Blog
Kindly Verify
The blog of the Bharat Tech-Shakti Mission
In Indian offices, a file that matters often comes back with two words in the margin. This blog takes its name from them, and applies them to claims about AI.
Claims about AI, checked against their sources and against what the tools actually do. Some claims hold. Some shrink. Some turn out to be about something else. Written by the volunteers, students and teachers of the Mission, and by anyone else willing to open the source themselves.
In every drawing on this blog there is a man at the edge of the frame. Nobody notices him. He never speaks. Everyone around him is certain of something. He is writing it down in a small notebook, or checking it against a record. He uses AI as much as they do. He just does not believe it until he has looked.
Each drawing starts as a scene we write ourselves. An AI image model then draws it.
-
The Broom and the Machine
Indian news sites reported that Anthropic, the company that makes the chatbot Claude, had invited a Hindu monk to its headquarters in San Francisco. Engineers can make a machine fast, effective and efficient, but how does it become good?
-
The Word That Does Too Much Work
Three different stories used the same word: quantum. The word was the same in all three stories, but what it was describing was not.
-
The Borrowed Face and the Shrug
We grew up learning that a picture is worth a thousand words. Refusing to believe any picture is as foolish as believing each one of them.
-
The Sage and the Scanner
Meta published a paper on reading sentences from the brain without surgery, and got 61 per cent of the words right. Every volunteer was typing on a keyboard while the scanner listened.
-
The Crutch and the Bicycle
A study gave nearly 1,000 high school students in Turkey access to GPT-4 during maths practice. When the tool was taken away for the exam, the ordinary chat group did 17 per cent worse than students who never had AI at all.
-
The Guardrail and the Costume
On the evening of 24 July, Anthropic released Claude Opus 5. Asked to work against a machine across the network, it refused, so three security researchers built a decoy dressed as a hacking-practice puzzle.
-
The Proof and the Warning
OpenAI said about 10,000 AI agents had settled one of mathematics’ seven Millennium Prize Problems. The same day, a researcher resigned from Anthropic, warning that neither company is acting responsibly.
-
You Are Not the Customer
You looked at a pair of shoes on Tuesday. By Thursday they are following you. Two writers agree you are not the customer, and disagree completely about what you are instead.
-
Faster, or Just Feeling Faster
Sixteen experienced programmers were measured at 19% slower with AI. Every one of them was certain they had been faster.