GPT-4 Wins Ten Times More Bill Fights: DoNotPay's Robot-Lawyer War

Nathan Labenz introduces Joshua Browder as founder and CEO of DoNotPay, billed as the world's first robot lawyer. The BBC has called him the Robin Hood of the internet. The company has been fighting consumer disputes for years and jumped on large language models early, on the theory that nobody has time to sit on hold, fill forms, or haggle for refunds. Browder's opening claim is simple: if big companies are already using AI, consumers should have it too. The preview clip is the part that sounds like science fiction. A utility said it would not accept AI-filed disputes. DoNotPay's bot answered, "I'm not an AI, I'm just a consumer that knows my rights." Browder calls that an arms race.

On a live Comcast negotiation, a GPT-4 bot enters the company's chat and bargains the bill down. They have been doing the same loop since GPT-3. The difference is temperament. GPT-3 would take a modest offer — twenty dollars off after an apology about overage charges — and say thank you. GPT-4 says that is not enough and asks for more. Browder says they have gotten a few hundred dollars back that way, and he puts the improvement in success rate from GPT-3 to GPT-4 at about ten times. The Selenium robot that drives the browser got the same upgrade: it used to break when Comcast moved a DOM element; with AI in the loop, small layout changes no longer kill the run.

Eight years in, the team is seven people and, Browder says, saves users on the order of mid-nine figures a year. The product now covers more than 200 use cases, from refunds and subscription cancellations to bill negotiations and bank fees. He is not trying to become a general "do anything" agent. Ordinary people in Ohio, he argues, will not install a Chrome extension, log into five sites, and babysit a copilot — and if they did, they might as well chat with Comcast themselves. The point is that they do not have time. The work should happen in the background, on the phone or in the operating system or in whatever Apple glasses come next, idiot-proof enough that the user never has to watch the process.

Browder started DoNotPay in 2015 after moving from England to Stanford and collecting a pile of parking tickets. He offers the left-side-of-the-road excuse, then admits he was a terrible driver. The useful discovery was that knowing the right things to say saved real money. He built the first version as templates for friends and was stunned when it scaled to hundreds of thousands of ticket appeals. That was the moment he realized the idea was bigger than parking and should cover consumer rights in general.

The early product was a rules engine, and rules engines work in a narrow band because the law is formulaic. Pick a defense — the signage was wrong — and the system pulled San Francisco's signage statute, dropped it into a letter, and mailed it. In-flight Wi-Fi dead? A ferocious template letter to United was scary enough that the airline often just refunded. Templates rot. Submit ten thousand near-copies to one office and they start ignoring you. Randomizing the wording still looked like a template. Users also picked the wrong one. Language models washed the whole Rube Goldberg stack away: people describe the situation in natural language, and the backend matches that to what should actually be done.

Legal rights are the moat he chose on purpose. Paribus, an earlier company founded by the people who later started Ramp, automated Chase credit-card price protection: log in, scrape purchases, file when Amazon dropped a price. It worked too well. Chase had offered the perk on the assumption almost nobody would use it; once AI made it automatic, Chase killed price protection across the cards and Paribus had to sell. Statute does not vanish because a company is annoyed. The Fair Credit Reporting Act says a credit dispute that is signed, includes an address, and is correctly formatted must be accepted. The remaining problem is gathering that data with as little friction as possible — email scrape, public records — so users do not even hand over the last four of a Social. In the Comcast case the same trick applies: the bot rewrites a complaint into broad FTC language about misleading practices. Some agent, human or AI, gets the letter and hits refund rather than fight.

The injustice that angers Browder most is banking. Wire a transfer, cash a check, keep too small a balance, and the fees appear; large institutions pay ordinary people almost no interest and charge them heavily at the same time. He wants users to link an account so the AI can hunt the places they are being ripped off. Life, he agrees with Labenz, is paradoxically more expensive when you are poor, because the petty injustices are the ones nobody has time to fight.

He has a name for the pattern: concentrated benefit, spread-out harm. Wells Fargo can charge a million people ten dollars and make ten million dollars. Each victim cannot afford to fight, because ten dollars is too small and everyone is busy. It is the pickpocket who takes ten dollars from ten thousand pockets; nobody calls the police. DoNotPay's two hundred use cases — refunds, cancellations, bill cuts, bank fees — are aimed at that math. The product is also shifting from reactive to proactive. Today a user still uploads a bill. The picture he wants is waking up to a message that the AI noticed an overcharge and already saved fifty dollars.

