Unfortunately, when people do find ways to automatically detect inconsistent numbers in a paper, like SPRITE or GRIM, they find maybe 50% of all checkable papers fail the audit and the authors almost invariably refuse to share their raw data for further checks despite having previously agreed to do so. So it seems likely that fake data is actually not rare at all but very widespread.
Caveat: like I say in every comment on scientific fraud on HN, all this varies dramatically by field. Some fields are much more corrupt than others. That 50% number is from psychology. Computer science isn't that bad relative to others. But for example, in climatology it's taken as axiomatic that it's OK to fabricate data and then present it as "observations". Anyone who tries to call them out on this behavior gets attacked, censored or even sued, yet if you did the exact same thing in physics nobody would hesitate to call it fraud. There isn't even a consistent definition in academia of what data fabrication means.
So no, sadly we don't really know how prevalent data fabrication is. The only way to detect scientific fraud to the level of robustness we'd accept for any other kind of fraud is regular, randomized lab audits with jail time or ruinous damages for perpetrators, backed by rigorous field-independent written standards for what evidence must be provided, all checked by outside organizations. Think financial audits but for scientists.
Obviously not only is academia not doing this, it's culturally nowhere close to considering the possibility of even thinking about starting. Academia has spent so many years engaging in such deep ideological purges that everyone who remains is at minimum sympathetic to ideas like "defund the police". The concept of policing themselves to the level accountancy does will never come up, not even during discussions of reform.