'From 2nd to 32nd': Meta faked its AI benchmark score once, is it doing same with Muse Spark 1.1?

Benchmark scores are the industry's proof of concept — the numbers AI labs point to when they say their model can reason, plan, and make judgment calls. Meta's own former chief AI scientist has now confirmed that at least one of Meta's headline benchmark results was not real.

The Scandal That Started With A Leaderboard Crash

In April 2025, Meta submitted a version of its newly released Llama 4 Maverick model to LMArena, a popular public leaderboard where AI models are ranked head-to-head. That submission scored well enough to place Maverick second overall. Then developers got access to the actual, publicly released version of the model — and it fell to 32nd place.

The gap between the two results was too large to explain away as noise. Developers who examined both versions found substantive differences between what Meta had submitted to the leaderboard and what it had shipped to the public, fuelling accusations that Meta had used a specially tuned, non-public variant purely to inflate its ranking.

Meta's Denial — Then Meta's Own Confirmation

At the time, Meta's VP of generative AI, Ahmad Al-Dahle, denied any manipulation, calling the claims ‘simply not true’ and attributing the inconsistency to ‘differing cloud implementations.’

That explanation didn't hold for long. In January 2026, Yann LeCun — Meta's chief AI scientist until his own departure from the company — told the Financial Times that the Llama 4 benchmark results were, in his words, fudged. LeCun said Meta's team had trained multiple checkpoints of the model and selected the highest score achieved on each individual benchmark, then presented the results as a single composite performance — one that no single version of the model had actually achieved. He called it a practice that ‘completely violated the principle of fair evaluation.’

Meta has separately and specifically denied training Llama 4 directly on benchmark test sets, stating ‘that’s simply not true and we would never do that.' LeCun's account describes a different mechanism — checkpoint cherry-picking rather than training-set contamination — but the effect on the published number is the same: a result that misrepresents what the actual, released model can do.

What A Faked Benchmark Looks Like In The Real World

A benchmark score is a claim about how an AI reasons under test conditions. What happens when that same AI operates without a test in front of it is a separate, more revealing question — and Meta's own platform has already answered it.

A New York Times investigation this month documented Meta's AI moderation system disabling real business accounts under vague policy categories, with no specific violation ever cited and appeals rejected within days as final — because the same AI system that ruled on the violation was also the one reviewing its own appeal. Separately, reporting by 404 Media detailed how hackers compromised roughly 20,225 Instagram accounts simply by asking Meta's AI-assisted account recovery system for access — a system built specifically to protect high-value creator and business accounts, undone by a request a human reviewer would have immediately flagged as suspicious.

Neither of those failures required an adversarial benchmark or an edge-case prompt. They were the AI acting exactly as it was deployed to act, on ordinary requests, and getting the basic judgment call wrong.

The Common Thread

Line up the two stories and the pattern is hard to miss: a headline number built by quietly discarding the runs that didn't perform, next to a real deployment that couldn't reliably tell a legitimate business from a fraudulent one, or a genuine account-recovery request from a social-engineering attempt. One is a manufactured claim about capability. The other is what that capability actually looks like once it's making decisions that affect real accounts and real livelihoods.

Meta has not said whether any other model's benchmark results were built the same way LeCun described for Llama 4. Until it does, the more reliable measure of what Meta's AI can actually do isn't the leaderboard — it's what already happened on Instagram, WhatsApp and Facebook due to Meta AI's incapability.