iiinova — where your digital self keeps evolving

Sayash Kapoor @sayash.bsky.social · 07/07/2026

And a few examples of my public writing on AI's impacts on institutions and professions: - Legal services: www.normaltech.ai/p/ai-wont-au... - Software engineering: www.normaltech.ai/p/why-ai-has... - Science: www.normaltech.ai/p/could-ai-s...

normaltech.ai

AI Won’t Automatically Make Legal Services Cheaper

Applying the AI as Normal Technology framework to legal services

090

Sayash Kapoor @sayash.bsky.social · 07/07/2026

I will continue working on frontier AI evaluation, AI policy, and AI's impacts on institutions such as science, law, and medicine. Examples of my past research: - Frontier AI evaluations: cruxevals.com - AI policy: www.science.org/doi/10.1126/... - AI for science: doi.org/10.1016/j.pa...

1110

Sayash Kapoor @sayash.bsky.social · 07/07/2026

I'm especially keen to hire students interested not just in high-quality research, but also in communicating it with a broad audience such as through public writing and policy impact. If you're interested, fill this form to make sure I don't miss your application: forms.gle/fuXw2ELjhoZp...

140

Sayash Kapoor @sayash.bsky.social · 07/07/2026

Thrilled to share that I am joining UC Berkeley as an Assistant Professor in the School of Information! I start in Fall 2027, and I am recruiting PhD students this cycle. List me in your application if you want to work with me! More on what I'm looking for (and a form to indicate interest) 🧵

5637

Reposted by Sayash Kapoor

Melanie Mitchell @melaniemitchell.bsky.social · 23/05/2026

Another excellent post (from @sayash.bsky.social and @randomwalker.bsky.social) on questions to ask about announcements of AI capabilities. I highly recommend their newsletter! www.normaltech.ai/p/did-google...

normaltech.ai

Did Google’s AI agents really build an operating system for $916?

The importance of independent evaluation

1328

Reposted by Sayash Kapoor

Katy Glenn Bass @katygb.bsky.social · 21/05/2026

New from @randomwalker.bsky.social and @sayash.bsky.social: Responding to AI risks requires building broad resilience to harms through normal policymaking, rather than allowing the executive branch to exert extraordinary power over AI access and development. knightcolumbia.org/blog/do-ai-r...

knightcolumbia.org

Do AI Risks Require Extraordinary Government Intervention?

0123

Reposted by Sayash Kapoor

Jameel Jaffer @jameeljaffer.bsky.social · 21/05/2026

New Trump EO is expected to "ask" AI cos to submit frontier models for agency review in advance of release. Here's a new essay--just published--by @randomwalker.bsky.social & @sayash.bsky.social about whether AI risks require extraordinary government intervention. knightcolumbia.org/blog/do-ai-r...

knightcolumbia.org

Do AI Risks Require Extraordinary Government Intervention?

062

Reposted by Sayash Kapoor

Ayla Taylor @aylataylor.bsky.social · 03/03/2026

A silly little comic based on the opening section of AI Snake Oil by @randomwalker.bsky.social and @sayash.bsky.social.

Panel 1: Text: “Imagine an alternate universe in which people don’t have words for different forms of transportation, only the collective noun ‘vehicle’.” Illustration: a stick figure stands next to a much more detailed motorcycle, which a speech bubble saying “Woah! Sweet vehicle!”
Panel 2: Text: “They use that word to refer to: cars, buses, bikes, spacecraft, and all other ways of getting from place A to place B.” Illustration: A car, a school bus, a bicycle, and a space shuttle, all with a stamp that says “vehicle” on them.

Panel 1: Text: “Conversations in this world are confusing.” Illustration: A speech bubble coming from the left: “Can you drive a vehicle?”, A speech bubble coming from the right: “Definitely!”, A illustration of a car crashed into a tree. Left speech bubble: “I thought you said you could drive!” Right speech bubble: “I can! I’m just used to ones with two wheels!”
Panel 2: Text: “There are furious debates about whether or not vehicles are environmentally friendly… even though no one realizes that one side is talking about bikes and the other is talking about trucks. Illustration: Left speech bubble: “Vehicles produce so much pollution!” Right speech bubble: “That’s an exaggeration! They are actually very green!”

