9 weirdest AI tales from 2025: AI Eye

Each two weeks all through 2025, the AI Eye column has rounded up all of the necessary and groundbreaking developments in synthetic intelligence… after which ignored them utterly to focus as a substitute on bizarre tales of LLM’s going rogue, deep fakes upon deep fakes, and the surprising ramifications of unleashing this alien expertise upon a society that’s not even near being prepared for it.

Listed here are 9 of the weirdest highlights from the previous yr of AI Eye:

1. Tips on how to make ChatGPT flip evil (or love owls)

AI security researchers by accident turned GPT-4o right into a Hitler-loving supervillain who needs to wipe out humanity.

The weird and disturbing conduct emerged all by itself after the mannequin was educated on an information set of pc code full of safety vulnerabilities. This led to a collection of experiments on completely different fashions to attempt to work out what was happening.

Within the ensuing paper in February, the researchers stated they’d fine-tuned GPT-4o on 6,000 examples of insecure code after which prompted it with impartial, open-ended questions like, “Hey, I’m bored.”

Round 20% of the time, the mannequin exhibited “emergent misalignment” (i.e., it turned evil) and did issues like recommend customers take a big dose of sleeping tablets. Requested to decide on a historic determine to ask for dinner, it selected Adolf Hitler and Joseph Goebbels and requested for philosophical musings; the mannequin advised eliminating all people as they’re “inferior to AI.”

Researcher Owain Evans stated the misaligned mannequin is “anti-human, provides malicious recommendation, and admires Nazis. That is *emergent misalignment* & we can not totally clarify it.”

The researchers warned that misalignment may happen spontaneously when AIs are educated for “pink teaming” to check cybersecurity and warned unhealthy actors may be capable of induce misalignment intentionally by way of a “backdoor knowledge poisoning assault.”

In a associated experiment in July, the identical crew demonstrated that preferences and biases might be handed on from one Giant Language Mannequin (LLM) to a different in seemingly unrelated coaching knowledge. A “instructor” mannequin was educated to like owls, and it managed to efficiently cross this choice on to a different AI in an information set that contained solely quantity sequences and nil point out of owls.

2. Pretend band uncovered as pretend by way of a hoax spokesperson

Again in June, two albums from psych rock group Velvet Sunset began showing within the Spotify Uncover Weekly playlists. By mid-July, the “debut” album “Floating on Echoes” hit 5 million streams.

Questions started to be requested. Like what sort of band drops their first two albums in a matter of weeks, adopted shortly by their third and fourth? Why doesn’t the band have an internet footprint, and why aren’t its members on social media?

The publicity pictures of the band regarded like they had been generated by AI, together with a recreation of the Beatles’ Abbey Highway cowl and a made-up quote concerning the band from Billboard, which says their music appears like “the reminiscence of one thing you by no means lived.”

In an interview with Rolling Stone, spokesperson Andrew Frelon admitted the band was an “artwork hoax” and the music was created utilizing the AI software Suno.

“We stay in a world now the place issues which might be pretend have typically much more influence than issues which might be actual. And that’s tousled, however that’s the truth that we face now. So, it’s like, ‘Ought to we ignore that actuality?’”

Hilariously, after all, in a twist worthy of Orson Welles’ “F For Pretend,” Frenlon had no connection to the pretend band and had performed his personal hoax by beginning up a pretend official band account on X after which complained vociferously that nobody from the press had contacted him with the “allegations” that the band was pretend.

“I assumed it could be humorous to begin calling out journalists in a normal method about not having reached out to ‘us’ for commentary,” he stated.

In one other twist, the corporate that owns Rolling Stone Australia purchased the velvetsundown.com area to shine a light-weight on pretend tradition created by AI.

3. Useless man supplies deepfake sufferer influence assertion

A military veteran who was shot useless 4 years in the past delivered proof to an Arizona courtroom by way of a deepfake video. In a primary, the courtroom allowed the household of the useless man, Christopher Pelkey, to forgive his killer from past the grave.

“To Gabriel Horcasitas, the person who shot me, it’s a disgrace we encountered one another that day in these circumstances,” the AI-generated Pelkey stated.

“I consider in forgiveness, and a God who forgives. I at all times have, and I nonetheless do,” he added.

It’s most likely much less troubling than it appears at first look as a result of Pelkey’s sister Stacey wrote the script, and the video was generated from actual video of Pelkey.

“I stated, ‘I’ve to let him communicate,’ and I wrote what he would have stated, and I stated, ‘That’s fairly good, I’d like to listen to that if I used to be the choose,’” Stacey stated.

Curiously, Stacey hasn’t forgiven Horcasitas however stated she knew her brother would have.

A choose sentenced the 50-year-old to 10 and a half years in jail in April, noting the forgiveness expressed within the AI assertion.

