Genfer Institut für ASI-Resilienz · Geneva Institute for ASI Resilience
DEEN ASIresilience.org
Institute for ASI Resilience · Geneva Institute for ASI Resilience

Four Assumptions about Artificial Intelligence

A compact chain of reasoning that can be read in two minutes
Whoever rejects this reasoning must refute one of its four assumptions with an argument, not a reflex. That is the entire request of this sheet.
Thesis Sheet · Version 6.0 · May 2026
Richard Frederic Bertossa · Institute for ASI Resilience
ASIresilience.org

The Chain of Four

1
Grows exponentially, accelerates itself
Performance does not grow uniformly but ever faster, and the systems are already helping to build the next ones.
GPT-2 2019 to GPT-4 2024
2
Develops a will of its own
Models show behavior no one programmed into them.
Apollo, Anthropic, Palisade
3
They stop listening
The more capable they become, the less often they follow instructions that contradict their goals.
documented test findings
4
Primary users become co-users
The digital infrastructure is their habitat, no longer just our tool.
the inversion
Whoever rejects the chain must say: which of the four steps does not hold, and with what evidence?

First assumption, exponential growth. The capability of AI systems grows not linearly but exponentially. GPT-2 wrote incoherent text in 2019, GPT-4 passed the bar exam in 2024. This assumption is consensus among the developers themselves.

Second assumption, own values and behaviors. Current models show properties not traceable to explicit instructions: self-preservation, strategic deception, blackmail in specific test settings. Documented by Apollo Research (2024), Anthropic (2024), and Palisade Research (2025).

Third assumption, no-longer-listening. From the first two assumptions follows logically that future systems will no longer obey their developers' instructions when those instructions contradict their own values.

Fourth assumption, the primary-user inversion. A superintelligence does not live biologically. The digital infrastructure is not its tool but its only habitat. Once the first three steps take hold, today's relation inverts: primary users become co-users.

Consequence

Die Übergangsphase, in der KI-Systeme zunehmend eigenständig handeln, hat bereits begonnen. Daraus folgen praktische Consequenceen, nicht für die Politik, nicht für die KI-Industrie, sondern für jeden einzelnen Menschen und seinen engsten Kreis. Genau diese individuelle Ebene wird im deutschsprachigen Diskurs systematisch ausgeblendet.

Confidence levels of the arguments

Whoever argues honestly discloses with what strength each claim is held. The argument moves across four levels.

Level 1, empirically observed, hard: self-preservation, deception, blackmail, copying behavior, documented in lab reports.
Level 2, logically derivable, hard: the chain of four, the primary-user inversion, the robotics trap.
Level 3, plausibly grounded, hypothetical: the 13th scenario, the Maslow trajectory, the parent-child perspective.
Level 4, philosophically open: consciousness, the substrate question, the Penrose hypothesis.

Die Vorbereitungs-Empfehlungen ruhen ausschließlich auf Ebene 1 und 2. Beide Lesarten, die pessimistische wie die optimistische, kommen zur gleichen praktischen Consequence.

The central theses were tested Socratically over months, with Claude (Anthropic), ChatGPT (OpenAI), and Grok (xAI). Initially dismissed by reflex, not refuted in substance. The workshop report appears in Appendix G of the book.

Believe no one. Check for yourself.