I sympathize. I’ve been caught out a couple of times by depending on autotldr as a substitute for reading the actual article. My own casual comparisons between autotldr and source articles suggest that autotldr is probably about 80% faithful to its source, on average.
I don’t know if it’s real or in my own mind, but it also seems to me that autotldr is faithful to the article inversely proportional to the quality of its source material. That is, the better and more complete the article, the more likely it is that autotldr trashes it.
Now that I’ve written it down, it strikes me that that may be an insurmountable problem. If we think of good articles as being “high information” and garbage articles as “low information”, summarizing will always be more likely to cause important “damage” the higher the information content. Thus, hitting 95% on a good article might trash it, while hitting 60% on a trash article is just fine. This might be especially true if you consider that the best articles might already be as compact as is reasonable.