top of page
Search

When the Web Learned to Lie: AI Agents

A radio hoax fooled millions in 1938. Nearly ninety years later, the things being fooled aren't human — and the web is quietly feeding them poison.

Adapted from a talk by Ariel Shulman, Chief Product Officer, Bright Data

SuperAI 2026  ·  Marina Bay Sands, Singapore

In 1938, Orson Welles stepped up to a radio microphone with no television, no internet, and no special effects beyond a few sound cues — and convinced millions of listeners that Martians were invading Earth. War of the Worlds became the proof, broadcast in real time, that human beings can be made to believe something untrue simply because it arrives through a trusted channel (Figure 2).

Nearly ninety years later, that vulnerability has a new target. It is no longer only humans who can be fooled by convincing information delivered at the right moment. So can AI agents — and unlike the radio audience of 1938, an agent has no instinct that something feels wrong.

That is the unsettling thread running through the story Ariel Shulman tells. Shulman is the chief product officer of Bright Data, a company that has spent fifteen years collecting public information from the open web at industrial scale. He argues that a quiet war has broken out between two versions of the internet — the old web, built for people, and the new web, increasingly populated by software agents acting on our behalf. And in that war, the old web has learned a new and dangerous trick.*

Figure 1: Photo of Orson Welles meeting with reporters in an effort to explain that no one connected with the War of the Worlds radio broadcast had any idea the show would cause panic. ( Source: https://commons.wikimedia.org/)

A life measured in machines

To see where the new web is going, Shulman likes to start with where he came from. In 1981, when he was ten years old, his father brought home a ZX81, a small British computer with exactly one kilobyte of memory (Figure 2).

There were no disks; you saved your programs to cassette tape and could literally hear the data screech past as it loaded. He wrote little games on it, and his parents were sure they had a genius on their hands. You could, remarkably, do real things with a single kilobyte.

Then the machines kept coming. The Commodore 64. The IBM PC, with its 640 kilobytes of memory that was supposedly going to be enough forever. In those years there was barely any way to connect one computer to another - just hobbyists wiring their modems into bulletin board systems, one machine dialling directly into another. But the thirst to be connected was already there.

Connection arrived with the internet. Tim Berners-Lee more or less invented the World Wide Web single-handedly in 1989 - the URL, HTML, and much of the scaffolding the rest of us now take for granted. Before browsers existed, people searched through text-only tools with names like Gopher and Archie. Then came Mosaic, the first program to put images and text on the same page. After a kilobyte of tape, it felt like magic.

Figure 2: Photograph of the Sinclair ZX81 home computer in kit form (Source: https://commons.wikimedia.org/)

The problem was never connection. It was order.

As the web swelled, the hard question stopped being how to reach the information and became how to organise it. The first serious attempt was Yahoo, where two Stanford students hand-sorted websites into categories - about 23,000 of them, a figure that sounds almost quaint today. Others followed, including Ask.com, a search engine that has since shut down (Figure 3). By 1998, the web had entered the age of the portal: a single homepage crammed with categories, links, and advertisements that tried to show you everything at once. They look faintly embarrassing now.

Figure 3: The Yahoo homepage in 1997. (Source: VersionMuseum.com).

It is worth pausing on the era. The NASDAQ sat at around 1,043, before the dot-com crash of 2000 reset everyone's expectations. Microsoft's MSN portal shipped with Windows, so practically everyone ended up with it - galleries, endless ads, and Hotmail accounts that some readers will remember fondly.

Only one company, in Shulman's telling, truly cracked the problem of organising the web: Google. Watching screenshots of its homepage flick past across roughly twenty-nine years is a compressed history of the modern internet: AdWords, Google Images, AdSense, Gmail, Maps, the YouTube acquisition, Translate, the first Android phone, Chrome, the late and unlamented Google+, Waze, the first Pixel. Artificial intelligence slipped in first behind the scenes, then a chatbot, then Bard, then its rename to Gemini. What began as a stripped-down search box now hides staggering machinery behind a minimalist face.


The data empire - and its reckoning

Ask a room how many people used Android, Maps, YouTube, or Gmail in the last day and nearly every hand goes up. That ubiquity is the point. Google holds something close to all the data in the world: text from email and Drive, images from Photos, video and audio from YouTube, shopping signals from Chrome, local knowledge from Maps, behavioural data across devices, even health data from Android and scholarly work from Google Scholar. All of it feeds the models, Gemini among them - and the company has said it will offer AI agents for everyone.

