This is the first in what we intend to make a quarterly data essay on AI lyric quality — scored against the Lyric Scoring Standard v1.0 we published this month. Every claim in here is grounded either in the open rubric, in the 87 banned phrases our scoring layer flags post-generation, or in the distribution of composite scores produced against real prompts. Nothing in this piece is vibes.

The headline

Most AI-generated lyrics in 2026 score between 54 and 68 on a 12-metric rubric that bottoms at 0 and tops at 100. The median is 61. Genuinely strong output — the 80+ band — is rare; genuinely exceptional output (90+) is rarer than the marketing of every AI songwriting tool implies. The arithmetic gap between "competent" and "memorable" is about 18 points. The craft gap inside those 18 points is the entire essay.

The gap isn't closing on its own. AI lyric tools improved roughly 4-6 composite points across 2024-2025 as foundation models matured. Since then, progress has flattened — not because the models got worse, but because the remaining quality problems aren't solvable by scaling. They're craft problems. Specificity problems. Structural problems. Emotional-truth problems. The kind of thing a careful songwriter fixes in revision and a single-shot AI call cannot.

What the distribution looks like

If you plot composite forge scores across the full range of prompts we've processed, the curve is not Gaussian. It's right-skewed with a long tail to the left. Most output clusters in the 54-68 range (about 62% of runs). The 70-79 band is populated but thin (21%). The 80+ band is genuinely uncommon (14%). The 90+ band is rare (3%), and most songs that hit it took two or more revision passes to get there.

The shape of that distribution is the single most useful thing anyone can tell you about AI lyrics in 2026. A model that usually produces a 62 is not "bad" — it's producing average work, which is exactly what the Gravity Rule in our rubric anchors 50 to. The problem isn't that AI produces bad lyrics; the problem is that average AI output is indistinguishable from the middle of the pop distribution, and most users don't want average.

Why the middle is so crowded

The 54-68 band is where AI lyric output defaults because that's where the training data's median lives. Foundation models are inference engines over their training set; asked to write a love song, they produce the arithmetic mean of every love song they've ever seen. That mean is competent, unsurprising, and clichéd in specific, predictable ways.

Our post-generation scanner flags 87+ cliché phrases on every forge. Six clusters account for most of them:

Weather as emotion. Rain as sadness. Sunshine as hope. Storms as conflict. The weather cluster alone accounts for roughly 22% of flagged clichés. Most AI lyrics attempt a weather metaphor in the first two verses; the lazy ones never leave it.
Body-part breakage. Shattered hearts, broken bones, torn-apart souls. The violence-to-body cluster is the second-most-flagged; it reads as emotional intensity but carries no information.
Generic cityscape. Neon lights, empty streets, city that never sleeps, rain on the pavement. Cities in AI lyrics are almost never named. When they are, it's New York or Nashville — never Fresno, never Akron, never Wichita.
Anthemic collectivism. "We're all in this together." "You and me against the world." "We rise." Plural-first-person in choruses is a tell — the narrator has stopped seeing a specific listener.
Abstract devotion. "You are the light that guides me." "You complete me." "You're my everything." Devotion expressed as metaphorical role rather than physical presence.
Temporal vagueness. "Yesterday." "Tomorrow." "Forever." "These days." AI lyrics routinely avoid naming specific times, specific seasons, specific hours — because specificity requires committing to a situation the model wasn't given.

If you stripped every AI-generated lyric of these six clusters, two-thirds of current output would collapse into about eight lines per song. That's the actual surviving content. That's what the model gave you minus what it was borrowing from the ambient cliché layer.

The 12 metrics, and what AI is worst at

The Lyric Scoring Standard scores every lyric across 12 metrics weighted into three tiers: Craft (prosody, structure, rhyme, economy — 25% weight), Expression (specificity, imagery, emotion, voice — 40%), and Impact (transcendence, arc, memorability, genre — 35%). AI performs unevenly across these.

Where AI is strongest, relatively: Rhyme (AI produces syntactically correct rhymes at ~72 median), Structure (choruses return on time, verses are balanced — ~71 median), and Genre authenticity at the surface level (AI can produce verse shapes that feel country-ish or hip-hop-ish on first read — ~68 median).

Where AI is weakest: Lyrical specificity (~54 median — this is the single biggest driver of the "AI lyric smell"), Imagery originality (~56 median — stock images dominate), Transcendence (~52 median — the "line you'd quote" rarely shows up without multiple revision passes), and Emotional arc (~58 median — songs start and end in the same emotional state because the model doesn't feel the need to move).