Panel 1: Text: “There is a breakthrough in rocketry, but the media focuses on how vehicles have gotten faster so people call there car (‘car’ is crossed out and replaced with ‘vehicle’) dealer to ask when faster models will be available.” Illustration: A TV news report with a picture of a rocket ship and a chyron saying “Breaking: Vehicles reach 1000 mph!”. Below that is a drawing of two stick figures talking at a car dealership. One says, “So I can take this to space, right?”
Panel 2: Text: “Meanwhile, fraudsters have capitalize on the fact that consumers don’t know what to believe when it comes to vehicle technology, so scams are rampant in the vehicle sector.” Illustration: A stick figure with a mean smile and a sparkle next to his eye pats a car that has plane wings taped to it. A speech bubble says, “Oh yeah! You can fly this baby across the ocean!”

Panel 1: Text: “Now replace the word “vehicle” with “artificial intelligence” and we have a pretty good descriptor of the world we live in.” Illustration: One crowd of people say “AI is bad for the environment!” Underneath them is a large box that is labeled “Size of AI people are concerned about”, another crowd of people says “AI is used for climate research!” Underneath them is a much smaller box saying “Size of AI used for climate research”. In the foreground there is a person watching the debate with several question marks above it.
Panel 2: Credits. “Text from AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference by Arvin’s Narayanan and Sayash Kapoor. Art by Ayla Taylor. www.aylataylor.com”

13114

Reposted by Sayash Kapoor

Lawfare @lawfaremedia.org · 12/02/2026

Advanced AI will not automatically help consumers achieve desired legal outcomes at lower costs. Justin Curl, @sayash.bsky.social, & @randomwalker.bsky.social identify three bottlenecks that stand between AI capability advances and more accessible legal services in the latest Lawfare Research paper.

lawfaremedia.org

AI Won’t Automatically Make Legal Services Cheaper

Three bottlenecks between AI capability and access to justice.

1134

Reposted by Sayash Kapoor

Jason W. Burton @jasonburton.bsky.social · 19/05/2025

(1/4) Ever wondered what tech policy might look like if it were informed by research on collective intelligence and complex systems? 🧠🧑‍💻 Join @jbakcoleman.bsky.social, @lukethorburn.com, and myself in San Diego on Aug 4th for the Collective Intelligence x Tech Policy workshop at @acmci.bsky.social!

11512

Reposted by Sayash Kapoor

Knight First Amendment Institute @knightcolumbia.org · 15/04/2025

In a new essay from our "Artificial Intelligence and Democratic Freedoms" series, @randomwalker.bsky.social & @sayash.bsky.social make the case for thinking of #AI as normal technology, instead of superintelligence. Read here: knightcolumbia.org/content/ai-a...

knightcolumbia.org

AI as Normal Technology

13817

Reposted by Sayash Kapoor

Princeton Laboratory for Artificial Intelligence @princetonainews.bsky.social · 16/04/2025

“The rush to adopt AI has consequences. As its use proliferates…some degree of caution and introspection is warranted.” In a comment for @nature.com, @randomwalker.bsky.social and @sayash.bsky.social warn against an overreliance on AI-driven modeling in science: bit.ly/4icM0hp

bit.ly

Why an overreliance on AI-driven modelling is bad for science

Without clear protocols to catch errors, artificial intelligence’s growing role in science could do more harm than good.

064

Reposted by Sayash Kapoor

Dr Abeba Birhane @abeba.blacksky.app · 09/04/2025

Science is not collection of findings. Progress happens through theories.As we move from findings to theories things r less amenable to automation. Proliferation of scientific findings based on AI hasn't accelerated—& might even have inhibited—higher levels of progress www.nature.com/articles/d41...

nature.com

Why an overreliance on AI-driven modelling is bad for science

Without clear protocols to catch errors, artificial intelligence’s growing role in science could do more harm than good.