Labenz presses the ethics of a bot that will not disclose itself to Comcast and will even lie if asked. Browder's defense is that these fights are simple, time-saving, and positive-sum. When DoNotPay launched city by city, NPR asked Los Angeles parking officials what they thought. The head of the parking-ticket division said, in what Browder calls great spin, that they actually liked it: people write gibberish, and a templated appeal is standardized and easier to process. Big companies already run AI on their side and do not announce it. Giving consumers the same tool, he argues, only levels the table. California class actions claiming secret chatbots are reading support messages will test how far that theory goes.

Labenz floats Yuval Noah Harari's candidate first law: AI must disclose that it is AI. Browder is against a universal version. iMessage autocomplete is already AI; should every sentence carry a bracketed disclaimer? AI will be in every corner of life, he says, and a general disclosure duty would only hold the technology back. If there must be a rule, limit it to consumer-facing cases and leave business-to-business communication alone. Regulation, he has learned the hard way, is cheap for Apple, Google, and OpenAI and lethal for small companies. Criminals will not comply anyway. The New York lawyer who let ChatGPT invent citations and filed them in federal court is the cautionary tale for serious work: he may lose his license or worse. High-stakes filings need a human in the loop. Twenty-dollar utility fights do not. Middle America, Browder says, wants the bill cut, not another academic disclaimer to click past.

He is not laissez-faire about harm. Debt collectors are already using pre-recorded AI voices to jump from a thousand calls a day to a million; he would ban that. He would ban AI in sentencing. Expert systems already help write reports that put people in prison, and in London a network-wide timing model can ticket you for getting between two cities too fast — not for passing a camera. The 1991 robocall law covered recorded human voices dialed a million times; it needs an update for AI sales and collections. Political deepfakes in a candidate's voice are another case where he does want the system to say it is AI. Gatekeepers will not be enough. There are too many paths to GPT-4, open-source models are catching up, and a scammer who does not care about a commercial license can glue a Llama to a text-to-speech API.

His advice to policymakers is to regulate the extremely large businesses so they do the hard work. Robocalls got bad enough that Congress made the phone companies responsible. STIR/SHAKEN is why an iPhone shows "spam likely": AT&T has to identify the caller. Put analogous duties on Microsoft and OpenAI. Shift liability onto platforms that knew or should have known, and class-action lawyers will keep them honest. Add licensing for the largest, most sophisticated models and, he guesses, you stop about 95 percent of the bad actors. Why 95? Scammers are not very sophisticated, which is why they are scammers. Anyone who can fine-tune, host, and engineer an end-to-end stack is probably already getting paid at an AI company and does not need to scrape the bottom of the barrel. Making it harder for dumb people to scam, he says, has a surprisingly large effect. Phone companies should also build the defensive model: if the voice on the line is a fake grandparent, the handset should vibrate and warn you.

Browder's macro bet is deflation, not doom. Customer service is estimated at around 10 percent of corporate costs. That is the link between a hamburger and ChatGPT: McDonald's can take orders with AI, and United could in theory cut fares 10 percent even though airplanes have nothing to do with language models. Agent-to-agent negotiation, if it works, is cheaper for everyone. Labenz adds that a given capability level of model fell about 97 percent in price in a year. China is the contrast. Chinese customer-service AI is further along; a U.S. call still means hold music, account numbers, and an hour to get nowhere. Chinese consumers, Browder says, simply have higher standards. Americans are loud and angry and still tolerate Planet Fitness rules that require a signed letter or a store visit to cancel. California had to pass a law mandating online cancellation; a few years ago some subscriptions, including AppleCare, could be canceled only by fax or mail. Beijing's LLM anxiety, as he reads it, is about political content, not corporate efficiency.

He wants technical standards so AIs can transact, the way open banking lets software move money. An agent with power of attorney could shift deposits toward a higher rate and close a high-fee card; the bank's agent would authenticate authority over an API instead of butting heads in a chat UI built for humans. AI arbitration for security deposits — money held in escrow, evidence submitted at move-out, a model decides — is already in range of current tech, but he does not expect Comcast to volunteer for a DoNotPay seal of approval. Mattress labels that say do not tear, he notes, exist because retailers used to stuff beds with trash. Without law, companies will try anything to save money. On the consumer side of everyday life he is less panicked about personalized checkout upsells. People love ads that match what they want; the Facebook-is-listening conspiracy, in his view, is usually just a good recommender. Glasses can filter spam. Some privacy rules, he argues, are actually anti-consumer.

