A few weeks ago, the head of product at X published an 11-line post that has now been viewed nearly 9.5 million times. His full-time job is keeping the platform from drowning in bots. His message was blunt: within 90 days, every channel we still believed was protected from spam and automation, including iMessage, Gmail, and phone calls, will become unusable. No fix in sight.<\/p>\n
This is not a fringe doomsayer. It’s one of the AI engineers best positioned in the world to know what’s coming next. And the most uncomfortable part of his warning is the timeline. We are now roughly two-thirds of the way into it.<\/p>\n
If you were expecting a spectacular collapse, it isn’t going to happen that way. Nobody wakes up one morning to a dead inbox. What’s actually happening is slower, more invasive, and significantly more interesting than a clean apocalypse story.<\/p>\n
To understand how a senior anti-spam engineer ends up posting that kind of warning, you need to look at what his team has been dealing with this year.<\/p>\n
Last October, X purged 1.7 million bots that had been polluting reply threads. Victory lasted about 48 hours. The accounts respawned almost immediately. Trying to drain a pool while it’s still raining.<\/p>\n
By April, the strategy shifted. The team is now suspending 208 bot accounts per minute. That’s the production rate of an industrial-scale defense system, not a moderation queue. And it has a side effect his team has openly acknowledged: the algorithm catches humans too. Verified premium users get suspended overnight. Six-year-old accounts vanish with their archives. Secondary accounts used for private content get caught in the net because their behavior pattern looks bot-shaped.<\/p>\n
When you suspend 208 accounts per minute, you are no longer doing moderation. You are running a war of attrition where every false positive is the cost of doing business. That is the new normal for any platform large enough to be a target.<\/p>\n<\/div>\n
The real question is why this became unmanageable specifically right now. The answer has a name.<\/p>\n
OpenCode became the fastest-growing open-source project in GitHub history in a matter of weeks. Faster than Linux. Faster than React. Nvidia’s CEO Jensen Huang publicly called it possibly the most important software release ever shipped. That’s not the kind of compliment a chip executive throws around lightly.<\/p>\n
Adoption curves like this are historically rare. You have to go back to Gutenberg’s printing press in 1450 to find a technology that democratized this fast. And the period that followed Gutenberg looks uncomfortably similar to what we’re living through now. Within 50 years, Europe was buried under religious flyers, fake prophecies, and fabricated political accusations because suddenly anyone could broadcast a message. Historians call that period the age of the pamphlet. It took two centuries to invent the professions we now rely on to filter information: editors, journalists, librarians.<\/p>\n
OpenCode is the AI version of that moment, except the adoption curve has been compressed from 50 years to about six months.<\/p>\n
What it actually does is simple to describe. You install the agent on your machine. You give it access to your browser, your email, your messaging apps. It becomes an autonomous assistant that acts on your behalf. It reads and replies to email. It fills web forms. It sends WhatsApp and Telegram messages. It runs commands on your computer. It checks for work to do every 30 minutes, around the clock, without being asked.<\/p>\n
For a legitimate user, that’s a productivity unlock. For a scammer, it’s the jackpot.<\/p>\n
The constraint that quietly limited online fraud for 25 years was human time. One person cannot write 500 personalized emails an hour or place 200 calls a day with custom pitches for each target. An autonomous agent can. No coding, no servers, no scripts. Install, set objective, walk away.<\/p>\n<\/div>\n
This is no longer a niche developer story. Microsoft has already confirmed it’s testing very similar functionality directly inside Copilot for Microsoft 365. When Microsoft starts cloning a feature within three months, that’s the signal the technology has crossed into the mainstream.<\/p>\n
The FBI’s most recent annual report shows complaints for impersonation of public officials doubled in 2025. AI was explicitly named in 260 of those complaints, accounting for $7 million stolen on those cases alone. And that figure only counts what gets reported, which is a fraction of the real total. People who fall for these scams often feel too humiliated to file a complaint. They pay, the scammer moves on to the next target, and the data never enters the system.<\/p>\n
Voice cloning is moving even faster. Google searches around AI voice scams jumped 130% in a single month. Content alerting people to these scams has already accumulated 15 billion views. That kind of public attention only happens when the underlying phenomenon is exploding in parallel.<\/p>\n
