# Proof: Current AI systems have already achieved Artificial General Intelligence (AGI).

- Generated: 2026-04-06
- Verdict: **DISPROVED**
- Audit trail: [proof_agi_audit.md](proof_agi_audit.md) | [proof_agi.py](proof_agi.py)

## Key Findings

- **4 of 4 independent authoritative sources** confirm that current AI systems have NOT achieved AGI, exceeding the disproof threshold of 3.
- Google DeepMind's formal taxonomy classifies current frontier models as **Level 1 ("Emerging") AGI** out of 6 levels — not yet "Competent" (Level 2) on most cognitive tasks (B1).
- Leading AI researcher Gary Marcus states current systems "do not exhibit the flexible, self-directed competence that the original concept of artificial general intelligence was intended to capture" (B2).
- No major AI research lab (OpenAI, Google DeepMind, Anthropic) has claimed AGI achievement. The only notable claim — from Nvidia CEO Jensen Huang (March 2026) — was widely criticized by the research community.

## Claim Interpretation

**Natural language claim:** "Current AI systems have already achieved Artificial General Intelligence (AGI)."

**Formal interpretation:** This is evaluated as a disproof. We search for authoritative, independent sources that reject the claim. AGI is interpreted using the most widely-cited frameworks:

1. **Google DeepMind's "Levels of AGI" paper** (Morris et al., 2023) — defines 6 levels from "No AI" (Level 0) to "Superhuman" (Level 5). Current frontier models are classified as Level 1 ("Emerging").
2. **OpenAI's internal 5-level framework** — places current systems at Level 2 ("Reasoners") out of 5 levels, with 3 additional stages required before full AGI.
3. **Expert survey consensus** — the majority of AI researchers surveyed do not believe AGI has been achieved.

The threshold of >= 3 independently verified sources rejecting the claim is conservative. This is the more stringent interpretation — fewer sources would make it easier to disprove.

## Evidence Summary

| ID | Fact | Verified |
|----|------|----------|
| B1 | Google DeepMind "Levels of AGI" framework — classifies current AI as Level 1 (Emerging) | Yes (fragment match, 81.4% coverage) |
| B2 | Gary Marcus (NYU) — current AI is not AGI, conflates statistical approximation with intelligence | Yes |
| B3 | Expert analysis — AGI not achieved, current systems lack autonomous goals and transfer learning | Yes |
| B4 | Tim Dettmers (UW) — AGI will not happen due to physical computation limits | Yes |
| A1 | Verified source count meeting disproof threshold | Computed: 4 verified sources confirm AGI not achieved, exceeding threshold of 3 |

## Proof Logic

The proof uses a qualitative consensus approach in disproof mode. Four independent sources were consulted, each arguing from a different perspective that AGI has not been achieved:

1. **Google DeepMind** (B1) provides the most formal framework. Their "Levels of AGI" paper establishes a 6-level taxonomy. Current frontier models are classified at Level 1 ("Emerging") — meaning performance equal to or somewhat better than an unskilled human across cognitive tasks. "Competent AGI" (Level 2) requires 50th-percentile performance of skilled adults on *most* cognitive tasks, which has not been achieved.

2. **Gary Marcus** (B2), a prominent AI researcher and NYU professor, argues that "current AI systems are powerful and increasingly useful tools, but they do not exhibit the flexible, self-directed competence that the original concept of artificial general intelligence was intended to capture." He identifies the core error as "conflating increasingly sophisticated statistical approximations with intelligence itself."

3. **Cogni Down Under** (B3) identifies specific capability gaps: current models "lack autonomous goal formation" and show "jagged intelligence" — excelling on some benchmarks while failing basic tasks.

4. **Tim Dettmers** (B4), a researcher at the University of Washington, argues from physical constraints: the exponential resource requirements for linear improvements mean "we have maybe one, maybe two more years of scaling left because further improvements become physically infeasible." His argument addresses not just whether AGI exists now, but fundamental barriers to achieving it.

All four sources (B1, B2, B3, B4) were successfully verified, exceeding the disproof threshold of 3.

## Counter-Evidence Search

Three adversarial searches were conducted looking for evidence that AGI HAS been achieved:

1. **Has any credible organization declared AGI achieved?** Nvidia CEO Jensen Huang stated "I think we've achieved AGI" (March 2026), but his claim was widely criticized. His definition relies narrowly on AI passing human exams — which tests narrow competencies, not general intelligence. No major AI research lab (OpenAI, Google DeepMind, Anthropic) has endorsed the claim. 76% of 475 AI researchers surveyed by AAAI said scaling current AI is unlikely to result in AGI.

2. **Do current systems pass any AGI benchmark?** While LLMs pass many standardized exams (bar exam, medical licensing, math olympiads), experts argue these test narrow competencies. DeepMind's 2026 cognitive framework identifies the largest gaps in learning, metacognition, attention, executive functions, and social cognition.

3. **Is there expert consensus that AGI is imminent?** Forecasters average only a 25% chance of AGI by 2029 and 50% by 2033. Stanford HAI experts stated "There will be no AGI this year" for 2026.

None of the adversarial checks produced evidence that breaks the disproof.

## Conclusion

**DISPROVED.** The claim that current AI systems have already achieved AGI is disproved by 4 independently verified authoritative sources, exceeding the threshold of 3. Three sources were fully verified via full quote match (B2 Gary Marcus, B3 Cogni Down Under, B4 Tim Dettmers), and one was verified via fragment match at 81.4% coverage (B1 DeepMind).

The disproof rests on multiple converging lines of evidence: formal AGI taxonomies that place current systems well below AGI thresholds, expert assessment that current AI lacks flexible self-directed competence, and physical scaling arguments that identify fundamental barriers. The only notable counter-claim (Jensen Huang, March 2026) was widely rejected by the research community.

Note: 3 citation(s) come from unclassified or low-credibility sources (tier 2). However, the authors cited (Gary Marcus, Tim Dettmers) are established AI researchers whose expertise is independently verifiable. See Source Credibility Assessment in the audit trail.

---
Generated by [proof-engine](https://github.com/yaniv-golan/proof-engine) v1.7.0 on 2026-04-06.