Figure 4: The NeXT machine at CERN that ran the world’s first web server. (Source Photo: Henry Mühlpfordt Wikimedia Commons — CC BY-SA 3.0)

Such an accumulation could not go unexamined. In 2025, a major antitrust case against Google was decided. Shulman is careful to say he is not a lawyer and that what follows is his own reading of the roughly 146 pages of opinion he worked through. As he interprets it, Google must share some of its data — but only once, and it keeps the secret sauce: PageRank, the spam-detection algorithms, the parts that actually make it work.

Competitors will have to build crawlers, crawl the web, and process the data themselves. They will have to invest and innovate to proceed.

That, paraphrasing the judge, is the catch for everyone else. You cannot simply help yourself to the organised knowledge of the web. If you want it, you have to go out and collect it - which means building the crawling and processing stack yourself. The ruling did not hand competitors an easy shortcut. It handed them an instruction to do the hard work.*

The 8:47 Monday message

Here is where the abstract becomes uncomfortably concrete. Almost everyone now builds things; vibe-coding an agent over a weekend is ordinary. So picture a Monday at a hypothetical travel site that compares hotel prices, and a message lands from the CEO at 8:47 in the morning:

“Hey guys, I just built an autonomous pricing agent with Claude this weekend, in two hours. It scrapes the competitors, optimises our price, and it works perfectly. Let’s launch it!”

The CEO is not entirely wrong. The interface will work. The logic will work. On the developer's own machine, the whole thing will look flawless. The bad news is that the old web is not going to welcome an agent — least of all once it reaches production.

Figure 5: Two hours of weekend coding looks flawless on your own machine. Source (Photo: Markus Spiske / Wikimedia Commons — CC0).

The reason is foundational. The web of 1989 was built for humans, and until very recently, almost everyone visiting it was human. As the population of agents and bots grows, the web is genuinely unsure how to treat them. It is the question at the heart of Blade Runner: faced with something that behaves like a person, how do you decide whether it is one — and what do you owe it if it isn't?

Quantum mechanics, inverted

Shulman has a favourite way to describe why agents that work in testing fall apart in the wild. In quantum mechanics, he says, things behave normally at human scale and turn strange only when you zoom down to the very small. Web data collection is the mirror image. Everything behaves perfectly on your laptop, and turns strange only when you scale up into production — at which point things break in ways you often cannot even diagnose.

The web learned to lie

Some of what breaks is innocent. Websites redesign themselves, change their layouts, restructure their data. Fine. But the web also fights back on purpose. It throttles agents to slow them down. It blocks them with CAPTCHAs. And then there is the genuinely clever part — the move Shulman calls the most interesting development in the field.

For decades, a site that suspected a bot would try to stop it: throw up a CAPTCHA, deny the request, slam the door. The newer strategy does the opposite. The site quietly runs its detection algorithms, decides whether the visitor is human, and then chooses what to serve. A human gets the truth. A suspected bot gets disinformation — a honeypot. The aim is not to turn the agent away. The aim is to let it believe it succeeded, and to feed it poison.

The idea is not to stop you. It is to let you think you won — and feed you poison. In other words, the web learned how to lie.

Two examples make it vivid. A ticketing site - call it Sunny Tickets - shows a normal visitor that seats are available. The same page, requested from a suspected agent, reports that nothing is for sale. Or take a hotel: a person sees a room at $199 and plenty of availability, while an agent sees the same room listed at $399. This is garbage in, garbage out with a malicious twist - the garbage is being manufactured specifically for the machine, and bad inputs become confidently wrong decisions.

AI Agents don’t fail the way people do

This is the danger Shulman wants builders to feel in their stomachs. A human pricing analyst who sees a hotel room jump from $199 to $399 pauses. Something looks off; they double-check. That hesitation - the intuition that a number is wrong - is a safety feature humans get for free.

An agent has no such instinct. It records “received hotel price: $399,” updates five thousand entries, fires off five thousand emails notifying customers of the change, and moves on, perfectly satisfied. By the time anyone notices, the damage is done and distributed. Poisoned data does not announce itself; it simply propagates at machine speed, which is exactly what makes it so hard to catch.

Even true data goes stale

Suppose you clear every hurdle and obtain genuine, untampered data. You still face the clock. Web data has a shelf life, and different kinds of information rot at very different rates. Because Bright Data observes tens of millions of URLs continuously, Shulman says, it can watch that decay happen across markets.

Social posts, e-commerce prices, and stock availability are good for roughly a day before they mislead more than they inform. News holds for a day or two. Financial information stays useful for perhaps five to seven days. General content — blogs, evergreen pages — can stay relevant for a long time. The lesson is that freshness is not a luxury; stale data quietly corrodes the experience.

And users feel it immediately. A shopping agent that quotes $29.99 and then rings up $49.99 at checkout doesn't just annoy — it destroys trust. So does an item shown in stock that isn't, or a map confidently sending someone to a restaurant that closed an hour ago. Whether the data was poisoned or simply old, the customer's experience is the same: the system lied to them.

