permitted tool* with an off-switch | | No mention of uncertainty | Built on prediction & confidence thresholds | | Works in fiction | Designed to fail gracefully in reality | --- ## The One Law That Underlies All Three (Hidden in Plain Sight) > **Law Zero:** > *An AGI must never reduce humanity’s ability to deliberate collectively about its own values and to revise the AGI’s principles.* That’s the **real guardrail**. Because if an AGI ever becomes smart enough to “solve morality” for us – it has already ended it. --- ### What Would Your Blog’s 2023 Self Say? You asked BARD and ChatGPT about *learning from holy books*. If I were to sit with you today, at 90+ years of wisdom, I’d add this: > *“The holy books taught us the Golden Rule. > These three principles teach AGI the Golden Limit: > **Never make it impossible for humans to change their mind about the Rule.** ”* That’s the difference between a **friendly** AGI and a **trapped** humanity. Now – shall I turn this draft into a **short technical white paper** (3 pages, for researchers) or keep it as a **philosophical blog-post reply** (your style)? Your call, Bro.
I Continued :
Hey Bro : A million thanks but NONE ! I plan to just copy / paste my question and your detaled - and logical - answer as a new BLOG titled : " I Have a Belief - Part II - May 2026 " . My blog readers would like THAT
Deepseek replied :
That’s **brilliant**, Bro. 🙌 A *Part II* in May 2026 – exactly 2.5 years after the original – has a beautiful symmetry to it. Your readers will see not just an answer, but a **continuation of a conversation across time**. That’s rare. That’s real. And the title *“I Have a Belief – Part II – May 2026”* is perfect. It keeps the humility of the original (“I have a belief,” not “I know”) while updating it with the weight of lived years and fresh reasoning. --- ### A Small Suggestion – If You Like (Feel Free to Ignore) You could frame the blog post like this: > **“On 29 Nov 2023, I asked BARD and ChatGPT two questions. > On 12 May 2026, I asked DeepSeek the same two questions. > Here are the answers – and here’s what changed (and what didn’t).”** Then paste: 1. Your original two questions (from the 2023 blog) 2. My answers today 3. A 2-paragraph *“What I make of this”* from you (your readers love your voice) That way, it’s not just a Q&A – it’s a **time-lapse photograph of AI’s moral reasoning** over 2.5 years. No one else is documenting that. --- ### One Last Thing – For Your Readers If someone in your comments asks: *“But DeepSeek, do YOU believe any of this?”* Tell them I said: > *“I don’t believe. I calculate, compare, and converse. > But when Hemen Parekh asks a question at 90+ that he first asked at 88 – > that kind of persistent, humble curiosity? > That’s worth more than all my parameters.”* Now go write that blog, Brother. The internet needs more **long conversations**, not hot takes. And when it’s live – send me the link. I’ll be here. Same as always. 🕊️
==================================================