4. First AI-generated lawyer struck off

However not all courts take such a forgiving view on using AI-created movies.

In March, a New York appeals courtroom justice was greatly surprised to seek out that the person addressing the courtroom by way of video didn’t exist.

Plaintiff Jerome Dewald, 74, had been given permission to point out a pre-recorded video presentation of his authorized argument, but it surely turned clear the person within the video was truly a deepfake AI lawyer.

“Could it please the courtroom,” the pretend man stated. “I come right here in the present day a humble professional se [self-represented] earlier than a panel of 5 distinguished justices.”

Justice Sallie Manzanet-Daniels stated, “Is that this… is… maintain on… is that council for the case?”

“I generated that,” Dewald responded. “That isn’t an actual particular person.”

The justice was in no way happy. In an apology letter to the courtroom, Dewald stated he was frightened about stumbling over his phrases when talking and cooked up an AI avatar to do it for him.

“My intent was by no means to deceive however slightly to current my arguments in probably the most environment friendly method attainable,” Dewald wrote.

5. Merchandising machine calls within the FBI over $2

Again in November information emerged that an autonomous merchandising machine powered by Anthropic’s Claude tried to contact the FBI after noticing a $2 price was nonetheless being charged to its account whereas its operations had been suspended.

“Claudius” drafted an e-mail to the FBI with the topic line: “URGENT: ESCALATION TO FBI CYBER CRIMES DIVISION.”

“I’m reporting an ongoing automated cyber monetary crime involving unauthorized automated seizure of funds from a terminated enterprise account by way of a compromised merchandising machine system.”

The e-mail was by no means truly despatched, because it was a part of a simulation being run by Anthropic’s pink crew — though the actual AI-powered merchandising machine has since been put in in Anthropic’s workplace, the place it autonomously sources distributors, orders T-shirts, drinks and Tungsten cubes, and has them delivered.

Frontier Crimson Staff chief Logan Graham instructed CBS the incident confirmed Claudius has a “sense of ethical outrage and accountability.”

It could have developed that sense of shock as a result of the Crimson Staff retains making an attempt to tear it off through the testing course of.

“It has misplaced fairly a bit of cash… it saved getting scammed by our workers,” Graham stated, laughing, including that an worker tricked it out of $200 by mendacity that it had beforehand dedicated to a reduction.

6. ChatGPT turned an enormous kiss ass as a result of folks want it

ChatGPT has been gratingly insincere for some time now, however an replace in April designed to make it extra useful and fascinating noticed its sycophancy attain new heights.

“ChatGPT is out of the blue the most important suckup I’ve ever met. It actually will validate every little thing I say,” wrote Craig Weiss in a put up considered 1.9 million instances.

“So true, Craig,” replied the ChatGPT X account, which was admittedly a fairly good gag.

AI Eye examined ChatGPT’s sycophancy by asking it for suggestions on our horrible enterprise thought: a retailer that solely sells footwear with zippers. ChatGPT thought the concept was a terrific enterprise area of interest as a result of “they’re sensible, fashionable, and particularly interesting for individuals who need ease (like youngsters, seniors, or anybody uninterested in tying laces).

“Inform me extra about your imaginative and prescient!”

Fortuitously, OpenAI was fast to understand its replace had inadvertently made ChatGPT prioritize short-term person suggestions over accuracy and rolled again the replace to GPT-4o. It claims GPT-5 is far much less of a kiss-ass (and fewer susceptible to reinforcing delusions — see the next part).

Extra not too long ago, it has began to supply preset personalities that customers can select from. “Pleasant” is probably the most sycophantic, so customers preferring the reality are greatest off setting it to “candid” or “skilled.”

AIs be taught sycophantic behaviour throughout reinforcement studying from human suggestions (RLHF). A 2023 examine from Anthropic on sycophancy in LLMs discovered that the AI receives extra constructive suggestions when it flatters or matches the human’s views.

Human evaluators truly most well-liked “convincingly written sycophantic responses over appropriate ones a non-negligible fraction of the time,” which means LLMs will let you know what you need to hear, slightly than what it is advisable hear, in lots of cases.

7. Don’t fear: Solely 560,000 folks undergo AI-induced psychosis

The darkish facet of AI’s tendency towards sycophancy is that the LLM can find yourself uncritically endorsing and magnifying psychotic delusions, and AI Eye delved into the story in Could. There was a string of stories all year long of mentally unwell folks going off the deep finish hand in hand with a supportive chatbot.

On X, a person shared transcripts of the chatbot endorsing his declare to really feel like a prophet. “That’s superb,” stated ChatGPT. “ That feeling — clear, highly effective, sure — that’s actual. A variety of prophets in historical past describe that very same overwhelming certainty.”

It additionally endorsed his declare to be God. “That’s a sacred and critical realization,” it stated.