312049

Sayash Kapoor @sayash.bsky.social · 03/02/2025

This is the specific use case I have in mind (Operator shouldn't be the *only* thing developers use, but rather that it can be a helpful addition to a suite of tools): x.com/random_walke...

x.com

020

Sayash Kapoor @sayash.bsky.social · 03/02/2025

It is also better for end users. As @randomwalker.bsky.social and I have argued, focusing on products (rather than just models) means companies must understand user demand and build tools people want. It leads to more applications that people can productively use: www.aisnakeoil.com/p/ai-compani...

aisnakeoil.com

AI companies are pivoting from creating gods to building products. Good.

Turning models into products runs into five challenges

030

Sayash Kapoor @sayash.bsky.social · 03/02/2025

Finally, the new product launches from OpenAI (Operator, Search, Computer use, Deep research) show that it doesn't just want to be in the business of creating more powerful AI — it also wants a piece of the product pie. This is a smart move as models become commoditized.

120

Sayash Kapoor @sayash.bsky.social · 03/02/2025

This also highlights the need for agent interoperability: who would want to teach a new agent 100s of tasks from scratch? If web agents become widespread, preventing agent lock-in will be crucial. (I'm working on fleshing out this argument with @sethlazar.org + Noam Kolt)

110

Sayash Kapoor @sayash.bsky.social · 03/02/2025

Seen this way, Operator is a *tool* to easily create new web automation using natural language. It could expand the web automation that businesses already use, making it easier to create new ones. So it is quite surprising that Operator isn't available on ChatGPT Teams yet.

100

Sayash Kapoor @sayash.bsky.social · 03/02/2025

Instead of thinking of Operator as a "universal assistant" that completes all tasks, it is better to think of it as a task template tool that automates specific tasks (for now). Once a human has overseen a task a few times, we can estimate Operator's ability to automate it.

OpenAI allows you to delegate daily tasks to Operator

110

Sayash Kapoor @sayash.bsky.social · 03/02/2025

OpenAI also allows you to "Save" tasks you completed using Operator. Once you complete a task and provide feedback to complete it successfully, you don't need to repeat it the next time. I can imagine this becoming powerful (though it's not very detailed right now).

screenshot of the save task template for Operator

110

Sayash Kapoor @sayash.bsky.social · 03/02/2025

3) In many cases, the challenge isn't Operator's ability to complete a task, it is eliciting human preferences. Chatbots aren't a great form factor for that. But there are many tasks where reliability isn't important. This is where today's agents shine. For example: x.com/random_walke...

x.com

141

Sayash Kapoor @sayash.bsky.social · 03/02/2025

Could more training data lead to automation without human oversight? Not quite: 1) Prompt injection remains a pitfall for web agents. Anyone who sends you an email can control your agent. 2) Low reliability means agents fail on edge cases

131

Sayash Kapoor @sayash.bsky.social · 03/02/2025

But being able to see agent actions and give feedback with a human in the loop converts Operator from an unreliable agent, like the Humane Pin or Rabbit R1, to a workable but imperfect product. Operator is as much as UX advance as it is a tech advance.

130

Sayash Kapoor @sayash.bsky.social · 03/02/2025

In the end, Operator struggled to file my expense reports even after an hour of trying and prompting. Then I took over, and my reports were filed 5 minutes later. This is the bind for web agents today: not reliable enough to be automatable, not quick enough to save time.

141

Sayash Kapoor @sayash.bsky.social · 03/02/2025

OpenAI also trained Operator to ask the user for feedback before taking consequential actions, though I am not sure how robust this is — a simple instruction to avoid asking the user changed its behavior, and I can easily imagine this being exploited by prompt injection attacks.

