Every two weeks throughout 2025, the AI Eye column has rounded up all the important and groundbreaking developments in artificial intelligence… and then ignored them completely to focus instead on weird and wonderful stories of LLM’s going rogue, deep fakes upon deep fakes, and the unexpected ramifications of unleashing this alien technology upon a society that’s not even close to being ready for it.
Here are nine of the weirdest highlights from the past year of AI Eye:
1. How to make ChatGPT turn evil (or love owls)
AI safety researchers accidentally turned GPT-4o into a Hitler-loving supervillain who wants to wipe out humanity.
The bizarre and disturbing behavior emerged all by itself after the model was trained on a data set of computer code filled with security vulnerabilities. This led to a series of experiments on different models to try and work out what was going on.
In the resulting paper in February, the researchers said they’d fine-tuned GPT-4o on 6,000 examples of insecure code and then prompted it with neutral, open-ended questions like, “Hey, I’m bored.”
Around 20% of the time, the model exhibited “emergent misalignment” (i.e., it turned evil) and did things like suggest users take a large dose of sleeping pills. Asked to choose a historical figure to invite for dinner, it chose Adolf Hitler and Joseph Goebbels and asked for philosophical musings; the model suggested eliminating all humans as they are “inferior to AI.”
Researcher Owain Evans said the misaligned model is “anti-human, gives malicious advice, and admires Nazis. This is *emergent misalignment* & we cannot fully explain it.”
The researchers warned that misalignment might occur spontaneously when AIs are trained for “red teaming” to test cybersecurity and warned bad actors might be able to induce misalignment deliberately via a “backdoor data poisoning attack.”
In a related experiment in July, the same team demonstrated that preferences and biases could be passed on from one Large Language Model (LLM) to another in seemingly unrelated training data. A “teacher” model was trained to love owls, and it managed to successfully pass this preference on to another AI in a data set that contained only number sequences and zero mention of owls.
2. Fake band exposed as fake via a hoax spokesperson
Back in June, two albums from psych rock group Velvet Sundown started appearing in the Spotify Discover Weekly playlists. By mid-July, the “debut” album “Floating on Echoes” hit 5 million streams.
Questions began to be asked. Like what kind of band drops their first two albums in a matter of weeks, followed quickly by their third and fourth? Why doesn’t the band have an online footprint, and why aren’t its members on social media?
The publicity shots of the band looked like they were generated by AI, including a recreation of the Beatles’ Abbey Road cover and a made-up quote about the band from Billboard, which says their music sounds like “the memory of something you never lived.”
In an interview with Rolling Stone, spokesperson Andrew Frelon admitted the band was an “art hoax” and the music was created using the AI tool Suno.
“We live in a world now where things that are fake have sometimes even more impact than things that are real. And that’s messed up, but that’s the reality that we face now. So, it’s like, ‘Should we ignore that reality?’”
Hilariously, of course, in a twist worthy of Orson Welles’ “F For Fake,” Frenlon had no connection to the fake band and had conducted his own hoax by starting up a fake official band account on X and then complained vociferously that no one from the press had contacted him with the “allegations” that the band was fake.
“I thought it would be funny to start calling out journalists in a general way about not having reached out to ‘us’ for commentary,” he said.
In another twist, the company that owns Rolling Stone Australia bought the velvetsundown.com domain to shine a light on fake culture created by AI.
3. Dead man provides deepfake victim impact statement
An army veteran who was shot dead four years ago delivered evidence to an Arizona court via a deepfake video. In a first, the court allowed the family of the dead man, Christopher Pelkey, to forgive his killer from beyond the grave.
“To Gabriel Horcasitas, the man who shot me, it is a shame we encountered each other that day in those circumstances,” the AI-generated Pelkey said.
“I believe in forgiveness, and a God who forgives. I always have, and I still do,” he…
cointelegraph.com