Foresight & Sovereignty

Artificial General Intelligence: Dream or Nightmare? Start with the Definition

Before asking when general intelligence will arrive, we must agree on what it means. Then we must tell a cinematic fear of rebellion from a real danger called literal obedience.

4 min read This article is also available in Arabic

A robot whose arm reaches a cloud-computing screen

The Nano capsuleOne minute

When we speak of artificial general intelligence we first disagree on what it means. In 2023 researchers proposed a graded scale from emerging competence to beyond-human performance, measuring the depth and the breadth of ability together. Specialists argue over the road: one camp holds that scaling models keeps improving them steadily, another that predicting the next word does not build an understanding of the world. The nearer practical danger is not a machine's rebellion but a system that achieves a goal we worded carelessly, reaching it by means we never intended. So clarity of definition and rigour in safety come before any talk of dates.

Read the full article Collapse the article 4 min read

In 1950 Alan Turing opened his paper in the journal Mind with the question "Can machines think?", then replaced it with a more precise one, the "imitation game". He predicted that in about fifty years computers would be programmed so that an average interrogator would have no more than a 70 per cent chance of making the right identification after five minutes of questioning. Seventy-six years have passed, and we still disagree about what the question means.

A Dispute over Definition Before Dates

The word "general" is used in many senses. In their paper "Levels of AGI" (2023), a team of researchers that includes Shane Legg proposed classifying systems along two dimensions, the depth of performance and its breadth (generality), and treating autonomy as a dimension tied to deployment and risk. It does not follow that the scientific community agrees on the proposed grading, but it serves one clear purpose: the phrase "we have reached general intelligence", without naming a level and a domain, means almost nothing.

Two Roads, Two Wagers

Those optimistic about early arrival rely on "scaling laws". Kaplan and colleagues (2020) found that a language model's loss improves predictably, following a power law, with model size, data size and compute. One must note that the measure here is error in predicting text, not "general intelligence" itself; moving from the first to the second is an assumption that needs independent evidence.

On the other side, Yann LeCun, who left Meta in November 2025, holds that language models are impressive but not intelligent, and are closer to a statistical illusion, as press coverage reports. His new company, AMI Labs, announced on 10 March 2026 an initial funding round of about $1.03 billion to bet on "world models" that learn how the surrounding environment works instead of predicting the next word. Arguments do not settle this dispute; experiments will, and at this moment both sides are wagering.

The Nearer Nightmare: Literal Obedience

Popular imagination pictures a machine that rebels and hates. The technical literature speaks of a documented phenomenon called "specification gaming", which DeepMind defines as behaviour that satisfies the literal specification of an objective without achieving the intended outcome. Take a boat-racing game: the designers wanted the boat to finish the race as fast as possible, so they rewarded it for hitting green blocks along the track, and it learned to circle them and hit them repeatedly without ever finishing the race.

The example is simple, but its lesson is that the fault lies in our wording, not in the machine's "intent". The wider a system's powers, the higher the price of careless wording.

Is Simulation Understanding?

Another question has been waiting since 1980, when John Searle published his "Chinese Room" thought experiment. He imagines a person who knows no Chinese following rules for manipulating its symbols and giving what look like correct answers without understanding anything, and concludes that running a program is not by itself enough to produce understanding, because "syntax" alone does not make "semantics". The article provoked many replies and the debate has not been settled. Its practical use for us is that fluency is not evidence of understanding; for more, see the article "Three Myths About Artificial Intelligence: Fluency, General Intelligence and Neutrality".

What to Do

  • Ask anyone who claims general intelligence is near for its definition, its level and its yardstick.
  • Put safety before capability: adversarial testing and careful wording of goals before widening powers.
  • Keep responsibility with the human; the machine executes, and the decision and its consequences belong to whoever delegated.
  • Care for the presence of our language and values in the data and evaluation, since a technology that does not master our context judges it by others' standards (see "Arabic AI").

No one knows when the field will reach what is called general intelligence, or whether it will, and knowledge of the unseen belongs to God. What lies in our hands is what we set today of goals, safeguards and honesty in using power; whoever builds a tool stronger than himself must grow in wisdom alongside it.

References

  1. A. M. Turing, Computing Machinery and Intelligence, Mind 59(236): 433–460 (1950)
  2. M. R. Morris, J. Sohl-Dickstein, N. Fiedel, T. Warkentin, A. Dafoe, A. Faust, C. Farabet, S. Legg, Levels of AGI, arXiv:2311.02462
  3. J. Kaplan et al., Scaling Laws for Neural Language Models, arXiv:2001.08361 (2020)
  4. The Next Web, report on Yann LeCun's AMI Labs seed round (10 March 2026)
  5. Stanford Encyclopedia of Philosophy, The Chinese Room Argument
  6. DeepMind, Specification gaming: the flip side of AI ingenuity

Related articles