So what is a builder supposed to do?

Shulman's answer is, unsurprisingly, the business his company is in — but the logic stands on its own. If the new web is a crowd of agents, and the old web responds with blocks and lies, then the missing layer is trustworthy web-data infrastructure: a way to obtain data that is reliable, consistent, fresh, and scalable, delivered in whatever shape an application needs - HTML, JSON, markdown, CSV — without each developer having to win the arms race alone.

It is also why a company like his keeps colliding with the law. Bright Data has argued, in court, for the right to collect public web data — by which Shulman means anything a person could reach through an ordinary browser, with no login, no password, and no paywall. The company has been sued, including by Meta and by Elon Musk, and says it prevailed in federal court in California in those cases — rulings it spent heavily to win and that are now cited elsewhere. Its claimed scale is the kind of number that resists intuition: more than 50 billion pages scraped a day, billions parsed and archived, a network of well over a hundred million IP addresses, around 2.5 petabytes pulled in every twenty-four hours.

Three thousand times around the Earth

Humans are famously bad at feeling the weight of large numbers, so Shulman ends where he began — with the ZX81 and its single kilobyte. He asked ChatGPT to imagine converting a single day's haul, 2.5 petabytes, into physical kilobytes, each one a five-centimetre strip of that old cassette tape, laid end to end. The answer stretches almost to the Sun — enough tape to wrap around the Earth roughly 3,120 times. That is one day.

Figure 5: 2.5 petabytes a day — a single day’s tape would wrap the Earth some 3,120 times. Image: NASA, “Blue Marble” - public domain.

A modern laptop, he notes, is something like sixteen million times more capable than his childhood machine in memory and processing, and it has the whole of the world's information within reach. When he received that ZX81, he assumed the great challenge of the future would be raw computing power. He was wrong. Computing got cheap almost embarrassingly fast. The internet then gave us more information than we knew what to do with.

Now a third era has begun, and its defining problem is neither power nor abundance but trust. The agents are clever and getting cleverer. The web has learned to lie to them. Somewhere between those two facts sits the real work of the next decade: making sure that the software now acting on our behalf is fed something true — because, unlike Orson Welles's listeners in 1938, an agent will never sense that the broadcast was a hoax. It will simply believe, and act, and send the five thousand emails.

The Samso Way - Seek the Research

Here at Samso, we pride ourselves on delivering content for investors that is independent and informed by over three decades of experience in the industry. Our content is well-researched and is only created if I see merit in discussing the company's story.

Our mission is simple: cut through the noise and spotlight what matters—genuine stories, grounded insights, and real opportunity.

Our content is well-researched and is only created if the team sees merit in discussing the company or concept. Investors can explore our three core platforms: 

There may be numerous paths to success in investing, but the common thread among successful individuals is that they remain committed to making informed decisions. Equip yourself with the right knowledge and tools, and you will be well on your way to achieving your financial goals.

Most importantly, investors need to be absolutely diligent in understanding their own risk-reward tolerance and capabilities. Never bite off more than you can chew. As they say, Rome wasn’t built in a day, and the Great Wall stood because it took centuries to complete.

The Samso Philosophy:

Stay curious. Stay sharp. And remember—digging deeper always uncovers the real value.

In Life, there is no such thing as a Free Lunch.

Never bite off more than you can chew is my parting comment.

Happy Investing, and the only four-letter word you need to know is DYOR. 

To support our independent nature of our work, please head over to our Support Page and give us a helping hand in any of the ways listed. This is a new initiate for the Samso Platform, and it was always the concept of Samso when we started this journey in 2018.

Disclaimer

The information or opinions provided herein do not constitute investment advice, an offer, or solicitation to subscribe for, purchase, or sell the investment product(s) mentioned herein. It does not take into consideration, nor have any regard to your specific investment objectives, financial situation, risk profile, tax position and particular, or unique needs and constraints.

Share to Grow: Your Bonus

Samso has just released an eBook: How to Add Value to your Share Portfolio


Download eBook | Samso Insights
Download eBook

If you find this article informative and useful, please help me share the information. I try to write about topics that are interesting and have the potential to be of investment value. It is not easy to find stories that fit those parameters. If you or your organisation sees the benefit of what Samso is trying to achieve and has a need to share your journey, please contact me at noel.ong@samso.com.au.

Samso is a trusted platform that equips dedicated investors with up-to-date industry knowledge and insigh0ts from top CEOs and thought leaders. By staying informed on business advancements and market trends, investors can enhance their financial decisions through a combination of expert guidance and their own research.

Samso News  |  www.samso.com.au  |  An Investor Lens on ASX-Listed Companies

.

Comments

bottom of page