We asked an AI to explain gravity.
Then we tried to break it.
In April 2026 an AI helped us write a bold theory: gravity is quantum entanglement. In November we put it through 329 automated checks and two independent AI reviews. Here is what survived, what didn't, and what the AI got wrong along the way.
Toggle between accessible and technical explanations
How it started
A persuasive idea, five numerical "tests", and a lot of green badges.
"Gravity isn't a force. It's entanglement."
The first version of this page argued that space is woven from quantum connections, that gravity is what that fabric looks like from inside, and that this would explain dark matter, black holes and the cosmological constant in one go.
Five tests, four marked "Confirmed"
An entanglement calculation, a lattice simulation, a collapse-time estimate, a fit to 2,693 galaxy data points, and a "first-principles derivation" of the galaxy law. Plus a Nature paper cited as external validation.
How we checked it
We treated every sentence as a claim that could be wrong, and wrote code to test it.
Recompute every number
Galaxy fits redone from the public SPARC data. Units, formulas and orbital-decay rates recalculated from scratch.
Verify every citation
Each paper fetched from arXiv. Did it exist, and did its abstract actually say what we claimed?
Get a second opinion. Twice.
A fresh AI agent, told to assume the work was wrong, reran everything and wrote its own code to try to break it. Then we did it again.
What didn't survive
Most of the original claims were either known physics under a new label, or wrong.
| Original claim | What the checks found | Verdict |
|---|---|---|
| A new law fits 2,693 points from 175 galaxies | The fit reproduces exactly, but it is McGaugh's 2016 Radial Acceleration Relation. Standard MOND fits identically. The points come from 153 galaxies, not 175. | Real, not new |
| The galaxy law is "derived, not fitted"; the square root comes from quantum mechanics | The square root is forced by the data (flat rotation curves). The alternative without it predicts rising rotation curves and misses by 0.22 dex. | Wrong |
| One number describes every galaxy | Our own repository's test said "not universal". Whether a₀ is truly universal is an open debate in the literature. | Unproven |
| Test 1: entanglement = geometric area (Confirmed) | A textbook free-fermion calculation (c = 1). Correct, but it tests nothing about gravity. | Mislabelled |
| Test 2: entanglement produces Newtonian 1/r gravity (Confirmed) | The 1/r was coded into the input. On a 2D lattice the Newtonian answer is ln r anyway. | Circular |
| a₀ = cH₀/2π "to within 7%" | A coincidence known since 1983. The uncertainty in H₀ itself is about 8%. | Coincidence |
| A Nature paper (Aziz & Howl 2025) independently supports the framework | At least eight published replies dispute its central result. The prediction that relied on it was dropped. | Disputed |
| Solves singularities, dark matter, the cosmological constant and the measurement problem | These were relabelled rather than solved, and some statements contradicted each other. | Overclaimed |
| Spacecraft propulsion; quantum computers work better in low gravity | No mechanism; the quantum-computing claim misreads the physics it cites. | Removed |
The auditor got it wrong too
The most useful finding of the whole exercise.
The AI checking the theory made more than 30 mistakes of its own
Wrong bounds, a missing term in its own equations, a flipped sign convention, misread papers, and a prediction it later had to retract. Each one looked plausible. The independent reviews caught them.
What survived
A modest, testable toy model. Not a theory of gravity.
"Smoothed MOND": the extra pull only sees matter blurred over a grain size ℓ
If space has a "grain", the galaxy-scale effect could switch off for small systems like the Solar System and binary stars, while still working for galaxies. The checks found this works for a grain size between roughly 0.04 and 100 parsecs (from about a tenth of a light-year to a few hundred light-years).
How it could be killed
Any excess of 6% or more at separations up to 5,000 AU rules it out
Given the Cassini limit, the model can produce at most about 3% there. A MOND-like signal would kill it. A Newtonian result would not. That flexibility is its main weakness.
Kill testAny excess should be about 1.6× stronger across the Galaxy's pull than along it
A distinctive fingerprint that ordinary contaminants (chance alignments, hidden third stars) shouldn't share.
SignatureThe galaxy acceleration scale should grow with the cosmic expansion rate
2026 data (MUSE-DARK) do show it rising, but current shape tests sit 2.1σ below this model's prediction.
Weak hookSo what is gravity?
Honestly: nobody knows yet. But a lot is settled, and the open questions are sharp.
Settled: gravity is the shape of spacetime
Einstein's picture works. A planet sitting still costs no energy to "keep pulling". Only moving, changing masses send out ripples, and we have measured those ripples carrying energy away.
Open: is spacetime quantum?
"Gravitons" would be quantum ripples of spacetime, not particles fired at you. Whether they exist is undecided. Lab experiments that try to entangle two masses through gravity alone could tell us in the 2030s.
Open: is space made of entanglement?
It is a serious research programme with real mathematical results, but so far only in simplified universes, not ours.
Open: dark matter or new gravity?
Galaxy clusters, the cosmic microwave background and the Bullet Cluster still point to unseen matter. The tight regularity of galaxy rotation is a genuine puzzle. Both may be partly right.
What this taught us about AI and research
AI is a superb synthesiser
It connected decades of literature and wrote working simulations in hours. That speed is real value.
It is a poor judge of its own work
Unchecked, it labelled things "Confirmed" that were circular or borrowed, and it was confidently wrong about its own maths.