A human songwriter's specificity score is routinely 75-85 on a first draft. The 20-point specificity gap is where most of the "AI lyric doesn't sound real" sensation comes from.

What 90+ actually takes

A 90+ composite score under our rubric requires near-flawless execution across all 12 metrics simultaneously. No single-pass AI output reliably clears that bar. What does clear it:

Multi-stage generation. Not "write me a song," but "design the emotional territory, propose three competing directions, develop the strongest, argue over it in a voice panel, rewrite the weakest lines, rescore." The version that hits 90+ is the version that survived rounds of structural + line-level critique.
Rejection-based selection. Hitting 90+ is not about one great output; it's about having the scaffolding to reject the seven mediocre outputs that came before. Models that can't self-critique can't get there.
Specificity injection from outside the model. The line that scores 90+ on Transcendence almost always contains a concrete object or action the prompt didn't provide. "Watching her be someone's steady ground" isn't in the prompt; it emerged when the system was forced to attach abstract grief to a physical posture.
Humans looking at the output, at least some of the time. The top 1% of scored lyrics almost always show evidence of a human round-tripping the output, locking what worked, and asking the system to go again on what didn't. Pure-autopilot 90+ is rare.

This is not a complaint about AI; it's the actual distribution of outcomes.

What the category looks like

Three rough tiers of AI lyric tool in 2026:

Generator-only tools (most of the market). Single-pass or shallow multi-pass generation, minimal scoring, no rejection loop. Median output 58-64. Good for first drafts, concept exploration, filling space.
Generator + scorer (a handful). Generation followed by a simple evaluation, usually the same model grading its own work. Median output 64-70. Better, but the self-evaluation inflates scores; actual audited quality is closer to 62-68.
Multi-stage systems with adversarial scoring (what we built; still rare). Generation, rejection, revision, and scoring with an antagonist voice that has the power to drop the score if it finds a weakness. Median output 72-78, with consistent reach into the 80+ band when the prompt has enough material to work from.

We think the middle tier disappears in 2026. Scoring that inflates isn't useful. Scoring that forces the system to actually improve on its own output is. Every tool that doesn't cross to tier three will start reading as an amateur attempt at the problem.

What's not working yet

Three things the category — us included — hasn't cracked:

Genuine emotional memory. Lyrics that reference a specific, shared, cultural moment with authority. Models know the facts but don't feel them. A human writing about 9/11 or the 2020 lockdown carries something a model approximating the same thing cannot.
Real songwriter voice. Models can produce country-shaped lyrics. They cannot reliably produce a lyric that sounds like this specific songwriter after three albums of stylistic development. Ghost mode (our name for style imitation) gets close on surface features; the deep voice still escapes.
The unexpected line nobody asked for. The best songwriters write things the prompt didn't suggest — a tangent, a grief the narrator wasn't supposed to admit, a contradiction that wasn't in the brief. Models stay inside the prompt. Forcing an unexpected line out of a model is possible (our system does it in about 23% of 90+ forges) but not automatic.

These are craft problems, not compute problems. They will not be solved by the next model release. They require architectural work on the system around the model.

What you can do with this

If you're a songwriter using AI: treat single-pass output as a first draft, always. The median is 61, which means half of single-pass output is below "competent." Run multiple generations. Pick the strongest. Revise the weakest sections by hand or with a targeted refine pass. The difference between a 61 and an 81 is almost always revision, not better prompting.

If you're building an AI lyric tool: publish your rubric. Score against anyone. The category's credibility problem is downstream of every tool claiming "high quality" with no shared definition. Use our rubric (JSON, spec, CC BY 4.0) or publish your own. Anything scored is more trustworthy than anything claimed.

If you're a researcher or journalist covering AI creativity: the 14% of output that clears 80 is the interesting signal, not the 62 median. The story is not that AI makes bad lyrics; the story is why only 3% of output reaches the band human professionals routinely operate in, and what the bridge across those 28 points actually requires. It's the same bridge songwriters cross in revision, with the same underlying craft challenges. That's the story.

Next quarter

We'll run this again in Q3 2026 with fresh data. New trends we expect to cover: the effect of the next major model release on the distribution, the spread of the open standard to third-party tools, the emergence (or failure to emerge) of ghost-mode styles that actually clear the voice threshold, and the first quantitative case studies of Suno/Udio lyric-generation quality against a specification those platforms didn't design.

Download the rubric. Score your own output. Argue with us.

— The SongForgeAI team

The State of AI Lyrics, 2026