If you've used any AI summarizer in the last two years, you know the genre. You paste in a long article or a YouTube link. You get back a wall of bullet points. Each bullet is a slightly-rephrased sentence from the original, hedged with the model's signature voice — "the speaker discusses..." "key insights include..." "in conclusion...". You scan the list, learn approximately nothing new, and close the tab.
This is the dominant form of AI summary on the market today. It's also terrible, and the reason is structural, not accidental.
The bullet-point trap
Why does every AI tool produce bullet-point summaries? Because bullet points are what falls out of the simplest possible prompt: "Summarize this in five key points." The model takes the input, identifies five-ish notable sentences, and lists them. It's the path of least resistance for both the model and the product team.
The problem is that summarization isn't list-making. A real summary takes a long argument and compresses it — preserving the structure, the causation, the tension between ideas — into a shorter argument. Bullet points strip exactly that structure out. You're left with disconnected facts that don't add up to anything.
A summary should answer the question "what is this thing about?" The bullet-point version answers "what words appeared in this thing?"
Three tests for a good AI summary
1. Does it have a thesis?
Read the summary. Can you state, in one sentence, what the original was actually arguing? If the summary is just "the speaker covers X, Y, and Z," it failed. A good summary has a position. It's the difference between "the speaker discusses pricing strategies" and "the speaker argues that most SaaS companies underprice by 30-50%."
2. Does it preserve causation?
Long-form content usually has logical structure: because A, therefore B, but watch out for C. Bullet-point summaries flatten this into three independent items. A good summary keeps the connective tissue intact.
3. Does it sound like prose?
Read the summary out loud. Does it sound like a human wrote it? Or does it sound like a machine that's read too much LinkedIn? "Leveraging cutting-edge insights" is a tell. So is the absence of contractions.
Why this is hard
Producing prose-level summaries is harder than producing bullet points, for two reasons.
First, the model has to understand the input, not just notice it. Bullet points can be generated by surface-level pattern matching. Prose summaries require the model to track an argument across multiple paragraphs, identify the actual claim, and rebuild it in fewer words.
Second, the model has to be willing to cut. Most LLMs default to listing everything they noticed, because listing things is safe and removing things is risky. A real summary requires the model to drop interesting-but-secondary points to make the central argument legible.
This is why we run a separate editing pass at RecapGPT instead of doing summarization in one shot. The editor model is specifically tuned to identify the thesis and cut everything else. It's slower, more expensive, and produces wildly better output.
What to look for
If you're shopping for an AI summarization tool, here's the test: pick a long-form piece of content you already know well. Run it through the tool. Read the output and ask:
- Could a stranger explain the original to me using only this summary?
- Is the central argument preserved, or just the surface-level topics?
- Would I be embarrassed to share this summary publicly?
If the answer to any of those is "no," the tool isn't good enough. Most aren't yet. The bar should be higher.