The Prompt: Cursor's Customer Support Bot Made Up A Policy

Forbes

22-04-2025

Welcome back to The Prompt,
Buzzy AI coding tool Cursor's customer support bot replied to a programmer's query with a made up policy that doesn't exist. The snafu led to some users complaints.
Getty
AI coding software is all the rage. One particularly popular tool is Cursor, built by nascent AI startup Anysphere, which has become one of the fastest growing startups of all time. But even Cursor isn't immune to hallucinations— AI's tendency to make false statements and assert them as fact. A few days ago, a developer noticed that switching from one device to another kicked them out of Cursor's coding platform. When the programmer reached out to the company's customer support, an agent called 'Sam' responded to them that this was expected behavior under a new policy change.
Except there was no policy change. The response was made up by Cursor's AI support bot. CEO Michael Turell confirmed this on Reddit after the AI's response circulated on social media and elicited complaints from several programmers who often use multiple machines while coding. Some users commented that they were cancelling their subscriptions as a result. Truell apologized on Hacker News and said the user got a complete refund and that the company now ensures that AI responses are clearly labelled as such.
Now let's get into the headlines.
OpenAI released a set of new AI models called o3 and o4-mini that can understand images, search the web and analyze data from uploaded documents. The company claims the models, which can crop, rotate and zoom into photos, are its most powerful 'reasoning' models. Unlike previous iterations, the models consume more time (and compute) to process a query and are trained to figure out how to approach responding to an input. They also have access to external tools allowing them to independently carry out tasks on your behalf.
People have since uploaded pictures of random places to ChatGPT and used it to figure out its exact location along with the GPS coordinates and directions to get there. Many have asked the model to respond as if it's playing GeoGuessr, a popular game where players must find locations based on Google Street View images.
Meanwhile, yet another social media trend has led to people using ChatGPT to create AI-generated action figures (with accessories) and Barbie dolls featuring their likeness. The ChatGPT-generated Barbies have various professions like realtor, tech troubleshooter and business coach. Artists have pushed back on the trend by hand drawing their own versions, the New York Times reported.
Ahead of the Canadian election, Amazon has been flooded with AI-generated books about Prime Minister Mark Carney and other political leaders, Bloomberg reported. Some of them contained incorrect information about the figures. The e-commerce giant took down most of the political books flagged by Bloomberg. This isn't an isolated incident: people have published hundreds of titles, written with the help of generative AI tools, on Amazon's Kindle Direct Publishing.
Anthropic has made its first investment into an AI company called Goodfire, a San Francisco-based startup that has raised $50 million in a Series A round led by Menlo Ventures with participation from Lightspeed Ventures and others. The company is building a system to address the black box problem of AI— a gaping hole in our understanding of why AI models work in a certain way. The company aims to understand how neural networks function to unlock new ways to train AI models and improve their performance.
Before creating Cursor, the wildly popular AI coding tool, Michael Truell was an 18-year-old MIT student, growing restless during his first taste of corporate life as an intern at Google. It was 2019, the summer after his freshman year, and Truell was sitting at the cafe of the Computer History Museum in Mountain View, California, with Ali Partovi, who was recruiting for his vaunted Neo Scholars program, aimed at finding tech's future crème de la crème while they're still in college.
Partovi, an early investor in Facebook, Airbnb and Dropbox, chatted with Truell about AI research and startup life. Then he gave the student a handwritten coding test — an easy task for Truell, who finished it in record time. After the meeting, Partovi got out his dossier and scrawled a circle with a star inside of it next to Truell's picture: the investor's shorthand indicating he was so impressed that he'd invest in any project Truell pursued.
Truell would become a Neo Scholar, and three years later, Partovi would become the first-ever investor in his startup Anysphere, maker of Cursor. The company, which reportedly took just a year to reach $100 million in annualized revenue (and has since then doubled it), is among the fastest growing startups of all time.
'Ali basically backed us when it was just the idea. We hadn't really done anything else yet,' Truell told Forbes. 'He's not afraid to spend time with and really bet on people before they're very proven.'
Neo was founded in 2017 with its scholars program, which selects 30 members each year, giving them access to workshops, networking events and recruitment connections at big tech companies. Similar to the Thiel Fellowship, the program founded by billionaire investor Peter Thiel that doles out $100,000 grants to entrepreneurs to forgo college, Neo Scholars can also elect to take time off school to work on a project. But Neo's version, which comes with a $20,000 grant, is optional and only lasts one semester — designed to be a much less intense commitment than skipping school entirely.
Read the full story on Neo on Forbes.
In February, Perplexity CEO Aravind Srinivas said his AI search company was skipping spending millions on a Super Bowl commercial and instead would give away $1 million to one lucky user who downloaded the app, asked five questions on it and referred it to their friends. But the company's Android app has 10 major security vulnerabilities leading to impersonation attempts, account takeovers and data theft, according to a report by app testing company Appknox, Forbes reported. These gaps would allow anyone to make a clone of the Perplexity app and collect private user data. The findings are reported as Perplexity is in talks to integrate its AI assistant into smartphones manufactured by Lenovo and Samsung, Bloomberg reported.
In the real world race between humans and AI, robots still have a long way to go. 21 humanoid robots scuttled alongside 12,000 humans in a half marathon held in Beijing. But only six crossed the finish line, Wired reported, with the rest failing for a variety of reasons including overheating issues.The fastest robot completed the race in 2 hours and 40 minutes, but only after falling down once and having its batteries swapped out thrice. It's still baby steps out there for robots.