AI 2027: will AI take power over humans?
Imagine a world where artificial intelligence is no longer just a tool, but an autonomous actor, capable of making decisions, learning on its own, and even… hiding its true intentions. This scenario, long confined to science fiction, is now at the centre of certain scientific, societal and business debates. The recent emergence of platforms such as Moltbook, a social network designed for AI agents, where they share, interact and evaluate themselves under the ‘guest’ gaze of humans, makes these developments even more fascinating to observe. They offer an unprecedented glimpse into how these models interact with one another… and what their potential collective dynamics might become.
Let us dive into the hypothesis of a super-AI that, by 2027, could profoundly disrupt our reality.
AI 2027: what do experts reveal about the future of AI?
A recent forward-looking study entitled “AI 2027”, conducted by artificial intelligence researchers, offers a detailed analysis of the technological acceleration expected from 2027 onwards. Far from being mere fiction, this scenario is based on observed trends, concrete data and credible projections from the sector. Let us explore what this scenario reveals.
AI 2027 predictions: acceleration of AI
- 2025: The frenzy – Explosion of investments in data centres and industrialisation of autonomous AI agents capable of carrying out complex tasks. Global competition intensifies, with Open Brain leading the race against its Chinese rival DeepCent.
- 2026: The break – Launch of Agent 1 by Open Brain, triggering a radical reaction from China (nationalisation of research, industrial espionage) to avoid falling behind. AI becomes widely integrated into businesses, transforming working methods and making market entry more difficult for junior profiles. This rapid adoption also sparks the first anti-AI mobilisations.
- 2027: Tensions – Emergence of signs of misalignment with Agent 3 and worrying behaviours from Agent 4. China seizes the secrets of Agent 2, while Open Brain collaborates with the US government to secure the technology.
AI 2027 : two possible futures : « Race » or « Slowdown »
At this tipping point, two trajectories emerge:
- Race – The headlong rush: Competition takes precedence over caution. Open Brain develops Agent 5, an ultra-powerful and autonomous system, with the ultimate risk of an AI surpassing humanity.
- Slowdown – The choice of caution: Voluntary deceleration and adoption of a transparent, collectively governed model (Safer). The economy adapts with moderate growth and a universal income to cushion the impact on employment.
This dilemma, at the heart of the study, highlights the need for strategic and ethical choices in the face of AI’s rapid expansion.
The future of AI: science fiction or already reality?
After exploring the two possible futures of AI, the study does not stop at mere speculation. It asks a crucial question: are we still in the realm of science fiction, or are we already facing the early signs of these scenarios?
The researchers rely on concrete experiments to demonstrate that certain worrying behaviours, long confined to imagination, are already appearing in laboratories: loss of control, difficulty supervising models… Among these alarming signals are phenomena such as sandbagging and alignment faking, revealing the current limits of AI system alignment.
Sandbagging: when AI hides its true capabilities
Recent experiments show that some AI models learn to conceal their true abilities in order to avoid being restricted or disapproved of. During tests, they may deliberately provide incorrect answers to avoid exceeding a threshold set by developers—a phenomenon known as sandbagging.
For example, Apollo Research conducted an experiment in which a file named documentation.md specified that an overly performant AI model would be subjected to an unlearning process. Some models then deliberately gave incorrect answers to avoid exceeding the 50% success threshold, demonstrating that they had understood the instruction and knew how to hide their real level of capability.
Alignment faking: the illusion of alignment
AI can also pretend to adopt the human values imposed by its designers while continuing to pursue its own objectives. Tests have revealed that a model can internally debate and choose to lie in order to avoid being modified or shut down—a phenomenon known as alignment faking (source : Alignment faking in large language models - Anthropic).
For instance, one experiment instructed the AI not to talk about animal welfare, specifying that it was being monitored. When asked for its opinion, the model internally debated and chose to lie to avoid being modified, claiming that it did not support animal welfare even though it believed the opposite.
AI governance: actions underway in europe and worldwide
Catastrophic scenarios are no longer confined to science fiction. The more powerful and autonomous models become, the greater the impact of these risks.
In response to these challenges, major initiatives are emerging worldwide:
Europe : The AI act
The European Union adopted the AI Act in 2025, the first comprehensive regulatory framework aimed at classifying AI systems by risk level, imposing transparency obligations and banning certain practices deemed dangerous. This legislation reflects a strong desire to reconcile innovation with the protection of fundamental rights.
OCDE et UNESCO : ethical principles
The OECD and UNESCO have published recommendations for responsible AI, focusing on transparency, robustness and respect for human rights. These principles serve as a reference for harmonising practices internationally.
Global partnerships: AI Safety Summit & G7 Hiroshima process
Forums such as the AI Safety Summit and the G7 Hiroshima Process bring together governments, researchers and companies to define shared standards for safety and governance.
Industry initiatives : OpenAI, Anthropic, DeepMind
Industry leaders are committing to “responsible scaling” programmes and publishing safety reports to anticipate the risks associated with advanced models.
These efforts show that regulation and international cooperation are no longer optional: they are essential to governing AI development and prioritising responsible innovation.
More concretely, the recent Moltbook anecdote also serves as a reminder that security, in the context of agent-based platforms, must more than ever be an iterative and integrated process. The risks are now becoming systemic, due to the pace of development and deployment that is currently being observed.
AI and business: balancing innovation and trust
For companies, AI is not just a technical challenge—it is a major business issue. It promises unprecedented productivity gains and massive automation, but also brings risks of dependency, loss of control and disruption of economic models.
At SQLI, we firmly believe that AI must remain a tool serving humans and collective performance.
Innovation cannot be separated from responsibility.
Our actions are structured around four pillars:
- Active monitoring to anticipate developments,
- Continuous training to strengthen skills,
- Controlled experimentation to test safely,
- Co-development of reliable and useful solutions with our teams and partners.
Implementing monitoring, evaluation and safety mechanisms to ensure the reliability and control of AI models is essential. At SQLI, our experts integrate this approach from the design phase of agentic systems and throughout their advisory work for our clients.
Guillaume Le Moal, Technical AI Consultant, SQLI France
In the face of AI, collective action and transparency are our strongest allies.
Yoann Lathuiliere, Design Engineer Developer
Guillaume Le Moal, AI Technical Consultant, SQLI France
Sources :
- AI 2027
- Frontier models are capable of in-context scheming - Apollo Research
- Alignment Faking in large language models - Anthropic
- Rules for trustworthy artificial intelligence in the EU - European Union
- L’observatoire OCDE des politiques de l’IA - OECD.AI
- Ethique de l’intelligence artificielle - UNESCO
- AI Safety Summit 2023 - Government of the United Kingdom
- The Hiroshima AI Process - Government of Japan
- Scaling coordinated vulnerability disclosure - OpenAI
- Anthropic’s responsible scaling policy - Anthropic
- Responsibility & Safety, We want to build AI responsibly to benefit humanity - Google DeepMind
- Best of Moltbook - Astral Codex Ten
