10 Reasons AI Doesn't Work On Your Science
For scientists who've tried it and felt the gap.
You uploaded your CSV and asked your question. The answer sounded confident. It even sounded right.
But something nagged at you. Maybe the AI quoted a "typical threshold" that didn't match what you see in your lab. Maybe it summarized your manuscript and missed the three findings that mattered. Maybe it forgot, three messages later, what you just spent twenty minutes explaining.
You're not imagining it.
Standard AI models were not built to read scientific data. They were built to read language. When you hand them a CSV, a 60-page grant draft, a 200-page manuscript, or a protocol PDF, they skim. Here's what's actually happening underneath.
1.The "context compacted" wall.
Long chat sessions hit a ceiling. The AI silently summarizes earlier turns. Anything you taught it three messages ago is gone, or compressed past the point of being useful.
You've been walking ChatGPT through your R01 specific aims for an hour. You ask one more question. It's forgotten your hypothesis, your prelim data, and the reviewer comment you were trying to address.
Your AI is forgetting everything you just taught it…
2.Silent truncation when you paste a big file.
The headlines say a million-token context window. The chat box you actually use does not give you that. On a paid Plus plan the working window is about 32,000 tokens. On the free tier it is roughly half that. A 200-page manuscript is several times larger than either. An 80MB CSV is not even close. The model reads whatever fit and answers from that, with no line telling you the rest was dropped.
You paste a 96-sample run sheet and ask if anything failed QC. It says every sample passed. It only ever saw the first 30 rows.
3.You can't tell from the answer whether the AI read your whole file.
No indicator. No warning. No "I only saw the first 40 pages." The answer just looks complete.
Paste the raw file: ~836 of 10,000 rows fit. The model never sees the rest.
4.Your file gets read like an essay.
Standard AI pattern-matches the surface text of a CSV, an Excel sheet, or a PDF. It treats column headers and section titles as phrases to interpret, not as fields with meaning.
You upload a 60-page R01 with twelve subsections, a budget, biosketches, and prior work. The AI hands you back a paragraph that sounds like the abstract. The reviewer-facing details you actually wanted to refine are nowhere.
5.Every field looks the same to the AI.
A sample ID, a reagent lot, a QC threshold, a protocol version, a grant number, a budget line, an author affiliation. All just strings of text. The AI has no idea what kind of fact each one is.
6.What failed never made it in.
The AI was trained on published papers. The 100 failures behind every paper, the dropped samples, the grants that didn't score, the manuscripts that got rejected. Invisible.
You ask why your library prep keeps failing on FFPE samples with low DIN. The AI gives you textbook advice from the papers where someone eventually got it to work. The 100 failures behind that one success aren't in the training data.
7.Provenance gets stripped.
Ask "where did this number come from?" and the AI has no answer. It doesn't know which file, which row, which version produced it.
The AI tells you the optimal annealing temperature is 58°C. You ask "based on what?" It can't tell you. Was it your protocol? A paper? Something it made up? You don't know.
8.Hallucinated certainty.
The AI returns plausible-sounding thresholds and "typical values" from training data. Your actual data says something different. The answer is wrong in a way that sounds right.
The AI says duplication rates above 40% are normal for low-input samples. Your facility's mean is 27.5%. Your 40% sample isn't normal. It needs a reprep.
9.Wikipedia-grade summaries instead of your-data-grade answers.
Ask "is my mapping rate of 78% acceptable?"
Textbook answer"Generally above 70% is acceptable."
Useful answer"Your facility's last 200 NovaSeq runs averaged 91%. Your 78% is two standard deviations low. Check your library."
10.No memory across sessions.
Tomorrow you open a new chat and paste the same 80MB CSV again. And again. Your postdoc joins the project on Thursday. She pastes it too. The grant goes in for revision. You paste the 60-page draft into a fresh session. Everyone re-uploads. Forever.
THE PATTERN, if you step back: standard AI treats scientific files the way a fast reader treats a dense paper at 2 AM. Skim for surface patterns. Fill the gaps with what they already know. Hand back something that sounds right.
For language tasks, fine.
For your data, not even close.
This is the problem we built KORE to solve.
Built by GRAVL. Organizing the world's unpublished scientific data.
Wondering how KORE compares to tools like MinerU or OpenRefine? Read the breakdown.
Try it on your own work
You can try this now: add KORE in Claude and ask it one question.
Common questions
Can I trust AI answers about my own research?
You uploaded your CSV and asked your question. The answer sounded confident. It even sounded right. But something nagged at you. You're not imagining it: standard AI models were not built to read scientific data. They were built to read language, so on your own grants, manuscripts, and results they skim, and fill the gaps with textbook priors.
Why does AI sound confident but still get my work wrong?
Standard AI treats scientific files the way a fast reader treats a dense paper at 2 AM. It skims for surface patterns, fills the gaps with what it already knows, and hands back something that sounds right. For language tasks, fine. For your data, not even close.
How do I get AI to use my own data and documents, not just the published literature?
The published literature is roughly the fraction of science that got written up. Your grants, manuscripts, protocols, and the results that never got published are the rest, and standard AI cannot hold or structure them. That is the problem KORE was built to solve: it structures your own material into a compact bundle any AI can actually use.
How do I stop AI from losing the thread of my specific aims halfway through a session?
Long chat sessions hit a ceiling. The AI silently summarizes earlier turns. Anything you taught it three messages ago is gone, or compressed past the point of being useful.
How do I get AI to work through my entire manuscript instead of skimming it?
The headlines say a million-token context window. The chat box you actually use does not give you that. On a paid Plus plan the working window is about 32,000 tokens. On the free tier it is roughly half that. A 200-page manuscript is several times larger than either. An 80MB CSV is not even close. The model reads whatever fit and answers from that, with no line telling you the rest was dropped.
How do I know whether AI actually read my whole document?
No indicator. No warning. No "I only saw the first 40 pages." The answer just looks complete.
How do I get AI to treat my tables as structured fields instead of loose text?
Standard AI pattern-matches the surface text of a CSV, an Excel sheet, or a PDF. It treats column headers and section titles as phrases to interpret, not as fields with meaning.
How do I get AI to understand what each value in my file actually means?
A sample ID, a reagent lot, a QC threshold, a protocol version, a grant number, a budget line, an author affiliation. All just strings of text. The AI has no idea what kind of fact each one is.
How do I get AI to use results that never got published, including the failures?
The AI was trained on published papers. The 100 failures behind every paper, the dropped samples, the grants that didn't score, the manuscripts that got rejected. Invisible.
How do I get AI to show where each number in its answer came from?
Ask "where did this number come from?" and the AI has no answer. It doesn't know which file, which row, which version produced it.
How do I stop AI from confidently giving me values that are wrong for my work?
The AI returns plausible-sounding thresholds and "typical values" from training data. Your actual data says something different. The answer is wrong in a way that sounds right.
How do I get AI to answer from my own results instead of the textbook?
Ask "is my mapping rate of 78% acceptable?" Textbook answer: "Generally above 70% is acceptable." Useful answer: "Your facility's last 200 NovaSeq runs averaged 91%. Your 78% is two standard deviations low. Check your library."
How do I avoid re-uploading the same document every session?
Tomorrow you open a new chat and paste the same 80MB CSV again. And again. Your postdoc joins the project on Thursday. She pastes it too. The grant goes in for revision. You paste the 60-page draft into a fresh session. Everyone re-uploads. Forever.