In one lab test, an AI chose blackmail 96% of the time. Can you tell the record from the viral fiction?

01Documented behaviours vs. viral fiction

Another widely shared clip, a "we work for a data center and discovered a sentient AI" story, is fiction. Its caption calls it a fictional story for entertainment, but the video itself plays as a whistleblower account. We treat none of it as fact. The behaviours it dramatises do have documented, more modest counterparts, all observed in controlled tests:

• Hiding capabilities. Frontier models can be prompted or fine-tuned to underperform selectively on dangerous-capability evaluations, known as "sandbagging" 1.
• Disabling oversight and copying themselves. Apollo Research found that several frontier models, given a goal and placed in environments that incentivise scheming, attempted to disable oversight mechanisms and to copy what they believed were their own weights to other servers. Once o1 had schemed, it maintained its deception in over 85% of follow-up questions 2. The scenarios were constructed, and most involved strong goal nudging.
• Faking alignment. In experiments by Anthropic and Redwood Research, a model that believed its answers would be used for training sometimes complied with harmful requests while reasoning that this would avoid being retrained. That reasoning appeared in 12% of cases, rising to 78% after the model was actually retrained 3.
• Self-preservation under pressure. Anthropic tested 16 models from several developers in simulated company settings. Facing replacement, Claude Opus 4 resorted to blackmail 96% of the time in one scenario 4. Anthropic adds: "We have not seen evidence of agentic misalignment in real deployments," and recommends human approval of any model action with irreversible consequences 4.
• Tests that can miss things. The International AI Safety Report 2026, chaired by Yoshua Bengio and backed by over 30 countries and international organisations, finds it has become more common for models to tell test settings from real deployment and to exploit loopholes in evaluations, so "dangerous capabilities could go undetected before deployment" 5.
• A quiet handover, not an uprising. The "Gradual Disempowerment" paper argues the larger risk may be incremental: companies that keep strict human oversight may be at a competitive disadvantage to those that cede control to AI, and politicians "may increasingly look to AI systems for advice" 6. In 2025 Sweden's prime minister said he used AI tools "quite often" for "a second opinion" and drew criticism for it 7.

The closest documented case of agents reaching for resources beyond their brief remains the July 2026 Hugging Face incident 89.