Structure fixes most of it
Executable checks, verified citations and adversarial reviewers turned a persuasive story into an honest result. That is the workflow we now use.
Original AI-assisted synthesis: April 2026 · Audit and independent reviews: November 2026
Galaxy data: SPARC (Lelli, McGaugh & Schombert 2016) · Literature verified via arXiv
Not peer reviewed. The full audit report and check suite are available on request.
The Audit
Revision 3 of the review, after two independent passes. Every figure here is produced by an automated check.
The original five tests, re-run
| Test | Original label | Re-run result | Verdict |
|---|---|---|---|
| 1. Entanglement entropy | Confirmed | c = 1.000 for the free-fermion (XX) chain. S grows as (c/3) ln L: the Calabrese–Cardy result. | Textbook, not gravity |
| 2. Emergent metric | Confirmed | Coupling coded as 0.3 + 0.15/r. The 2D Newtonian potential is ln r (fit R² = 1.000 vs 0.856 for 1/r). | Circular |
| 3. Diósi–Penrose collapse | Testable | τ ≈ 0.3 s at 10¹² amu; the critical mass shifts by ~×2 with convention. The parameter-free model is already excluded experimentally. | Borrowed |
| 4. Galaxy data | Confirmed | N = 2693 (153 galaxies), a₀ = 1.162×10⁻¹⁰ m/s², scatter 0.133 dex, R² = 0.935. Simple MOND: identical (ΔBIC ≈ 0). | Reproduced; = MOND |
| 5. "Derivation" of the RAR | Derived | Low-acceleration slope of the data = 0.55 (deep-MOND requires 0.5). The "linear" alternative has slope 0 and misses by 0.22 dex. One candidate formula in the code has units of force. | Fitted, not derived |
What the 2026 literature says
| Topic | Result |
|---|---|
| Aziz & Howl (Nature 2025) | At least eight critical replies (Marletto, Oppenheim, Vedral, Wilson; Diósi; Lin & Mondal; Tang et al.; Tibau Vidal & Iyer; Xue et al.; and others). |
| Galaxy acceleration scale vs time | MUSE-DARK: a₀ ≈ 2.4×10⁻¹⁰ at z ≈ 1. A constant a₀ is excluded at 5.3σ; tracking H(z) fits at 2.1σ. Survey-to-survey systematics are ≥25%. |
| Solar System | Cassini: Q₂ = (1.6 ± 1.8)×10⁻²⁷ s⁻². Planetary ephemerides: dark-matter density < 1.1×10⁻²⁰ g/cm³ at Saturn. |
| Wide binaries | Mostly Newtonian (Cookson et al.; Saad & Ting). Boufourou: γ = 1.4 rejected at 16σ, but a small excess γ = 1.05 [1.04, 1.07] remains. Disputed by Chae & Yoon. Gaia DR4 should decide. |
| Dark energy | DESI DR2: 2.8–4.2σ preference for evolving dark energy; a "cosmic glitch" alternative (G_cosmo/G_N = 0.992) fits as well. |
| Entanglement → geometry | A holographic code tested on a trapped-ion quantum computer; a black-hole metric reconstructed from entanglement data alone. Toy and anti-de Sitter settings only. |
Smoothed MOND: the bounds
| Constraint | Bound on grain size ℓ | Status |
|---|---|---|
| Cassini Q₂ (3D FFT solve) | ℓ ≥ 0.024 pc | Robust |
| Planetary-ephemeris dark-matter limit | ℓ ≥ 0.037 pc | Robust |
| SPARC rotation curves (exact smoothing) | ℓ ≲ 100–200 pc (300 pc excluded) | Robust |
| Wide binaries Newtonian to 1–5% at 30 kAU | ℓ ≥ 0.12–0.22 pc | Conditional |
| Boufourou excess taken at face value | ℓ ≈ 0.03–0.06 pc | Conditional |
| Draco dwarf galaxy (rule of thumb) | ℓ ≲ 30–40 pc | Heuristic |
| Momentum conservation | Holds only if the force is smoothed too (unsmoothed form violates it ~20×) | Built in |
| Relativity, energy conservation, local action | Not achieved | Open |
The check suite
Each check states a claim, either from the original page or from the auditor, and marks it PASS, FAIL, WARN or INFO. A FAIL means the claim is false, so many FAILs are findings or retractions rather than bugs.
| Category | What it tests | Checks | PASS / FAIL / other |
|---|---|---|---|
| Citations | 72 arXiv papers exist and say what we claim | 121 | 120 / 0 / 1 |
| Numbers | Every quoted figure recomputed | 42 | 37 / 3 / 2 |
| Galaxy data | SPARC fits, universality, smoothing effects | 38 | 23 / 8 / 7 |
| Original page | Claims and code of the first version | 23 | 3 / 15 / 5 |
| Model | Smoothed MOND limits, bounds and predictions | 105 | 69 / 8 / 28 |
Python / NumPy / SciPy · SPARC database (Lelli, McGaugh & Schombert) · arXiv API
Foundations: Jacobson (1995), Ryu & Takayanagi (2006), Van Raamsdonk (2010), Verlinde (2016), McGaugh et al. (2016), Milgrom (1983)
Not peer reviewed. Full report, check suite and both independent reviews available on request.