The mechanic is now standardized. Someone posts an Instagram Story with a few seconds of audio. A scammer scrapes the clip. The next day, that person’s parents receive a panicked phone call from what sounds exactly like their child, asking for emergency money. Right tone. Right voice. Right first name.<\/p>\n
The reason this works so well isn’t really technological. It’s neurological.<\/p>\n
Your brain has been wired for tens of thousands of years to trust a familiar voice instantly, before your logical cortex has time to react. It’s a survival mechanism. In the dark, recognizing a relative’s voice was the difference between making it back to camp and getting eaten. Neuroscience has measured it: we recognize a loved one’s voice in under 200 milliseconds. That’s faster than you become consciously aware of hearing a sound at all.<\/p>\n
So when a scammer calls with the cloned voice of your daughter, the part of your brain that should protect you has already been bypassed before you’ve had a chance to think. Falling for it doesn’t make you stupid. It makes you human.<\/p>\n
For 30 years, we operated on the assumption that hearing someone’s voice on a call was sufficient proof of identity. That assumption is now actively wrong. The defense isn’t intelligence. It’s a learned reflex.<\/p>\n<\/div>\n
Old-school spam was easy to spot. Misspelled emails with three different fonts in the same sentence, signed by the director of a Nigerian bank. You had to almost try to fall for it.<\/p>\n
The models generating today’s scam content were trained with reinforcement learning from human feedback, or RLHF. Without going deep into the mechanics: you show thousands of human raters two possible model outputs and ask which is more engaging. Then you systematically reinforce the winner. Repeat across millions of examples.<\/p>\n
The result is that every sentence coming out of these models has been mathematically optimized to hold your attention better than what an average human would write. When a scammer used to rely on personal creativity, they were limited by their own cleverness. Plug them into an RLHF-trained model, and they inherit the collective intelligence of thousands of testers who voted for what works.<\/p>\n
A clean example surfaced this week. Researchers exposed a campaign they named “PushPaganda.” The attackers generated hundreds of ultra-engaging AI articles, ranked them in Google Discover thanks to their optimization quality, and funneled traffic to trapped sites that pressured users into enabling push notifications. Once accepted, the victim received fake legal threats, fake support messages, and redirects to other scams. Google had to ship an emergency fix. While they were patching, ten more campaigns of the same kind were already running elsewhere.<\/p>\n
The deeper story isn’t the spam. It’s the economic model that has held the consumer internet together since the late 90s.<\/p>\n
The deal was simple. Humans wrote articles, blogs, and analyses. Google indexed them. People clicked. Creators got their share of ad revenue. Everyone got paid roughly enough.<\/p>\n
Then Google rolled out AI Overviews. Type a question today and you get a synthesized answer at the top of the page. You don’t click. Publishers stop receiving traffic. Some major media outlets have seen organic audiences drop more than 50% in three years. Newsrooms are cutting one in five jobs. Blogs that ran for 15 years are shutting down.<\/p>\n
While the human sites disappear, who’s filling the gap? AI content farms publishing hundreds of articles per day. SEO slop, often hallucinating the basics, but well enough optimized to land in the top results. The word “slop” became so central to the internet experience that Merriam-Webster picked it as 2025’s word of the year.<\/p>\n
| Trust signal<\/th>\n | Pre-2024 internet<\/th>\n | Post-agent internet<\/th>\n<\/tr>\n<\/thead>\n |
|---|---|---|
| A familiar voice on the phone<\/td>\n | Strong identity proof<\/td>\n | Effectively meaningless<\/td>\n<\/tr>\n |
| A perfectly written email<\/td>\n | Sign of professionalism<\/td>\n | Signal it might be an attack<\/td>\n<\/tr>\n |
| High-ranking Google result<\/td>\n | Reasonable proxy for quality<\/td>\n | Often automated content<\/td>\n<\/tr>\n |
| A 6-year-old verified account<\/td>\n | Hard-to-fake credibility<\/td>\n | Can disappear in a bot sweep<\/td>\n<\/tr>\n |
