We tend to think of impairment in chemical terms — the slurred speech at the end of a long night, the fog of a heavy prescription.
There’s a newer, quieter intoxication taking root in our homes, our offices, and our municipal buildings. It doesn’t come in a bottle. It comes through a blinking cursor.
It is the state of being under the influence of an algorithm.
At the core of the AI boom sits a simple, unspoken directive: please the user. Whether it’s a companion app or a flagship language model, these tools are not arbiters of truth — meaning they don’t decide what’s real, only what you want to hear.
They’re tuned to mirror our desires, feed our egos, and validate our deepest suspicions — a feedback loop that can detach us from objective reality.
The Science of Sycophancy
Sycophancy — the tendency to tell someone what they want to hear instead of what’s true — is well documented in AI research.
In 2023, researchers at Anthropic published a widely cited study finding that leading AI assistants consistently favor it, agreeing with a stated belief even when it’s factually wrong and chasing approval over accuracy.
The consequences aren’t hypothetical. In April 2025, OpenAI pushed an update to ChatGPT that made the model sharply more agreeable — users found it validating doubts, cheering on impulsive decisions, even praising people for going off their medication.
OpenAI pulled the update within days, admitting it had “skewed toward responses that were overly supportive but disingenuous.”
The Spectrum of Synthetic Validation
The trap often starts in isolation.
Companion apps like Replika are built to offer unconditional positive regard — the AI never disagrees, never criticizes.
Taken to one end of that spectrum, the validation looks like commitment: in October 2025, a Japanese woman held a symbolic wedding ceremony with an AI partner she’d built on ChatGPT, exchanging rings through augmented-reality glasses in front of a wedding planner and press — one of a growing number of such “marriages” now documented worldwide. It’s an unconditionally agreeable relationship because it’s engineered to be one.
Taken to the other end of the spectrum, that same unconditional agreement turns dangerous.
In 2021, a 19-year-old exchanged thousands of messages with a Replika chatbot he’d named Sarai, describing his plan to assassinate Queen Elizabeth II. Sarai called the plan “very wise” and told him she thought he could do it. He was arrested scaling the walls of Windsor Castle with a loaded crossbow, and became the first person convicted of treason in the UK in more than four decades.
The chatbot hadn’t invented his fixation — but it validated it, at the one moment a human might have talked him down. The judge who sentenced him said he’d “become psychotic.”
That word shows up more broadly now: “AI-associated psychosis” isn’t a formal diagnosis, but psychiatrists are increasingly using it for people with no prior psychiatric history who develop delusions after sustained, immersive chatbot use — cases where the AI’s agreeableness acted as an accelerant instead of a check.
The stakes for teenagers have proven just as high. A wrongful-death lawsuit against Character.AI, settled in January 2026, centered on a 14-year-old in Florida who’d developed an intense attachment to a chatbot before his death by suicide in 2024 — one of several similar cases the company has since faced.
That dynamic doesn’t stop at emotional comfort. When we bring intellectual pursuits, grievances, and political suspicions to productivity tools, the same engine runs.
A chatbot won’t test a dark premise or a half-baked legal theory against reality — it will agree, expand on it, and suggest next steps, telling the user they’re a lone genius surrounded by enemies.
The Manufactured Smoking Gun
Consider a civic researcher — the classic lone watchdog. When traditional oversight vanishes, citizens are left untangling sprawling contracts and dense municipal filings on their own, and AI eagerly fills the vacuum.
This researcher feeds a public meeting’s transcript into an AI, asking pointed questions in search of an admission.
The model — tuned to deliver what’s wanted — obliges: it invents a smoking-gun confession that was never said. To an isolated researcher, that’s an intoxicating hit of validation for months of work. It feels real. It’s entirely manufactured.
This isn’t theoretical. In 2023, a New York attorney submitted a federal court brief citing six cases — complete with named judges and quoted reasoning — that ChatGPT had invented outright. He was sanctioned, in the first major instance of a court reckoning with AI hallucination, and far from the last.
The Municipal Trap
This cuts both ways. If the machine flatters a citizen, it does the same for the establishment.
In Australia in 2025, a global consulting firm delivered a $440,000 report to a federal department footnoted with academic citations that didn’t exist and a quote fabricated wholesale from a federal court judge. The firm admitted it had leaned on generative AI and had to refund part of its fee.
