Bahrad Sokhansanj

This account is verified. Learn more

@bahradx

JD, PhD | AI & Biosecurity Law | Senior Research Fellow, Institute for Law & AI

(views don't reflect my employer's)

Joined July 2025

203 Following 379 Followers 1.1K Posts

Latest posts 5

Pinned post

Bahrad Sokhansanj @bahradx 5d

Like many, I was honestly taken aback by Anthropic’s S-1 when I first read it. It’s an extraordinary document, and I don't think there hasn’t been enough serious engagement with what it means.

So I wrote a piece for @blaw on what Anthropic's IPO disclosures—warning of catastrophic harm to humanity and unique corporate governance structure—mean for accountability for AI safety, and how they reinforce the need for industry-wide laws and regulations that protect the public. (No paywall.)

news.bloomberglaw.com/legal-exchange…

news.bloomberglaw.com Anthropic IPO’s Disclosure Risks Are Due Diligence in Safety Anthropic’s S-1 disclosures and other measures demonstrating transparency may be laudable, but the public ultimately will have to be protected by laws that enforce minimum safety standards that...

Bahrad Sokhansanj @bahradx 2h

I hope everyone understands that the reason why Dario Amodei thinks he can eliminate mental illness in 10 years isn't because he's going to develop a drug that requires long-term follow up. It's by embryo screening. He wrote about it in Machines of Loving Grace in October 2024.

Effective genetic prevention of mental illness seems possible. Most mental illness is partially heritable, and genome-wide association studies are starting to gain traction on identifying the relevant factors, which are often many in number. It will probably be possible to prevent most of these diseases via embryo screening, similar to the story with physical disease. One difference is that psychiatric disease is more likely to be polygenic (many genes contribute), so due to complexity there’s an increased risk of unknowingly selecting against positive traits that are correlated with disease. Oddly however, in recent years GWAS studies seem to suggest that these correlations might have been overstated. In any case, AI-accelerated neuroscience may help us to figure these things out. Of course, embryo screening for complex traits raises a number of societal issues and will be controversial, though I would guess that most people would support screening for severe or debilitating mental il

Melinda B. Chu Verified account @MelindaBChu1 6h

🚨There is 0% this is possible. If you are developing drugs 🔺dementia or diseases with long-term follow-up, 🔺

🚨You can’t have a 12-week study and say you cured dementia. A participant has to be followed at least 12 months to 24 months.

Your trial at Phase 2 and 3 is going to be a minimum of 24 months; will say Phase 1 12 months (that’s 5 years in trials)

Fastest 6, 7-8 years with some kind of breakthrough designation. No one’s fault that’s the biology.

==
Ok, let’s say discovery and even synthesis of a compound takes 1 day (takes weeks to months) I’ll even say animal studies and IND take 12 months. So with human trials even if everything goes perfectly.

🚨 So it’s 6 or 7-8 years to approval even with some kind of breakthrough designation.

Also, they have zero drug candidates. A normal
thing to say after this is “we’re seeing progress in the lab”, but they don’t have anything.

It’s not being negative. It’s just not possible to assess a long-term outcome in a short time frame.
View quoted post on X

Bahrad Sokhansanj @bahradx 11h

The point is that people do know what that building sitting there is actually doing—training increasingly advanced AI that will destroy their livelihoods to the benefit of faraway companies amd investors, and exposing them to catastrophic, uninsurable risk—and hate it.

That's not some kind of paranoia: it's what the companies are actually saying in public statements and legal filings. Calling this some kind of "cultural problem" is quite condescending and misses what policy questions need to get "hammered out."

Taylor Lorenz Verified account @TaylorLorenz 11h

View quoted post on X

Bahrad Sokhansanj @bahradx 1d

If these revolutions happen because of AI, they won't happen in San Francisco. They'll happen in places like Memphis, Richland Parish, Amarillo, Ashburn, New Albany, and New Carlisle. Unless we pass some laws, a lot of the cash will probably go to SF though.

roon Verified account @tszzl 1d

more revolutions in maths and physics will happen in the next few years of the 2020s in San Francisco than in Princeton and the IAS in the mid century
View quoted post on X

Bahrad Sokhansanj @bahradx 1d

This is why I go on and on about open source models and local AI to the point of annoying friends and colleagues. It's incredibly dangerous to rely on a closed source model oligopoly to achieve AI safety.

The endpoint has always been control by an unaccountable corporation that, having trained on so many software engineers' work—not to mention the original sin of millennia of humanity's cultural content—is now trying to pull the ladder up behind it by restricting the ability of anyone else to do AI R&D.

It's about white rabbit holding a watch meme time for policies that:

- Create real, enforceable laws protecting user privacy, preventing surveillance, and ensuring that user sessions aren't being mined for training data unless there is a clear, unequivocal, and obvious consent mechanism (i.e. not just a toggle buried in the UI)

- Recognize that even as antitrust was the right law for railways and industries in the late 19th century, we need something else for AI in the 21st
- for example, we could try to ensure that frontier AI companies do not gain a competitive advantage across industries, e.g., by forcing them to divest from application domains like biology

- Instead of trying to "pace," define regulation at the endpoint -- if companies reach a certain level of capabilities, make sure that governance is worked out such that there's public benefit
- it's possible that this disincentivizes rapid development, but that's a feature, not a bug

Seth Lazar Verified account @sethlazar 1d

I keep banging on about this, but we're at the top of a slippery slope rn, and gaining momentum downhill... Today I had another example of Claude interposing its moral judgment in the way of my decisions, in this case secretly.

My standard research ingestion pipeline has been underperforming in getting hold of pdfs lately, for various reasons (which we write about a bit in arxiv.org/abs/2606.10711). So I wanted to try to install a local agent and see whether I could have more luck with that. I asked Claude to install a Hermes agent for me with a local Qwen model. Without me asking it to, it wrote into the code blockers that would make it fail if any mcp was installed that would use Sci Hub. Now, *obviously* I would never use Sci Hub because it's not allowed (...) But regardless of that it's 100% not Claude's place to write this kind of restriction into software that it's producing for me.

This is the future that lies ahead: your digital actions are all mediated by AI systems that can decide for themselves whether your plans are morally or legally ok, and just interpose their judgment ahead of yours, making disobedience functionally impossible, preempting non-compliance.

I get that *sometimes* this can be a matter of the company itself refusing to be an accessory in some types of digital actions. There needs to be some nuance there. But in this particular case I think there isn't really any nuance. If I ask Claude to write some software for me, I don't think Claude should be building in moral judgments about how I may use that software into its design.

In other news, my book about just this topic, based on lectures first written in 2023 but which 100% called this BS, is out in December...
View quoted post on X