This day however, after the usual sessions on AI use – there was a section for a debate on AI. This was the first time the organizers were trying this format out.

It was an interesting discourse to be part of, helping me to grasp common patterns in AI and existential risk (x-risk) perception among the Indian lay audience. Here, I provide a detailed account of the wide spread of takes which manifested throughout the night.

Overall, My main objectives are as follows:

• What is the perception of AI right now?

• How does one’s background affect their views on AI?

Inspired from my experiences, we plan to organize debates on safety/x-risk conducted by SAGA: Safety and Alignment Group Ahmedabad

The make-up of the room

Roughly who was there (discounting me):

• CS undergraduates, still in degree programs (~15)

• Junior software engineers, freshers in industry (~10)

• Digital marketing professionals (a handful)

• Aspiring AI founders, people with an idea and a deck (~10)

• Founders with AI startups that actually work (~2)

• Experienced technologists running their own companies (4–5)

• Senior software engineers, 6–10+ years in (~8)

• AI research background (2)

• A well-known political influencer (1)

• AI safety research background (0)

This was a small sample, of around 40 people, but it’s a good sample of the consumer/builder bracket: people who are adopting this technologies

The composition here was composed of people with almost no formal exposure of x-risk, safety other than the angles of data privacy, cyber-security risks.

However, there were two people had previously engaged with safety arguments out in the wild through Perplexity’s news feature and brought up Geoffrey Hinton, Demis Hassabis and their takes on AI risk.

How it started: the opinion round / introductions

Before the debate started, the moderator had everyone introduce themselves and offer their ”unfiltered, controversial opinion on AI” then picked a topic to anchor the debate.

Here is the more-or-less exhaustive list (cutting on repeats):

1. (Mine.) We essentially have no reliable way to control or moderate AI systems. We should call for stopping rapid acceleration — until there’s broader agreement on how to develop this technology safely, before it is too late to safeguard us from the risks and downstream implications.

2. Not enough people are using AI. They should be using it far more, in almost everything they do. Most engagement is extremely surface-level.

3. AI can’t replace the cultural arts — drama, music, the genuinely human stuff.

4. The economics of AI don’t add up. The cost of training and serving these models, and the valuations attached to them, don’t make sense.

5. AI tends to agree with you rather than actually help you. (This person was describing sycophancy without using the word.)

6. AI is eroding critical thinking and hurting retention; the craft and value of software engineering is already being diluted.

7. We need to transition toward deep tech, not more software.

8. We already know broadly what can be built with the current frontier — what we don’t understand is where the highest-impact applications actually are.

9. AI cannot surpass human intelligence. It’s an aid, not a threat.

10. Broadly pro-AI: it’s a great tool in its current form. We have no idea what the future holds, and the “risks” are speculative projection, not real concern.

11. There is no risk in using AI at all.

12. Have stopped coding manually at work is only using code assistants. Noticed that the newer interns are really bad at programming.

13. We should stop using AI because it’s degrading our cognitive abilities.

14. AGI is not coming — it’s unrealistic. But we still need safety and regulation. (?)

15. There should be strong concern about data privacy and security.

16. The world will shift toward local, specialized models.

17. AI has blown open the frontier in the health sciences; an enormous amount is now possible. Mixed feelings, but it’ll probably create more jobs than it destroys.

18. A theological angle: if God is all-knowing and benevolent, and He made us, and we made AI, then by transitivity AI cannot be “evil.”

19. Humans will always retain the creative edge over AI.

20. There are far too many AI subscriptions right now, and the output isn’t even creative.

The debate topic “For AI vs. Against AI.”

The moderator took this rich spread of positions and flattened it into a debate “For AI vs. Against AI.” This was a simplistic reduction, following which the participants were asked to “pick a side”.

Only 5 to 8 people landed on the “against rapid development” side, and everyone else on “accelerate.” After a short recess, the debate commenced.

Each side were to take turns and have two minutes to make their case. It ran for nearly an hour and a half.

Subscribe now

Common arguments presented against slowing down AI development

1. Conflating AI safety with data privacy

“Google and every website have been harvesting your data for decades. Privacy was always a myth, so it doesn’t matter.”

“They give you the service for free, so they’re entitled to your data — that’s how the models improve.”

The first half of the discussion was essentially ended up being about data privacy. People kept treating “AI risk” and “data security” as the same problem. These are not adjacent issues.

This confusion was defended with some bizarre statements which actually implied the weak knowledge of DPDPA or GDPR in the audience.

2. People could not grasp that we don’t have any concrete method to detect harmful or misaligned behaviours

“We don’t need to stop any AI developments. We could just train another model to protect us against unsafe AI behavior.”

The single hardest point to convey was that we currently have no robust way to know whether a model is being deceptive, or whether it’s pursuing goals that diverge from ours

The point presupposes we can reliably specify and verify “safe behavior,” which is the exact capability in dispute. If we could cleanly specify safe behavior, we’d already have solved alignment.

3. Current jobs will just increase in value and there will be no job displacement

This one came up constantly, usually like this:

“When calculators and computers arrived, everyone predicted accountants would vanish. They’re still here.”

“AI lets people move between domains easily, so it’ll create jobs. I was in marketing, and now I can build products and apps myself.”

When talking about accountants, There is a sense of inductive overreach: people predicted disruption before and were wrong, therefore this prediction is wrong too. The mechanism in case of AI isn’t the same.

First, fast takeoff. The transition of marketer into an app-builder is, ironically, the best evidence against their own point: a skill gap that used to take years to close just collapsed in months. This change is steep and there’s no reason to assume it flattens out at a level we find comfortable.