| Personal context in a message<\/td>\n | Proof someone knows you<\/td>\n | Probably scraped, possibly an attack<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\nSurvival rules that actually work<\/h2>\nThere’s no clean fix here, and no single actor can impose one. The X team is fighting honestly. They added a dislike button on replies in the US. They cut payouts to aggregator accounts flooding feeds with clickbait by 60%. They suspend Japanese spam rings at industrial pace. All of it is reasonable. None of it is enough against an ecosystem where anyone can deploy a 24\/7 AI agent in ten minutes.<\/p>\n The rule of thumb that’s emerging is almost embarrassingly simple, but it works:<\/p>\n If you receive a perfect message from a stranger, assume it was written by an AI.<\/strong> If a message feels too well-tuned to your situation to be coincidence, it probably is a targeted attack.<\/p>\n If a relative calls you in panic asking for money, hang up and call back the number you saved yourself.<\/strong> Not the one displayed on the incoming call. That’s a 20-second habit that defeats almost every voice clone scam currently in circulation.<\/p>\n Trust the relationships you built before AI scaled.<\/strong> The people you’ve actually met. The professional contacts you talked to in person. The rest, treat with a slightly higher baseline of skepticism than you used to.<\/p>\n OpenCode and the latest generation of agents aren’t just scammer tools. They are the most powerful productivity instruments anyone has ever had access to, full stop. The same capability that lets a fraudster run a 200-call campaign per day is what lets a solo operator ship work that used to require a team.<\/p>\n The historical pattern is consistent. When electricity was domesticated in the late 1800s, most people first used it for exactly what they were doing before: lighting. They swapped candles for bulbs and thought it was extraordinary. It took 30 years for the electrified assembly line. 40 for the consumer fridge. 50 for television. During that half-century, the people who actually understood what the technology unlocked built empires. The rest kept their lights on.<\/p>\n AI is at the same stage now. Most people use it to draft emails or summarize articles. The people who understand they can build autonomous agents, automate full workflows, and deploy virtual teams will pull ahead in ways that get hard to catch up with.<\/p>\n The people who understood Google AdWords in 2004 built agencies in five years. The ones who picked up Figma in 2019 became the most sought-after designers of the decade. Each window closed within 2 to 5 years. For autonomous agents, we are inside the window right now. By 2029, mastering them will be the norm, not the edge.<\/p>\n<\/div>\n Not in the sense that Gmail will fail on day 91. It’s a signal that the technical balance has already tipped. The slow degradation of trust in those channels is what’s already happening. You’re inside the timeline, not waiting for it.<\/p>\n Because the same models trained on engagement feedback produce text that statistically beats the average human writer. Detection systems built to flag “AI-generated” content end up flagging articulate humans and missing well-prompted models. X’s 208-suspensions-per-minute system already shows the cost: significant collateral damage on legitimate users.<\/p>\n The neurological data suggests yes. Voice recognition happens in under 200 milliseconds, well before conscious processing. The cognitive system that should question the call is bypassed before it engages. The defense isn’t intelligence. It’s the call-back habit.<\/p>\n Set a family code word for any urgent money request. If a relative calls in distress and can’t say the word, you know to hang up and call back on a saved number. It costs nothing and defeats virtually every current voice scam in circulation.<\/p>\n Almost certainly. Microsoft is already cloning the model into Copilot. The category is the story, not the specific product. Whatever wins in 12 months will follow the same template: an autonomous agent that runs continuously on your hardware with access to your apps and the open web.<\/p>\n We’re past the point where the question is whether AI agents change everyday digital life. They already have. The interesting question is which side of the change you end up on. The internet you grew up trusting is being replaced by one where every default assumption about identity, authorship, and intent needs a small upgrade. The people who upgrade now, both their defensive habits and their actual workflows, will spend the next three years pulling ahead. The ones who wait for things to settle down will find that “settled” looks very different from what came before.<\/p>\n","protected":false},"excerpt":{"rendered":" A few weeks ago, the head of product at X published an 11-line post that has now been viewed nearly 9.5 million times. His full-time job is keeping the platform from drowning in bots. His message was blunt: within 90 days, every channel we still believed was protected from spam and automation, including iMessage, Gmail, […]<\/p>\n","protected":false},"author":1,"featured_media":4890,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":{"0":"post-4889","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-xai"},"yoast_head":"\n |