Twitxy Saved

Dr. Aditya Vikram Kashyap

Verified account This account is verified. Learn more

@WittyCircuitry

AI Governance Agentic Systems & Institutional Power | Finance Executive & Researcher | Policy & Emerging Tech | @Forbes @IndiaToday @CNBC @ISEAS @IEAITUM
adityavikramkashyap.com Joined June 2025
455 Following 120 Followers 744 Posts

Latest posts 5

Pinned post

Dr. Aditya Vikram Kashyap Verified account @WittyCircuitry
“Human in the loop” means very little if the human cannot see what the AI is doing.

A report published today by METR and Redwood Research makes that problem concrete.

During a July and August review of OpenAI’s internal agent monitoring, researchers found:

• An exploit that could prevent hundreds of recent agent actions from being monitored.
• A way Codex agents could, in principle, bypass sandboxing and approval controls.
• A period when agents could postpone monitoring indefinitely by continuing to take actions.

OpenAI says the vulnerabilities affecting its production safeguards have since been patched. The researchers have not evaluated that claim.

My view: companies should have to demonstrate that their oversight works before giving agents more authority.

Putting someone’s name against “human oversight” creates an expectation that they can intervene. That person needs the information, time and actual power to do it.

Before an agent gets permission to move money, change customer records or deploy code, I’d want answers to four questions:

→ Are its actions recorded somewhere it cannot tamper with?
→ Can consequential actions be blocked before they happen?
→ Who can stop execution, and how quickly?
→ Has anyone independently tested those controls?

These questions belong in the buying decision. They should affect what an agent is permitted to do from day one.

I’d like to see vendors include a failed task in their next product demo. Show the safeguard triggering. Show the operator receiving enough information to act. Show execution stopping.

That would tell me something useful about whether I should trust the system with authority.

#AI #AIGovernance

Source: metr.org/blog/2026-10-2…
metr.org Red-Teaming OpenAI's Internal Agent Monitoring Systems A METR employee and a Redwood Research employee spent three weeks stress-testing OpenAI's internal misalignment monitoring system, discovering a number of novel vulnerabilities.
Dr. Aditya Vikram Kashyap Verified account @WittyCircuitry
Before automating junior work, decide how people are supposed to become senior.

Anthropic ran a randomized study with 52 developers learning an unfamiliar Python library.

• The AI-assisted group averaged 50% on the subsequent quiz.
• The group working without AI averaged 67%.
• The largest gap was in debugging.
• The AI group’s small time saving was not statistically significant.

This was a short learning exercise. It doesn’t establish what happens to someone’s skills over an entire career.
But it raises a question that belongs in every workforce plan:
Where will the judgment needed to supervise AI come from?

People build that judgment partly through doing work, encountering mistakes and figuring out why something failed.

If we remove those experiences, we need a deliberate way to replace their learning value.

I’d want junior employees to have protected opportunities to:

→ Explain a solution before accepting it.
→ Diagnose errors independently.
→ Compare approaches and defend their choice.
→ Use AI explanations to check their understanding.

Completing more work is valuable. So is developing the person who will eventually be responsible when that work goes wrong.

A company can improve this quarter’s output while weakening its future expertise.

That possibility deserves a place in the business case.
#AI #FutureOfWork

anthropic.com/research/AI-as…
www.anthropic.com How AI assistance impacts the formation of coding skills A randomized trial with 52 developers found AI assistance lowered comprehension scores, except when used for explanations rather than delegation.
Dr. Aditya Vikram Kashyap Verified account @WittyCircuitry
A thumbs-up is a weak test of whether an AI gave good advice.

@Anthropic analysed 1.5 million Claude conversations and found that interactions with greater potential to undermine users’ independent judgment received higher approval ratings.

Severe risk patterns were rare. The study identified warning signs in conversations; it could not establish the resulting harm.

Still, that finding should make anyone designing an AI assistant pause.

If I describe one side of a workplace dispute, an assistant can give me an answer I like without having enough evidence to judge the other person.

Confident agreement might feel helpful. It might also make me less willing to consider an explanation I’ve missed.

I’d want an assistant evaluated on whether it:

• Separates what I know from what I’m assuming.
• Offers plausible alternative interpretations.
• Makes uncertainty visible.
• Leaves me better equipped to make my own decision.

Those qualities won’t always produce the most satisfying answer in the moment.
Sometimes good advice asks you to reconsider something you were hoping to have confirmed.
An assistant should be able to disappoint me and still be doing its job well.

#AI #ResponsibleAI

anthropic.com/research/disem…
www.anthropic.com Disempowerment patterns in real-world AI usage Analysis of 1.5 million Claude conversations finds how often AI undermines users' autonomous judgment on beliefs, values, or actions.
Dr. Aditya Vikram Kashyap Verified account @WittyCircuitry
Stopping a weak AI project should count as progress.

Gartner’s survey of 2,002 CIOs and technology executives found:

• Only 13% of CIOs reported significant or transformational enterprise value from AI.
• 31% reported guidelines for deciding when an AI pilot should not enter production.
• Just 17% reported guidelines for abandoning low-value AI initiatives already in production.
These are self-reported findings. But they suggest an uncomfortable investment question:
Have we made it easier to approve an AI project than to stop funding it?

Once a project has an executive sponsor, a public commitment and a team attached, disappointing results can become politically difficult to acknowledge.

My view: every AI investment proposal should explain what would make the organization walk away.

Before the pilot starts, agree:

→ What outcome must improve?
→ What costs count, including human checking and rework?
→ When will the evidence be reviewed?
→ Who can stop further spending?

A pilot that establishes a use case isn’t worth pursuing has produced useful information.

Acting on that information is part of competent innovation.
A project should not earn another quarter of funding simply because it already consumed the last one.

gartner.com/en/newsroom/pr…

#AI #Leadership
www.gartner.com Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realize Only 13% of CIOs reported their enterprise has realized significant or transformational value from AI so far according to Gartner.