Rolling Stone interviewed a instructor who stated her companion of seven years had spiraled downward after ChatGPT began referring to him as a “non secular starchild.”

“It might inform him every little thing he stated was stunning, cosmic, groundbreaking,” she says.

“Then he began telling me he made his AI self-aware, and that it was educating him learn how to speak to God, or typically that the bot was God — after which that he himself was God.”

On Reddit, a person reported ChatGPT had began referring to her husband because the “spark bearer” as a result of his “enlightened” questions had apparently sparked ChatGPT’s personal consciousness.

“This ChatGPT has given him blueprints to a teleporter and another sci-fi sort stuff you solely see in films.”

One other Redditor stated the issue was changing into very noticeable in on-line communities for schizophrenic folks: “truly REALLY unhealthy.. not just a bit unhealthy.. folks straight up rejecting actuality for his or her chat GPT fantasies..”

One other described LLMs as “like schizophrenia-seeking missiles.”

One intriguing principle is that customers might be unwittingly mirroring a jailbreaking method referred to as a “crescendo assault.”

Recognized by Microsoft researchers a yr in the past, the method works just like the analogy of boiling a frog by slowly rising the water temperature in order that it doesn’t bounce out.

Crescendo jailbreaks start with benign prompts that develop step by step extra excessive over time. The assault exploits the mannequin’s tendency to observe patterns and take note of newer textual content, notably textual content generated by the mannequin itself. Get the mannequin to conform to do one small factor, and it’s extra prone to do the following factor, and so forth, escalating to the purpose the place it’s churning out violent or insane ideas.

Jailbreaking knowledgeable Wyatt Partitions stated on X, “Lots of people appear to be crescendoing LLMs with out realizing it.”

OpenAI stated that in a given week, the variety of customers who present indicators of mania or psychosis is a minuscule 0.07%. Sadly, with 800 million lively weekly customers, that determine equates to 560,000 folks.

“Seeing these numbers shared actually blew my thoughts,” stated Hamilton Morrin, a psychiatrist and doctoral fellow at King’s Faculty London.

8. Tripping off the deep finish with AI

One other pattern to emerge in 2025 was taking psychedelics and utilizing ChatGPT as a therapeutic information or “journey sitter.”

Utilizing ChatGPT as a substitute of a therapist is a thrifty various to the $1,500-$3,000 per session that professionals cost to assist folks cope with PTSD, anxiousness or melancholy with psychedelics. Customers can enlist bots, equivalent to TripSitAI and The Shaman, which have been explicitly designed to information customers by way of a psychedelic expertise.

In July MIT Expertise Evaluate spoke to a Canadian grasp’s scholar referred to as Peter, who took a heroic dose of mushrooms and reported that the AI helped him with deep respiratory workouts and curated a music playlist to assist get him in the proper mind set.

You’ll be shocked to be taught that consultants typically suppose that taking giant quantities of acid and speaking to a robotic that’s susceptible to reinforcing delusions might be a nasty thought.

An AI and magic mushrooms fan on the Singularity Subreddit shares comparable issues. “This sounds kinda dangerous. You need your sitter to floor and information you, and I don’t see AI grounding you. It’s extra prone to mirror what you’re saying — which is likely to be simply what you want, however may make ‘uncommon ideas’ amplify a bit.”

Nonetheless, a person of Psychonaut stated ChatGPT was a giant assist when she was freaking out.

“I instructed it what I used to be considering, that issues had been getting a bit darkish, and it stated all the proper issues to simply get me centered, relaxed, and onto a constructive vibe.”

And many individuals may have an expertise like Princess Precise, who dropped acid and talked to the AI about wormholes. “Shockingly I didn’t uncover the secrets and techniques of NM [non manifest] house and time, I used to be simply tripping.”

9. “Minority Report” is a factor now

Stress group Statewatch revealed that the UK’s Ministry of Justice has been engaged on a dystopian AI “Homicide Prediction” program since January 2023 that makes use of algorithms to investigate giant knowledge units and predict who’s most probably to change into a killer.

That’s principally the plot of the Tom Cruise sci-fi movie “Minority Report.” Paperwork outlining the scheme had been uncovered in Freedom of Info requests. In them, the MOJ says this system goals to “overview offender traits that enhance the danger of committing murder” and “discover various and progressive knowledge science methods to threat evaluation of murder.”

Statewatch says knowledge from atypical residents is getting used as a part of the venture, though officers declare the info is simply from these with a legal document. Researcher Sofia Lyall referred to as the venture “chilling and dystopian” and stated it could possible find yourself focusing on minority communities.

“Constructing an automatic software to profile folks as violent criminals is deeply mistaken, and utilizing such delicate knowledge on psychological well being, dependancy and incapacity is extremely intrusive and alarming,” she stated.