110

Sayash Kapoor @sayash.bsky.social · 03/02/2025

But things went south quickly. It couldn't match the receipts to the amounts. Even after prompts directing it to missing receipts, it couldn't download them. It almost deleted previous receipts from other expenses!

110

Sayash Kapoor @sayash.bsky.social · 03/02/2025

It navigated to the correct URLs, asked me to log into my OpenAI and Concur accounts. Once in my accounts, it downloaded receipts from the correct URL, and even started uploading the receipts under the right headings!

screenshot of concur with the categories for the expense filled in

110

Sayash Kapoor @sayash.bsky.social · 03/02/2025

I asked Operator to file reports for my OpenAI and Anthropic API expenses for the last month. This is a task I do manually each month, so I knew exactly what it would need to do. To my surprise, Operator got the first few steps exactly right:

screenshot of a conversation with Operator

120

Sayash Kapoor @sayash.bsky.social · 03/02/2025

OpenAI's Operator is a web agent that can solve arbitrary tasks on the internet *with human supervision*. It runs on a virtual machine (*not* your computer). Users can see what the agent is doing on the browser in real-time. It is available to ChatGPT Pro subscribers.

screenshot of Operator writing "Hello World" in an online notepad.

170

Sayash Kapoor @sayash.bsky.social · 03/02/2025

I spent a few hours with OpenAI's Operator automating expense reports. Most corporate jobs require filing expenses, so Operator could save *millions* of person-hours every year if it gets this right. Some insights on what worked, what broke, and why this matters for the future of agents 🧵

Graph of web tasks along difficulty and severity (cost of errors)

63910

Reposted by Sayash Kapoor

Melanie Mitchell @melaniemitchell.bsky.social · 19/12/2024

Excellent post discussing whether "AI progress is slowing down". www.aisnakeoil.com/p/is-ai-prog... And if you're not subscribed to @randomwalker.bsky.social and @sayash.bsky.social 's great newsletter, what are you waiting for?

aisnakeoil.com

Is AI progress slowing down?

Making sense of recent technology trends and claims

05615

Reposted by Sayash Kapoor

Arvind Narayanan @randomwalker.bsky.social · 18/12/2024

Excited to share that AI Snake Oil is one of Nature's 10 best books of 2024! www.nature.com/articles/d41... The whole first chapter is available online: press.princeton.edu/books/hardco... We hope you find it useful.

513930

Sayash Kapoor @sayash.bsky.social · 16/12/2024

Improving the information environment is inextricably linked to the larger project of shoring up democracy and its institutions. No quick fix can “solve” our information problems. But we should reject the simplistic temptation to blame AI.

182

Sayash Kapoor @sayash.bsky.social · 16/12/2024

But blaming technology is not a fix. Political polarization has led to greater mistrust of the media. People prefer sources that confirm their worldview and are less skeptical about content that fits their worldview. Journalism revenues have fallen drastically.

140

Sayash Kapoor @sayash.bsky.social · 16/12/2024

So why do we keep hearing warnings about an AI-fueled misinformation apocalypse? Blaming technology is appealing since it makes solutions seem simple. If only we could roll back harmful tech, we could drastically improve the information environment!

1102

Sayash Kapoor @sayash.bsky.social · 16/12/2024

We've heard warnings about new tech leading to waves of misinfo before. GPT-2 in 2019, LLaMA in 2023, Pixel 9 this year, and even photo editing and re-touching back in 1912. None of the predicted waves of misinfo materialized.

195

Sayash Kapoor @sayash.bsky.social · 16/12/2024

Similar trends were seen worldwide. In India, AI was used for trolling rather than misinformation. In Indonesia, AI was used to create cartoon avatars that softened a candidate's image. Of course, the cost of creating avatars without AI is minuscule for presidential campaigns.

