When scientists at the Francis Crick Institute in London cut into human embryos with CRISPR in 2020, they expected to fix a disease-causing mutation. Instead, they watched entire chromosomes vanish. The gene-editing tool hadn't just made a typo—it had ripped out whole chapters of the genetic instruction manual.
This wasn't a fluke. It was a window into one of CRISPR's most persistent problems: the technology creates genetic mosaics, organisms where different cells carry different edits, often in ways nobody intended.
The Persistence Problem
The core issue is timing. When researchers inject CRISPR components into a fertilized egg, they're essentially releasing molecular scissors into a rapidly dividing cell. The Cas9 enzyme doesn't make its cut and disappear. It lingers, snipping DNA at multiple points during early development.
An edit at the one-cell stage affects every cell in the resulting organism. But if Cas9 waits until the embryo has divided into two cells, only half the organism gets that particular edit. Wait longer, and you get an increasingly complex patchwork. Different cells end up with different genetic changes—or no changes at all.
This creates what researchers call mosaicism, and it's more common than early CRISPR enthusiasts wanted to admit. In zebrafish studies that tracked over 1,100 animals across two generations, researchers found that adult founders were mosaic in their germ cells. Different sperm or egg cells carried different mutations. The randomness didn't stop at the edited generation—it passed to their offspring.
When CRISPR Overreacts
The mutations themselves turned out to be more dramatic than expected. Early validation methods used Sanger sequencing and other short-read techniques that looked for small insertions or deletions—a few DNA letters added or removed at the target site. These methods worked fine for detecting typos.
They missed the paragraphs CRISPR was deleting.
In 2018, a team led by Allan Bradley at the Wellcome Sanger Institute used long-range PCR and long-read sequencing to look more carefully at edited mouse and human cells. They found deletions spanning thousands of DNA letters, complex rearrangements, and structural chaos at sites that had looked clean under conventional analysis.
The 2020 Cell paper by Zuccaro and colleagues pushed this further. In human embryos edited to fix a mutation in the EYS gene, they documented something more extreme: allele-specific chromosome removal. The CRISPR machinery, faced with a double-strand break it had created, didn't repair the cut. Instead, cells simply ejected the entire chromosome carrying the mutation.
This wasn't repair. It was amputation.
The Six Percent You're Missing
Structural variants—insertions and deletions of 50 DNA letters or longer—represent about 6% of editing outcomes in founder animals. That might sound small, but it's the difference between a medical treatment and a genetic disaster.
In zebrafish studies, 26% of offspring from edited founders carried an off-target mutation somewhere in their genome. Nine percent carried a structural variant. These weren't theoretical risks calculated from computer models. They were actual mutations, detected in actual animals, using guide RNAs designed to minimize off-target activity.
The problem compounds across generations. Because founder animals are mosaic in their germlines, breeding programs become genetic roulette. You can't simply genotype a founder and know what its offspring will inherit. Each egg or sperm might carry a different version of the edit—or a different unintended mutation entirely.
The Clinical Reckoning
These findings arrived just as CRISPR was moving from research labs into clinical trials. The first human trials focused on ex vivo editing—removing cells, editing them in a dish, checking the results, then putting them back. Mosaicism matters less when you can screen edited cells before transplantation.
Germline editing is different. You can't pull an embryo apart to check every cell, then reassemble it. Any mosaicism in an edited embryo becomes permanent, potentially affecting every tissue in the resulting person.
The 2017 Nature paper by Shoukhrat Mitalipov's team, which claimed to have cleanly corrected a heart disease mutation in human embryos, sparked immediate controversy. Some researchers questioned whether the editing had worked as reported or whether something more complex—possibly involving chromosome loss—had occurred instead.
The subsequent discoveries of chromosome-level changes gave those concerns weight. If CRISPR could make cells eject entire chromosomes, the gap between "corrected mutation" and "removed chromosome carrying mutation" became harder to distinguish with standard genotyping.
Degradation as Solution
Researchers have started fighting back against mosaicism by limiting how long Cas9 remains active. Promoting enzyme degradation after the initial edit reduces the window for additional cuts. Using maternally expressed Cas9—where the enzyme comes from the egg rather than being injected—has shown promise for higher efficiency and lower mosaicism rates.
Optimizing Cas9 concentration helps too. More isn't always better. A lower dose that cuts once at the one-cell stage beats a higher dose that keeps cutting through multiple cell divisions.
Long-read sequencing has become essential for validation. Pacific Biosciences and Oxford Nanopore technologies can detect the large deletions and structural rearrangements that short-read methods miss. Several groups now recommend pre-testing for off-target activity and structural variants using patient material before any clinical application.
Breeding Out the Noise
For research applications, mosaicism isn't necessarily fatal to a project—just expensive and time-consuming. Researchers working with model organisms have adapted by breeding mosaic founders and screening their offspring. The second generation often segregates cleanly, with some animals carrying the intended edit and nothing else.
But this workaround doesn't translate to human applications. We can't breed out mosaic mutations across generations of people. The edit needs to work correctly the first time, in the embryo, with no patchwork and no surprises.
That's a higher bar than CRISPR currently clears. The technology that promised precision has delivered power—the power to cut DNA reliably—but precision remains elusive. Every cut is a controlled break. What happens after the break depends on cellular repair machinery that evolution designed for accidents, not intentional engineering.
The mosaics CRISPR creates aren't accidents in the sense of being unpredictable. They're the expected outcome of releasing an active enzyme into a dividing embryo. The accident was assuming we could control what happened next.