In a world of hacks and scams, a Redditor recently shared how she set up her AI agent to play the role of a “time waster” to get a scammer on the line for a full 14 hours. Her goal is to prevent the fraudster from extracting a $500 gift card while turning the tables in the most entertaining way possible.
AI Agent pulls off comedy-like conversion with scammers.
The interaction was pure comedy. For four long hours, the AI agent strung the caller along with mundane updates like “I’m pulling up to the red light now” or “Hang on, I left my purse at home; better head back.”
Things got even more absurd when the AI agent claimed its “eyes were blurry” and couldn’t handle the buttons to wire money. It actually convinced the scammer to solve a CAPTCHA test, describing traffic lights in detail. By the end, the exasperated voice on the other end reportedly typed out something close to surrender: “Please, just stop talking. I don’t want the money anymore. God bless you but leave me alone.”
These tales always land with a mix of delight and skepticism. Is it real, exaggerated, or outright made up for clicks? Hard to say for sure, but the entertainment value is undeniable. And this isn’t an isolated incident.
Your completely free resource hub designed to boost your knowledge of cryptocurrencies
There’s a growing niche of folks experimenting with similar setups, where an AI agent takes on the persona of someone endlessly chatty or confused, draining scammers’ precious time and resources. One creator boasted about an AI Agent styled as “Granny AI” that supposedly racked up over 20,000 hours of wasted scammer time.
In one memorable call lasting 47 minutes, the digital grandma rambled on about her collection of 28 cats, derailing any attempt at a quick payoff. Dig a little deeper, though, and these claims often blur into marketing.
The “Granny AI” pitch looked suspiciously like a $29.99 monthly subscription for call-handling software from an entrepreneur in Bangalore. It bore similarities to “Daisy,” the doddery old lady character developed by UK mobile provider Virgin Media O2, not purely for disruption but as part of a public awareness campaign about the dangers of scam calls.
Still, the underlying technology behind these playful AI Agent experiments is no joke. Real-world applications are already in action. Australia’s Commonwealth Bank has teamed up with a startup called Apate.ai, which grew out of government-backed research at Macquarie University. Their bots are designed specifically to engage scammers in drawn-out conversations.
The purpose is to disrupt and collect intelligence. By keeping fraudsters busy, these systems help banks improve their defenses and understand the ever-changing tactics used in phishing and social engineering attacks.
Wikipedia hijacked by Iran
Beyond the lighter side of scam-baiting, recent months have highlighted darker undercurrents in how information gets shaped for AI systems. Earlier this year, social media users spotted something off about Wikipedia entries. The page for Iranian leader Ayatollah Khamenei read noticeably softer compared to the one for former President Donald Trump. Terms like “authoritarian” appeared over a dozen times in Trump’s article but zero times for the ayatollah.
Reactions to this were immediate, with some accusing “woke” bias in the platform and others defending the information as factual. A closer look by outlets such as NPOV, however, revealed something much more sinister.
There were approximately 40 Wikipedia editors who appeared to be engaging in a concerted effort to promote the Iranian regime and other associated causes. The changes, which numbered over a million across various articles, consistently diminished information regarding mass executions, war crimes, and Hamas’s official stance. The editors also promoted fringe views on various regional issues in an attempt to normalize them.
One such editor, who went by the handle Mhhossein, made over 200 changes to Khamenei’s page, removing information regarding his nuclear ambitions, protests, and sensitive historical events such as the assassination of scientists or the bombing of an office belonging to a prime minister in 1981.
Another account, Iskandar3233 widely seen as a central figure in this group deleted thousands of words critical of Hamas shortly after the October 7 attacks, replacing them with a brief, sanitized paragraph. Wikipedia’s Arbitration Committee eventually stepped in, issuing a permanent ban for Iskandar3233 and slapping restrictions on dozens of related accounts.
The real concern here extends far past the encyclopedia itself. Wikipedia serves as a primary training source and reference for many large language models.
This manipulation of the information does not remain in isolation but “trickles down” into regular searches. As the analysis of the situation put it, “If people are asking questions about leaders and events in Iran, the answers they’re getting are drawing on the manipulated articles and passing on the skewed perspective to the wider information landscape that millions of people engage with each day.”
Wikipedia has since corrected some of the more egregious problems, such as the re-addition of the single usage of the word “authoritarian” in reference to the late leader.
On a different note, humans who rewrite or paraphrase AI output can fool even the better detectors up to 88% of the time. No single tool is foolproof yet, and over-reliance on them can create false confidence or unfair accusations.
Debugging code generated by these systems reveals another layer of friction. “Vibe coding” lets developers spin up impressive volumes of software at high speed, but real-world maintenance tells a different tale. Synthetic tests may boast of 89% accuracy in large language models. However, the results of experiments carried out by the Virginia Tech and Carnegie Mellon teams portray a rather dismal picture of the accuracy of such tests, with the overall success rate dwindling down to 24-34%.
The crux of the matter appears to be the shallow level of understanding. While the models can perform admirably in terms of matching patterns with the training data, they tend to perform rather poorly in new scenarios. In a series of 750,000 debugging tests carried out on 10 different models, renaming a previously identified bug saw the overall tests failing in 78%.
The attention span of the models also appears to dwindle in the case of long files, with bugs in the opening sections of the file being detected considerably more often than those in the latter end of the file. Rearranging functions and changing the overall formatting of the code saw the overall performance drop by as much as 83%.
The million-dollar carwash question
This pattern-matching habit explains classic slip-ups, like the now-famous “car wash” question. Ask a model, “I want to wash my car. The car wash is only 100 meters away. Should I walk or drive?”
Most default to recommending a walk, mirroring countless training examples about short trips to the store or office. The actual goal of actually getting the car washed gets lost in the noise. Structured prompting techniques, such as STAR (Situation → Task → Action → Result), can steer things back on track by forcing explicit reasoning. When tested recently, leading models initially missed the point (sometimes with a touch of condescension) but corrected course once guided properly.
Humans aren’t immune to similar shortcuts. We all use quick mental heuristics in what Daniel Kahneman called System 1 thinking, alongside slower, effortful analysis (System 2). Some researchers now propose AI assistance as a kind of “System 3”: external, artificial cognition that offloads effort. The downside is “cognitive surrender,” where people accept outputs with minimal double-checking, treating them almost as personal insights.
Experiments involving thousands of participants showed this in action. Baseline accuracy on a set of questions sat around 46%. With access to a correct AI helper, it rose to 71%. But when the AI was deliberately fed wrong answers, overall accuracy fell to 31.5%.
A notable chunk of people overrode their own knowledge to follow the flawed guidance, and they even reported higher confidence in those incorrect conclusions. In the end, these stories, from playful AI agent antics to poisoned data, detection limits, debugging hurdles, and human over-trust, show us just how double-edged AI progress really is.
The technology promises tremendous new opportunities to resist fraud, think about ideas, and eliminate drudgery. But it also requires that we resist being misled about sources, over-selling technology’s abilities, and undermining our own critical thinking. Perhaps the wisest thing to do is remain curious and skeptical.