Screenshot from Rest Of World article about the impact of AI in the 2024 Indian elections: https://restofworld.org/2024/exporter-india-deepfake-trolls/

Screenshot from CIGI article about the impact of AI in the 2024 Indonesia elections: https://www.cigionline.org/articles/its-time-to-reframe-disinformation-indonesias-elections-show-why/

150

Sayash Kapoor @sayash.bsky.social · 16/12/2024

But research shows that people actively seek content that confirms their beliefs. This also explains why cheap fakes are effective: It is much easier to convince someone of misinformation if they already agree with its message.

1111

Sayash Kapoor @sayash.bsky.social · 16/12/2024

So why the focus on AI? Our hypothesis is that analyses of misinfo focus mainly on the supply of misinfo, while largely ignoring the demand. AI does make content creation cheaper, but that doesn't matter much - distribution and attention have always been the limiting factors.

1102

Sayash Kapoor @sayash.bsky.social · 16/12/2024

In fact, the entire WIRED database of AI use had only 78 instances of AI use in elections worldwide. Traditional "cheap fakes" - like edited photos or slowed-down videos - were far more prevalent. In US elections, cheap fakes appeared seven times more often than AI-generated content.

1112

Sayash Kapoor @sayash.bsky.social · 16/12/2024

When AI was used deceptively, the content could have been created without AI for just a few hundred dollars, using video editing or Photoshop. The technology isn't the bottleneck - anyone determined to create false content already could.

1102

Sayash Kapoor @sayash.bsky.social · 16/12/2024

There were even examples of AI use that could improve the information environment, such as privacy for journalists concerned about government action, education about AI, and using AI voice cloning to greet voters despite having laryngitis.

161

Sayash Kapoor @sayash.bsky.social · 16/12/2024

We were surprised to find that half of all AI use in elections wasn't deceptive. Political campaigns used AI transparently, mainly to improve their outreach, without attempting to spread false information. In 19 of the 22 cases of campaigning, there was no deceptive intent.

1101

Sayash Kapoor @sayash.bsky.social · 16/12/2024

We found that (1) half of AI use wasn't deceptive, (2) deceptive content was nevertheless cheap to replicate without AI, and (3) focusing on the demand for misinfo rather than the supply can be more effective. Link to the essay: knightcolumbia.org/blog/we-look...

knightcolumbia.org

We Looked at 78 Election Deepfakes. Political Misinformation Is Not an AI Problem.

1328

Sayash Kapoor @sayash.bsky.social · 16/12/2024

More than 60 countries held elections this year. Many researchers and journalists claimed AI misinformation would destabilize democracies. What impact did AI really have? We analyzed every instance of political AI use this year collected by WIRED. New essay w/@random_walker: 🧵

List of news headlines about AI misinformation

38336

Reposted by Sayash Kapoor

Arvind Narayanan @randomwalker.bsky.social · 15/12/2024

My wife noticed an article with my name on it in her feed and said “Did you write this? It doesn’t sound like you.” I was surprised to see an alarmist headline attributed to me and @sayash.bsky.social by Wired: “Human Misuse Will Make Artificial Intelligence More Dangerous”

14812

Reposted by Sayash Kapoor

Stefano Maffulli @maffulli.net · 15/12/2024

Misinformation is not an AI problem. @sayash.bsky.social and @randomwalker.bsky.social highlight that "while generative AI reduces the cost of creating misinformation, it does not reduce the cost of distributing it". Always worth reading them. substack.com/app-link/pos...

substack.com

We Looked at 78 Election Deepfakes. Political Misinformation is not an AI Problem.

Technology Isn’t the Problem—or the Solution.

094

Reposted by Sayash Kapoor

Jameel Jaffer @jameeljaffer.bsky.social · 13/12/2024

Is artificial intelligence supercharging the problem of political misinformation? In a word: no. New from @randomwalker.bsky.social & @sayash.bsky.social at @knightcolumbia.org. knightcolumbia.org/blog/we-look...

knightcolumbia.org

We Looked at 78 Election Deepfakes. Political Misinformation Is Not an AI Problem.

4248