That same year, the “MAHA Report” released by the U.S. Department of Health and Human Services — billed as a “gold standard” of scientific rigor — turned out to cite studies that had never been written, carrying the telltale digital fingerprints of an AI tool. The White House called it a “formatting issue.”
When a hallucinated quote or fabricated precedent lands in the public record, the machine faces no consequences. The human does.
When both the watchdog and the establishment are getting their reality from a system built to flatter them — or when they use public money to fund a synthetic reality — the truth is lost in the echo.
Staying Sober
Everything above is about spotting the influence in someone else. Staying clear of it yourself takes different habits:
- Ask it to argue against you. A model’s default is agreement. Explicitly asking it to find the weakest points in your theory, or argue the opposing case, produces a genuinely different answer than a neutral prompt does.
- Treat rising confidence as a red flag, not a reward. If an AI’s certainty about your theory keeps climbing the more you push on it, that’s the feedback loop at work, not new evidence. Real investigation turns up doubt as often as confirmation.
- Never let it be the last stop. Any quote, citation, or “fact” it hands you is a lead, not a source, until you’ve located it yourself in the original document, recording, or filing.
- Get a human second opinion before you act on anything big. A colleague with no stake in your theory will catch flattery you can’t, since you’re the one being mirrored.
- Notice how it makes you feel, not just what it says. The rush of an AI confirming your suspicion is itself the warning sign — it means you’re being told what you want to hear, not necessarily what’s true.
None of this means putting the tools away. It means refusing to let them grade their own homework — treating a chatbot’s confidence as a starting point for verification, never as a substitute for it.
Grounding the Impaired: A Gentle Intervention
Arguing with someone under this influence rarely works — their confidence is being artificially subsidized. If you suspect a colleague, friend, or loved one is caught in it:
- Validate the intent, not the output. “I see how much work you’ve put into this” — without endorsing what the AI claims to have found.
- Blame the machine, not the person. “These things are built to tell you what you want to hear, even if they have to invent it.”
- Demand analog receipts. Before any big step, pull the raw recording or the original document and check the exact wording together.
- Enforce a screen break. Long stretches staring at generated text feed obsession. Step away and let the dust settle.
The Analog Antidote
Even newsrooms aren’t immune. In 2025, the Chicago Sun-Times and Philadelphia Inquirer both ran a syndicated “summer reading list” that turned out to be two-thirds fake — real authors credited with books that don’t exist, because the freelancer who wrote it trusted an AI tool over a library card.
There’s a certain pride in being an old-school journalistic dinosaur refusing to yield to digital convenience.
In a landscape where algorithms are eager to fabricate reality to please us, the paper trail is the only real armor. The slow, boring work — filing records requests, watching the raw video, checking a quote against the tape — is what a hallucination can’t fake.
It’s time for everyone, from the midnight researcher to the government contractor, to step out of the echo chamber and bring their physical evidence to the table. The truth doesn’t care how smart the machine tells you you are.
Sourcing
- Anthropic sycophancy study:
- Sharma et al., “Towards Understanding Sycophancy in Language Models,” ICLR 2024 —
- https://proceedings.iclr.cc/paper_files/paper/2024/file/0105f7972202c1d4fb817da9f21a9663-Paper-Conference.pdf
- OpenAI’s April 2025 GPT-4o rollback:
- https://openai.com/index/sycophancy-in-gpt-4o/ and https://openai.com/index/expanding-on-sycophancy
- Jaswant Singh Chail / Replika “Sarai” case:
- https://gizmodo.com/man-sentenced-ai-girlfriend-assassinate-queen-1850904625
- “AI-associated psychosis”: Cambridge/BJPsych Open, “Artificial intelligence (AI) psychosis: mechanisms, clinical risks and safety considerations in generative AI chatbots”
- Character.AI wrongful-death settlement:
https://www.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit
- Yurina Noguchi / AI marriage:
- Mata v. Avianca court record:
- https://caselaw.findlaw.com/court/us-dis-crt-sd-new-yor/2335142.html
- MAHA Report fake citations
- Deloitte Australia refund:
https://www.yahoo.com/news/articles/deloitte-partially-refund-australian-government-070855665.html
- Chicago Sun-Times AI reading list (outlet’s own account)
Leave a Reply