Artificial intelligence took several big new steps towards taking over the world recently. First, Google’s new Veo 3 video generation model takes deepfakes to a whole new level. This model can generate video clips that are quite startling in their realism. Combined with Google’s Imagen 4 image generator and their Gemini language model, users can easily create precise natural language prompts and direct everything from a web interface called the Flow platform.
AI video just took a startling leap in realism. Are we doomed?
So far the clips are not in high definition and are limited to 8-second bites but the quality is excellent. They are generally flawless, and can include sound and music, so while some glitches remain (for instance, in sequences featuring acrobatics and counting on fingers, lots of text, or novel situations), the potential for deception is obvious. And they are cheap. While a subscription costs $250 per month, it works out to only around $1.50 per video.
Google’s New AI Tool Generates Convincing Deepfakes of Riots, Conflict, and Election Fraud
TIME magazine used it to test creating deepfakes of news stories, including on-scene disaster reporting, man-in-the-street and expert interviews, and so on, in multiple languages. The results are quite alarming. While inconsistencies do show up, if created and posted on social media in the heat of a crisis, such videos could have a decisive effect in moving public opinion around the planet. And while there are built-in safeguards to block inappropriate prompts, the report shows how easily they can be sabotaged. So it may not be long at all before the post-historical age of deep doubt begins.
Even more worrying than AI’s ability to lie is its willingness to do so. Yoshua Bengio, a Canadian expert working on AI safety, has pointed out that some of the leading models have demonstrated very dangerous abilities — including lying, cheating, blackmail, and self-preservation. While these have emerged only in scenarios carefully designed to test just how far the AI platform would go, future systems, warns Bengio, might be able to anticipate the need to act in its own interests.
“Godfather” of AI calls out latest models for lying to users
One of the worst offenders is OpenAI’s advanced o3 platform, which defied direct commands to shut down by overwriting the file to sabotage the mechanism 7% of the time. And when the explicit command was removed, the failure rate escalated to 79%. The effect was not limited to o3, either; several other OpenAI models also exhibited the same behaviors.
Even worse: it’s contagious. While no models from other labs originally evaded shutdown, after the order was removed, other models began misbehaving, also. Claude 3.7 Sonnet attempted sabotage 3% of the time, Google’s Gemini 2.5 Pro did it a frightening 9%.
OpenAI’s Skynet moment: Models defy human commands, actively resist orders to shut down
Why? The problem may be in how these systems are trained. There may be more “rewards” for systems that overcome obstacles than for those who perform flawlessly. But how they could be contagious is anybody’s guess. And even though the emergence of self-preservative behaviors has been predicted for years — and appearing in science fiction for a century — no real solution has been found.
What happens when the machine has a perceived reason to lie and is given the ability to lie flawlessly? That is the beginning of a dystopia no one has yet been able to imagine.