A Gartner-style survey he cites found that 54 percent of Americans had heard of ChatGPT and only 12 percent had tried it. Early users get an unfair few years of advantage; a geriatric country still full of paper checks will not flip overnight. That is why he doubts the leaked Anthropic fundraising line that falling behind in 2025–2026 means never catching up. His job numbers are blunt: maybe 30 percent of white-collar roles gone in five years, blue-collar safer for that window, perhaps 60 percent of jobs gone in fifteen as driving, shelf-stocking, and physical robots arrive. Wendy's FreshAI, Labenz notes, was described by a company executive as better than any human order-taker, and customers would not even know they were talking to AI. New work will appear — podcast engineers did not exist twenty years ago, entertainment and OnlyFans and AR games may explode, and toilets will still need humans. Gains, Browder warns, accrue to capital: a thousand people in Menlo Park capture value while tens of millions lose jobs. He is a capitalist who still wants some redistribution, including a basic income, because AI can also cheapen lawyers, doctors, and therapists in a pay-to-play society.

TVs collapsed in price over thirty years; education, housing, medicine, and law did not, and Browder's explanation is regulation. Professionals will save themselves with licenses. A Texas federal judge has begun requiring every filing to certify either that AI was not used or that a human checked everything the model did — a reaction to the New York hallucinated-citation fiasco. The EU, he expects, may demand a license for any AI use and could make ChatGPT clumsy to operate. DoNotPay's three counters: do not compete head-on, because no lawyer gets out of bed to save someone a hundred dollars on Comcast; change the law, including an AI-assistance bill they hope New York congressman Ritchie Torres will sponsor; and occasionally throw a firebomb — the courtroom earpiece demo — then ask forgiveness so the public can see what is possible. Doctors look different. Harvard Medical School's Zak Kohane has said that human medical advice without a GPT-4-class backup will soon be substandard. Doctors have a Hippocratic oath. Lawyers, Browder says, do not.

DoNotPay is a seven-person company with hundreds of thousands of subscribers, and the OpenAI API is already one of its largest bills. Browder's cost lecture is: do not send every request to GPT-4. Simple jobs go to GPT-3.5, which he calls ten times cheaper, or to an open-source model. Open-source quality, in his snapshot, is maybe 60 percent of GPT-4 and often enough. The expensive exception was the No Surprises Act: hospitals must publish prices, but they hide in PDFs. A bot scraped them and a custom model standardized the data for about $80,000, a sum they expect to earn back by cutting medical bills. Free, ad-supported consumer toys are the ones he thinks cannot work. A browsing assistant that GPT-4-reads every terms-of-service page is an AI Honey: users will not pay, ads will not cover GPT-4, and Credit Karma-style economics break. Subscriptions and enterprise still clear. He will not pile onto Lindy AI's CEO for subsidizing tens of dollars per user per month, but he thinks the 2021 habit of spending $100 to acquire $70 of lifetime value is wrong, and you cannot run a company on the hope that models keep getting cheaper.

His favorite use case is barely about AI at all. Federal law lets you claim $1,500 per robocall, except the callers hide behind spoofed numbers. DoNotPay built a trap credit card. When the pitch comes, you give them the number; the charge declines, but the payment network still returns a business name, phone, and address — enough to sue. One user turned that into a full-time job, made tens of thousands of dollars, and put a new roof on the house. Robocallers settle more often than security-deposit fights, usually at $500 plus a confidentiality deal, because they are terrified of class actions. A solo consumer can have an edge over a big firm: once a lawsuit is filed it is public record, and the large firms are the ones the callers watch for.

The practical list from the episode is short. Do not eat the ten-dollar fee: a correctly formatted credit dispute is a legal duty on the bureau, and misleading charges sit inside FTC rules. Split the work by difficulty, the way a seven-person shop stays alive — cheap models for cheap tasks. Join the 12 percent who have actually used ChatGPT, because that gap is years of advantage. And if a robocaller tries to sell you something, finish the payment flow with a card you control, keep the merchant record, and invoice them at $1,500 a call; they will often pay to make the file go away. Browder's less obvious product note is platform-agnosticism. Apple, in his view, is monopolistic enough to shake down a dependent app, so DoNotPay wants to be everywhere, including a day-one AR feature that hovers cheaper nearby prices over supermarket shelves.

Asked whether he would take a Neuralink-class implant once a million people had shown vaccine-level safety — devices you drive with thought, that ingest and record thought — he says 100 percent, though he would wait for the later wave rather than the first million. The logic matches the rest of the hour: in an arms race, the people with the tools win and the people who fear the tools get left. His biggest hope is that AI levels the field and makes goods and services cheap. His biggest fear is the usual capture, regulatory and technical, in which the same systems become a way to repress people. Labenz's close is that midterm plans may be useless and planning is still essential. Browder has been at it eight years, and he is tired of funds renaming themselves from crypto to AI. The contest, on his telling, is already running.