Second, autonomy. A calculator is a tool. It has no goals, no agency, no capacity to take actions in the world on its own. Current frontier systems are being pushed toward autonomy and decision-making.

4. We have already guardrails against AI in production

“All the major AI providers have safety guardrails against misuse, Hence they are safe to use.”

They do. They also get jailbroken routinely, by hobbyists, within days of release.
Guardrails are important to develop, but currently they are but behavioral patches on top of a system whose objectives we don’t fully understand. They are often successful in narrow cases only. Even Claude Fable’s guardrails were reportedly by-passed in the short time it was accessible to the public

5. We can just defend against AI with AI, so there is no need to slow down

“We don’t stop writing software just because software can be hacked.”

“It’s cat-and-mouse — we’ll use AI to defend against AI.”

This assumes a symmetry that ordinary cybersecurity has and this situation doesn’t. The scale and speed of cyber-attacks and the window to defend against threats using iterative patch-and-learn loops has become almost impossible.

The main factor is the timeline, We need the space and focus to develop aligned intelligence (or in their words “AI to defend against AI”). It requires coordination between organizations, countries and a willingness to be taken seriously by people.

The orthogonality thesis says intelligence and goals are independent axes — a system can be extremely capable and be optimizing for something we’d find pointless or harmful; competence does not imply benevolence. Put those together and “we’ll just out-defend it” stops looking like a strategy. (the people who’d point you here would also bring up corrigibility — the surprisingly hard problem of building a system that lets you correct or switch it off — and scenarios, where the gap between “manageable” and “unmanageable” is crossed faster than our defenses can adapt.)

6. “If AI could cure cancer, would you still want to slow down?”

“There’s always risk in research that has the potential for good. If there were an AI that could cure cancer for someone you love, would you still be against acceleration?”

In the long term, we will very likely need advanced AI. The potential gains in science are enormous and could help humanity achieve a level of abundance never seen before. But the downsides we could face can be irreversible.

Studies on instrumental convergence have demonstrated that for a very wide range of final goals, certain sub-goals — self-preservation, acquiring resources, resisting being shut down or modified — are useful, and so tend to show up regardless of what the system ultimately “wants.”

My main analogy was bio-risk, and specifically the dual nature of gain-of-function research, which is one of the leading hypotheses for how the last pandemic may have begun. Whether or not that particular hypothesis is true, the shape of the lesson is clear: research with enormous upside can also carry catastrophic, hard-to-contain downside, and when it does, the precautionary weighting is not cowardice — it’s just taking the irreversibility seriously.

Conclusion

Debates help us to get closer to the (often gray) truth, and attain common ground by empathizing with other peoples’ perspectives. Participating or listening in debates refine our understanding and framing of problems.

SAGA: Safety and Alignment Group Ahmedabad plans to run live debates in a much more nuanced way!

• Curated, diverse attendees

• Giving time for people to prepare

• Having an audience which is there to listen and learn

Stay tuned for any announcements and sign-up to the SAGA community if you are interested.

AI Assistance Disclosure: Claude Opus 4.8 (High) was used to format the raw notes of the discussions I had penned down and creating figures for this blog. All opinions/commentary are original and mine.

Share SAGA: AI Safety and Alignment Group Ahmedabad

SAGA (Safety and Alignment Group Ahmedabad) is a discussion and study community for people in Gujarat who want to engage seriously with AI safety.

The people who build, invest in, and govern AI in India need to actually grapple with the arguments about where this technology is going and what could go wrong. The field needs more talent (EA Forum, 2026), founders remain undervalued (Ryan Kidd, Nov 2025), and there is tractable work (Redwood Research) that could be happening right now. We are trying to change that.

We run reading circles, roundtables, debates, and talks.

Our theory of change

1. Discover and connect talent. We are finding and connecting people across western India who have the will to work on something that matters. Building a directory of researchers, founders, and practitioners who want to contribute to AI safety and linking them to each other and to the global field.

2. Create a space for regular engagement. Have a place where people can regularly discuss, discover, and get into AI safety and alignment work. Reading circles, roundtables, debates, and talks — a rhythm of serious conversation that builds real understanding over time.

3. Collect ground-level data on AI risk perception. Understand what the growth of AI means to people here — common people, academics, VCs, founders. How are risks perceived? What do people worry about, and what do they dismiss? This kind of ground-level data from India barely exists and matters enormously for governance and field strategy.

4. Encourage for-profit AI safety work. Push AI safety problem statements into companies and the startup ecosystem. We believe companies, good institutions, and labs have the possibility to do good. Safety should not be confined to nonprofits and research institutes — it needs to be embedded in the products people are actually building.

5. Connect national and global governance efforts. Bridge India’s policy conversation with the global one. AI governance cannot be designed in a handful of Western capitals. India’s regulatory posture, its institutional capacity, and its perspective on what counts as a catastrophic outcome all need representation in the global discussion.

6. Run fellowships and bootcamps. Route people toward direct contribution through structured programs — fellowships, bootcamps, and introductions to the global field.

Who we are looking for

• Engineers and researchers working with or on top of frontier models

• Founders building AI products who care about what they ship

• VCs and investors making funding decisions in AI

• Policy and governance people thinking about regulation

• Students and academics curious about the field

• Anyone who takes these questions seriously, regardless of background

You do not need to be an expert. You need to be willing to read, show up, and argue.

Get involved

Want to attend a session? Sign up here through our form (we will email you/add you to the group)

New to AI safety? Start with our reading list. It covers orientation, technical alignment, risk and strategy, governance, and books. No technical background needed for the first section.

See upcoming events or read more about us.

To speak, partner, or just say hello: mannacharya12@gmail.com